US20020183971A1 - Diagnostic systems and methods for predictive condition monitoring - Google Patents

Abstract

A system for empirically diagnosing a condition of a monitored system. Estimates of monitored parameters from a model of the system provide residual values that can be analyzed for failure mode signature recognition. Residual values can also be tested for alert (non-zero) conditions, and patterns of alerts thus generated are analyzed for failure mode signature patterns. The system employs a similarity operator for signature recognition and also for parameter estimation. Failure modes are empirically determined, and precursor data is automatically analyzed to determine differentiable signatures for failure modes.

Description

BACKGROUND OF THE INVENTION
• 1. Field of the Invention[0001]
• The present invention relates generally to the field of early detection and diagnosis of incipient machine failure or process upset. More particularly, the invention is directed to model-based monitoring of processes and machines, and experience-based diagnostics.[0002]
• 2. Brief Description of the Related Art[0003]
• A variety of new and advanced techniques have emerged in industrial process control, machine control, system surveillance, and condition based monitoring to address drawbacks of traditional sensor-threshold-based control and alarms. The traditional techniques did little more than provide responses to gross changes in individual metrics of a process or machine, often failing to provide adequate warning to prevent unexpected shutdowns, equipment damage, loss of product quality or catastrophic safety hazards.[0004]
• According to one branch of the new techniques, empirical models of the monitored process or machine are used in failure detection and in control. Such models effectively leverage an aggregate view of surveillance sensor data to achieve much earlier incipient failure detection and finer process control. By modeling the many sensors on a process or machine simultaneously and in view of one another, the surveillance system can provide more information about how each sensor (and its measured parameter) ought to behave. Additionally, these approaches have the advantage that no additional instrumentation is typically needed, and sensors in place on the process or machine can be used.[0005]
• An example of such an empirical surveillance system is described in U.S. Pat. No. 5,764,509 to Gross et al., the teachings of which are incorporated herein by reference. Therein is described an empirical model using a similarity operator against a reference library of known states of the monitored process, and an estimation engine for generating estimates of current process states based on the similarity operation, coupled with a sensitive statistical hypothesis test to determine if the current process state is a normal or abnormal state. The role of the similarity operator in the above empirical surveillance system is to determine a metric of the similarity of a current set of sensor readings to any of the snapshots of sensor readings contained in the reference library. The similarity metric thusly rendered is used to generate an estimate of what the sensor readings ought to be, from a weighted composite of the reference library snapshots. The estimate can then be compared to the current readings for monitoring differences indicating incipient process upset, sensor failure or the like. Other empirical model-based monitoring systems known in the art employ neural networks to model the process or machine being monitored.[0006]
• Early detection of sensor failure, process upset or machine fault are afforded in such monitoring systems by sensitive statistical tests such as the sequential probability ratio test, also described in the aforementioned patent to Gross et al. The result of such a test when applied to the residual of the difference of the actual sensor signal and estimated sensor signal, is a decision as to whether the actual and estimate signals are the same or different, with user-selectable statistical confidence. While this is useful information in itself, directing thinly stretched maintenance resources only to those process locations or machine subcomponents that evidence a change from normal, there is a need to advance monitoring to a diagnostic result, and thereby provide a likely failure mode, rather than just an alert that the signal is not behaving as normal. Coupling a sensitive early detection statistical test with an easy-to-build empirical model and providing not only early warning, but a diagnostic indication of what is the likely cause of a change, comprises an enormously valuable monitoring or control system, and is much sought after in a variety of industries currently.[0007]
• Due to the inherent complexity of many processes and machines, the task of diagnosing a fault is very difficult. A great deal of effort has been spent on developing diagnostic systems. One approach to diagnosis has been to employ the use of an expert system that is a rule based system for analyzing process or machine parameters according to rules describing the dynamics of the monitored or controlled system developed by an expert. An expert system requires an intense learning process by a human expert to understand the system and to codify his knowledge into a set of rules. Thus, expert system development takes a large amount of time and resources. An expert system is not responsive to frequent design changes to a process or machine. A change in design changes the rules, which requires the expert to determine the new rules and to redesign the system.[0008]
• What is needed is a diagnostic approach that can be combined with model-based monitoring and control of a process or machine, wherein an expert is not required to spend months developing rules to be implemented in software for diagnosing machine or process fault. A diagnostic system that could be built on the domain knowledge of the industrial user of the monitoring or control system would be ideal. Furthermore, a diagnostic approach is needed that is easily adapted to changing uses of a machine, or changing parameters of a process, as well as design changes to both.[0009]
SUMMARY OF THE INVENTION
• The present invention provides diagnostic capabilities in a model-based monitoring system for machines and processes. A library of diagnostic conditions is provided as part of routine on-line monitoring of a machine or process via physical parameters instrumented with sensors of any type. Outputs created by the on-line monitoring are compared to the diagnostic conditions library, and if a signature of one or more diagnostic conditions is recognized in these outputs, the system provides a diagnosis of a possible impending failure mode.[0010]
• The diagnostic capabilities are preferably coupled to an empirical-model based system that generates estimates of sensor values in response to receiving actual sensor values from the sensors on the machine or process being monitored. The estimated sensor values generated by the model are subtracted from the actual sensor values to provide residual signals for sensors on the machine or process. When everything is working normally, as modeled by the empirical model, the residual signals are essentially zero with some noise from the underlying physical parameters and the sensor noise. When the process or machine deviates from any recognized and modeled state of operation, that is, when its operation becomes abnormal, these residuals become non-zero. A sensitive statistical test such as the sequential probability ratio test (SPRT) is applied to the residuals to provide the earliest possible decision whether the residuals are remaining around zero or not, often at such an early stage that the residual trend away from zero is still buried in the noise level. For any sensor where a decision is made that the residual is non-zero, an alert is generated on that sensor for the time snapshot in question. An alternative way to generate an alert is to enforce thresholds on the residual itself for each parameter, alerting on that parameter when the thresholds are exceeded. The diagnostic conditions library can be referenced using the residual data itself, or alternatively using the SPRT alert information or the residual threshold alert information. Failure modes are stored in the diagnostic conditions library, along with explanatory descriptions, suggested investigative steps, and suggested repair steps. When the pattern of SPRT alerts or residual threshold alerts matches the signature in the library, the failure mode is recognized, and the diagnosis made. Alternatively, when the residual data pattern is similar to a residual data pattern in the library using a similarity engine, the corresponding failure mode is recognized and the diagnosis made.[0011]
• The inventive system can comprise software running on a computer, with a memory for storing empirical model information and the diagnostic conditions library. Furthermore, it has data acquisition means for receiving data from sensors on the process or machine being monitored. Typically, the system can be connected to or integrated into a process control system in an industrial setting and acquire data from that system over a network connection. No new sensors need to be installed in order to use the inventive system. The diagnostic outputs of the software can be displayed, or transmitted to a pager, fax or other remote device, or output to a control system that may be disposed to act on the diagnoses for automatic process or machine control. Alternatively, due to the small computing requirements of the present invention, the inventive system can be reduced to an instruction set on a memory chip resident with a processor and additional memory for storing the model and library, and located physically on the process or equipment monitored, such as an automobile or aircraft.[0012]
• The diagnostic conditions library of the present invention is empirical, based on machine and process failure autopsies and their associated lead-in sensor data. The number of failure modes in the library is entirely selectable by the user, and the library can be added to in operation in the event that a new failure is encountered that is previously unknown in the library.[0013]
BRIEF DESCRIPTION OF THE DRAWINGS
• The novel features believed characteristic of the invention are set forth in the appended claims. The invention itself, however, as well as the preferred mode of use, further objectives and advantages thereof, is best understood by reference to the following detailed description of the embodiments in conjunction with the accompanying drawings, wherein:[0014]
• FIG. 1 shows a general arrangement for failure mode signature recognition using a database to identify likely failure modes from alert signals or residuals in accordance with the invention;[0015]
• FIG. 2 shows a prior art empirical model-based monitoring system with SPRT alert module;[0016]
• FIG. 3 shows a set of sensor signals, and the time-correlated sense of a “snapshot”;[0017]
• FIG. 4 is a chart showing a training method for an empirical model for use in the invention;[0018]
• FIG. 5 is a flowchart of the subject training method of FIG. 4;[0019]
• FIG. 6 illustrates a similarity operator that may be used for empirical modeling in a similarity engine with the present invention;[0020]
• FIG. 7 is a flowchart for carrying out the similarity operation;[0021]
• FIGS.[0022]8A-8D illustrate for a single sensor the actual sensor signal, estimate, alert index and alert decisions according to the monitoring system for use in the present invention;
• FIG. 9 illustrates a block diagram of a monitoring system according to the present invention, with three alternative avenues for using monitoring information for diagnostics;[0023]
• FIG. 10 is a flowchart for establishing a diagnostic library for a set of identical machines;[0024]
• FIG. 11 is a flowchart for establishing a diagnostic library for a process;[0025]
• FIGS.[0026]12A-12C illustrate alternative ranges from which to select failure mode signature information;
• FIG. 13 illustrates failure mode recognition by similarity operation;[0027]
• FIG. 14 illustrates similarity score generation for an input snapshot;[0028]
• FIG. 15 illustrates selection of a diagnosed failure mode on the basis of a highest similarity score;[0029]
• FIG. 16 illustrates selection of a diagnosed failure mode on the basis of a highest average similarity score;[0030]
• FIG. 17 shows failure mode recognition on the basis of an alert pattern; and[0031]
• FIG. 18 is a schematic block diagram of a hardware implementation of the present invention.[0032]
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
• Turning now to the drawings, and particularly FIG. 1, the preferred embodiment of the invention is set forth generally, in which a real-time[0033]data preprocessing module110 carries out monitoring operations on sensor data from a monitored machine or process, and outputs transformed data to a failure modesignature recognition module120. The transformed data can be alert patterns, residuals, and the like, derived from normal monitoring activities of themodule110. Therecognition module120 is connected to afailure mode database140, which contains signatures of transformed data and associated failure mode information. For example, if the transformed data is residual information, a signature can comprise a plurality of residual snapshots that are known to show themselves prior to that particular failure mode, and the associated failure mode information can comprise a description of the failure mode, a likelihood, an action plan for investigating the failure mode, or a corrective plan to fix the incipient failure. When signatures fromdatabase140 are recognize bymodule120, the associated identification and any corrective actions that should be taken are output in the failure mode diagnosis andactions output module160, which can communicate this to a display, or present the information in an object-based environment for automated action by a downstream control system or the like.
• The data preprocessing module can be any type of monitoring system, typically model-based, and more preferably empirical model-based. This is best understood with reference to FIG. 2, which illustrates a prior art empirical model-based monitoring system, such as that described in the aforementioned patent to Gross et al. Therein is shown a machine or[0034]process210 instrumented withsensors215 that have data acquisition means associated with them to provide the sensor data to any number of computing systems. Areference library230 of data characterizing the known or recognized states of operation of the machine or process is provided. Thereference library230 can reside in chip memory, or can be stored on a computer disk storage device. Anestimation model240 is implemented preferably in a computer as software, and receives sensor data fromsensors215 via a network or a data acquisition board. Theestimation model240 generates estimates of the sensor values in response to receiving the real-time values fromsensors215, using thereference library230, as described in greater detail below. Adifferencing unit250 receives both the estimates of the sensor values and the actual values and generates a residual for each sensor. Over successive snapshots, these residuals comprise residual signals that, as described above, should remain in the vicinity of zero with the exception of sensor and process noise, if the machine or process is operating normally (as characterized in the reference library data). ASPRT module260 receives the residuals and generates alerts if the residuals show definitive evidence of being other than zero.
• Therefore, the outputs of this prior art system include residual signals and SPRT alerts (which are really indications of difference), and one of each is provided for each sensor on the machine or process that is monitored.[0035]
• Turning to FIG. 3, the operation of the prior art system shown in FIG. 2 can further be understood in view of the multiple real-time sensor signals depicted therein. The[0036]vertical axis310 is a composite axis for the six sensor signals shown, and represents the signal amplitude.Axis320 is the time axis. The sensor signals in virtually all current industrial settings are sampled digitally, and are thus a sequence of discrete values, and a “snapshot”330 can be made at a point in time, which really represents a set ofvalues340 for each of the six sensors, each value representing the sensor amplitude at that time. Of course, in some industrial processes and machines, there is a time delay between cause and effect among sensors measuring physically correlated parameters of the process, and a time adjustment can be added to the data such that thesnapshot330 represents time-correlated, but not necessarily simultaneous, readings.
• An empirical model-based monitoring system for use in the present diagnostic invention requires historic data from which to “learn” normal states of operation, in order to generate sensor estimates. Generally, a large amount of data is accumulated from an instrumented machine or process running normally and through all its acceptable dynamic ranges. A method for selecting training set snapshots is graphically depicted in FIG. 4, for distilling the collected sensor data to create a representative training data set. In this simple example, five sensor signals[0037]402,404,406,408 and410 are shown for a process or machine to be monitored. Although the sensor signals402,404,406,408 and410 are shown as continuous, typically, these are discretely sampled values taken at each snapshot. As indicated hereinabove, snapshots need not be ordered in any particular order and so, may be ordered in chronological order, parametric ascending or descending order or in any other selected order. Thus, the abscissa axis412 is the sample number or time stamp of the collected sensor data, where the data is digitally sampled and the sensor data is temporally correlated. The ordinate axis414 represents the relative magnitude of each sensor reading over the samples or “snapshots.”
• In this example, each snapshot represents a vector of five elements, one reading for each sensor in that snapshot. Of all the collected sensor data from all snapshots, according to this training method, only those five-element snapshots are included in the representative training set that contain either a global minimum or a global maximum value for any given sensor. Therefore, the global maximum[0038]416 for sensor402 justifies the inclusion of the five sensor values at the intersections of line418 with each sensor signal402,404,406,408,410, including global maximum416, in the representative training set, as a vector of five elements. Similarly, the global minimum420 for sensor402 justifies the inclusion of the five sensor values at the intersections of line422 with each sensor signal402,404,406,408,410. Collections of such snapshots represent states the system has taken on. The pre-collected sensor data is filtered to produce a “training” subset that reflects all states that the system takes on while operating “normally” or “acceptably” or “preferably.” This training set forms a matrix, having as many rows as there are sensors of interest, and as many columns (snapshots) as necessary to capture all the acceptable states without redundancy.
• Selection of representative data is further depicted in the flow chart of FIG. 5. Data collected in[0039]Step500 has N sensors and L observations or snapshots or temporally related sets of sensor data that comprise Array X of N rows and L columns. InStep505, counter i (representing the element or sensor number) is initialized to zero, and observation or snapshot counter, t, is initialized to one. Moreover, Arrays max and min (containing maximum and minimum values, respectively, across the collected data for each sensor) are initialized to be vectors each of N elements which are set equal to the first column of X. Additional Arrays Tmax and Tmin (holding the observation number of the maximum and minimum value seen in the collected data for each sensor) are initialized to be vectors each of N elements, all zero.
• In[0040]Step510, if the sensor value of sensor i at snapshot t in X is greater than the maximum yet seen for that sensor in the collected data, max(i) is updated and set to equal the sensor value, while Tmax(i) stores the number t of the observation, as shown inStep515. If the sensor value is not greater than the maximum, a similar test is done for the minimum for that sensor, as illustrated inSteps520 and525. The observation counter t is then incremented inStep530. As shown inStep535, if all the observations have been reviewed for a given sensor (i.e., when the observation counter t equals the number of snapshots, L) then the observation counter t is reset to one and the counter i is incremented, as shown inStep540. At this point, the program continues to Step510 to find the maximum and minimum for the next sensor. Once the last sensor has been finished, at which point i=n, as shown inStep545, then any redundancies are removed and an array D is created from a subset of vectors from Array X. This creation process is discussed below.
• In[0041]Step550, counters i and j are both initialized to one. As illustrated byStep555, arrays Tmax and Tmin are concatenated to form a single vector Ttmp. Preferably, Ttmp has 2N elements, sorted into ascending (or descending) order, as shown inStep560 to form Array T. As shown inStep565, holder tmp is set to the first value in T (an observation number that contains a sensor minimum or maximum). Additionally, the first column of Array D is set to be equal to the column of Array X corresponding to the observation number that is the first element of T. In the loop starting with the decision box ofStep570, the ith element of T is compared to the value of tmp that contains the previous element of T. If they are equal (i.e., the corresponding observation vector is a minimum or maximum for more than one sensor), that vector has already been included in Array D and need not be included again. Counter i is then incremented, as shown inStep575. If the comparison is not equal, Array D is updated to include the column from X that corresponds to the observation number of T(i), as shown inStep580, and tmp is updated with the value at T(i). Counter j is then incremented, as shown inStep585, in addition to counter i (Step575). InStep590, if all the elements of T have been checked, and counter i equals twice the number of elements, N, then the distillation into training set or Array D has finished.
• Signal data may be gathered from any machine, process or living system that is monitored with sensors. Ideally, the number of sensors used is not a limiting factor, generally, other than concerning computational overhead. Moreover, the methods described herein are highly scalable. However, the sensors should capture at least some of the primary “drivers” of the underlying system. Furthermore, all sensors inputted to the underlying system should be interrelated in some fashion (i.e., non-linear or linear).[0042]
• Preferably, the signal data appear as vectors, with as many elements as there are sensors. A given vector represents a “snapshot” of the underlying system at a particular moment in time. Additional processing may be done if it is necessary to insert a “delay” between the cause and effect nature of consecutive sensors. That is, if sensor A detects a change that will be monitored by sensor B three “snapshots” later, the vectors can be reorganized such that a given snapshot contains a reading for sensor A at a first moment, and a reading for sensor B three moments later.[0043]
• Further, each snapshot can be thought of as a “state” of the underlying system. Thus, collections of such snapshots preferably represent a plurality of states of the system. As described above, any previously collected sensor data is filtered to produce a “training” subset (the reference set D) that characterizes all states that the system takes on while operating “normally” or “acceptably” or “preferably.” This training set forms a matrix, having as many rows as there are sensors of interest, and as many columns (snapshots) as necessary to capture the acceptable states without redundancy.[0044]
• According to this similarity operator-based empirical modeling technique, for a given set of contemporaneous sensor data from the monitored process or machine running in real-time, the estimates for the sensors can be generated according to:[0045]
• {right arrow over (Y)}_estimated={right arrow over (D)}·{right arrow over (W)}  (1)
• where the vector Y of estimated values for the sensors is equal to the contributions from each of the snapshots of contemporaneous sensor values arranged to comprise matrix D (the reference library or reference set). These contributions are determined by weight vector W. The multiplication operation is the standard matrix/vector multiplication operator. The vector Y has as many elements as there are sensors of interest in the monitored process or machine. W has as many elements as there are reference snapshots in D. W is determined by:[0046]$\begin{array}{cc}\overrightarrow{W}=\frac{\underrightarrow{\hat{W}}}{\left(\sum _{j=1}^N\hat{W}\left(j\right)\right)}& \left(2\right)\\ \underrightarrow{\hat{W}}={\left({\stackrel{\_}{D}}^T\otimes \stackrel{\_}{D}\right)}^{-1}·\left({\stackrel{\_}{D}}^T\otimes {\overrightarrow{Y}}_{in}\right)& \left(3\right)\end{array}$

  
• where the T superscript denotes transpose of the matrix, and Y[0047]_inis the current snapshot of actual, real-time sensor data. The improved similarity operator of the present invention is symbolized inEquation 3, above, as the circle with the “X” disposed therein. Moreover, D is again the reference library as a matrix, and D^Trepresents the standard transpose of that matrix (i.e., rows become columns). Y_inis the real-time or actual sensor values from the underlying system, and therefore is a vector snapshot.
• As stated above, the symbol ® represents the “similarity” operator, and can be chosen from a wide variety of operators for use in the present invention. Preferably, the similarity operation used in the present invention should provide a quantified measure of likeness or difference between two state vectors, and more preferably yields a number that approaches one (1) with increasing sameness, and approaches zero (0) with decreasing sameness. In the context of the invention, this symbol should not to be confused with the normal meaning of designation of {circumflex over (×)}, which is something else. In other words, for purposes of the present invention the meaning of {circumflex over (×)} is that of a “similarity” operation.[0048]
• The similarity operator, {circle over (×)}, works much as regular matrix multiplication operations, on a row-to-column basis. The similarity operation yields a scalar value for each pair of corresponding n[0049]^thelements of a row and a column, and an overall similarity value for the comparison of the row to the column as a whole. This is performed over all row-to-column combinations for two matrices (as in the similarity operation on D and its transpose above).
• By way of example, one similarity operator that can be used compares the two vectors (the i[0050]^throw and j^thcolumn) on an element-by-element basis. Only corresponding elements are compared, e.g., element (i,m) with element (m,j) but not element (i,m) with element (n,j). For each such comparison, the similarity is equal to the absolute value of the smaller of the two values divided by the larger of the two values.
• Hence, if the values are identical, the similarity is equal to one, and if the values are grossly unequal, the similarity approaches zero. When all the elemental similarities are computed, the overall similarity of the two vectors is equal to the average of the elemental similarities. A different statistical combination of the elemental similarities can also be used in place of averaging, e.g., median.[0051]
• Another example of a similarity operator that can be used can be understood with reference to FIG. 6. With respect to this similarity operator, the teachings of U.S. Pat. No. 5,987,399 to Wegerich et al. are relevant, and are incorporated herein by reference. For each sensor or physical parameter, a[0052]triangle620 is formed to determine the similarity between two values for that sensor or parameter. Thebase622 of the triangle is set to a length equal to the difference between theminimum value634 observed for that sensor in the entire training set, and themaximum value640 observed for that sensor across the entire training set. An angle Ω is formed above that base622 to create thetriangle620. The similarity between any two elements in a snapshot-to-snapshot operation is then found by plotting the locations of the values of the two elements, depicted as X₀and X₁in the figure, along thebase622, using at one end the value of the minimum634 and at the other end the value of the maximum640 to scale thebase622.
• [0053]Line segments658 and660 drawn to the locations of X₀and X₁on thebase622 form an angle θ. The ratio of angle θ to angle Ω gives a measure of the difference between X₀and X₁over the range of values in the training set for the sensor in question. Subtracting this ratio, or some algorithmically modified version of it, from the value of one yields a number between zero and one that is the measure of the similarity of X₀and X₁.
• Yet another example of a similarity operator that can be used determines an elemental similarity between two corresponding elements of two observation vectors or snapshots, by subtracting from one a quantity with the absolute difference of the two elements in the numerator, and the expected range for the elements in the denominator. The expected range can be determined, for example, by the difference of the maximum and minimum values for that element to be found across all the reference library data. The vector similarity is then determined by averaging the elemental similarities.[0054]
• In yet another similarity operator that can be used in the present invention, the vector similarity of two observation vectors is equal to the inverse of the quantity of one plus the magnitude Euclidean distance between the two vectors in n-dimensional space, where n is the number of elements in each observation.[0055]
• Elemental similarities are calculated for each corresponding pairs of elements of the two snapshots being compared. Then, the elemental similarities are combined in some statistical fashion to generate a single similarity scalar value for the vector-to-vector comparison. Preferably, this overall similarity, S, of two snapshots is equal to the average of the number N (the element count) of s[0056]_cvalues:$\begin{array}{cc}S=\frac{\sum _{c=1}^NS_c}{N}& \left(4\right)\end{array}$

  
• Other similarity operators are known or may become known to those skilled in the art, and can be employed in the present invention as described herein. The recitation of the above operators is exemplary and not meant to limit the scope of the claimed invention. The similarity operator is used in this invention as described below for calculation of similarity values between snapshots of residuals and the diagnostic library of residual snapshots that belie an incipient failure mode, and it should be understood that the description above of the similarity operation likewise applies to the failure mode signature recognition using residuals.[0057]
• Turning to FIG. 7, the generation of estimates is further shown in a flowchart. Matrix D is provided in[0058]step702, along with the input snapshot vector y_inand an array A for computations. A counter i is initialized to one instep704, and is used to count the number of observations in the training matrix D. Instep706, another counter k is initialized to one (used to count through the number of sensors in a snapshot and observation), and array A is initialized to contain zeroes for elements.
• In[0059]step708, the element-to-element similarity operation is performed between the kth element of y_inand the (ith, kth) element in D. These elements are corresponding sensor values, one from actual input, and one from an observation in the training history D. The similarity operation returns a measure of similarity of the two values, usually a value between zero (no similarity) and one (identical) which is assigned to the temporary variable r. Instep710, r divided by the number of sensors M is added to the ith value in the one-dimensional array A. Thus, the ith element in A holds the average similarity for the elemental similarities of y_into the ith observation in D. Instep712, counter k is incremented.
• In[0060]step714, if all the sensors in a particular observation in D have been compared to corresponding elements of y_in, then k will now be greater than M, and i can be incremented instep716. If not, then the next element in y_inis compared for similarity to its corresponding element in D.
• When all the elements of the current actual snapshot y[0061]_inhave been compared to all elements of an observation in D, a test is made instep718 whether this is the last of the observations in D. If so, then counter i is now more than the number of observations N in D, and processing moves to step720. Otherwise, it moves back to step706, where the array A is reset to zeroes, and the element (sensor) counter k is reset to one. Instep720, a weight vector W-carrot is computed from the equation shown therein, where {circle over (×)} represents a similarity operation, typically the same similarity operator as is used instep708. In step722 W-carrot is normalized using a sum of all the weight elements in W-carrot, which ameliorates the effects in subsequent steps of any particularly large elements in W-carrot, producing normalized weight vector W. Instep724, this is used to produce the estimated output y_outusing D.
• Examples of various preprocessed data that can be used for diagnostics as a consequence of monitoring the process or machine as described in detail herein are shown in connection with FIGS.[0062]8A-8D. FIG. 8A shows both the actual signal and the estimated signal for a given sensor, one of potentially many sensors that are monitored, modeled and estimated in theestimation model240 from FIG. 2.
• FIG. 8B shows the resulting residual signal from differencing the signals in FIG. 8A, as is done in the[0063]differencing module250 of FIG. 2. As can be seen on examination of FIG. 8B, the sensor residual takes on a series of non-zero values that lead to the eventual failure. In another failure mode, the series of values taken on may be different, such that the residuals for all the sensors in the monitored system contain information for differentiating the onset of one kind of failure from another, which is essentially a first step in diagnostics. The alert index of FIG. 8C and the alert decisions of FIG. 8D are discussed below, but also provide information that can be used to diagnose an impending failure. In FIG. 8D, each asterisk on thebottom line810 indicates a decision for a given input snapshot that for this sensor, the actual and the estimated value are the same. Asterisks on thetop line820 indicate a point in the series of snapshots for which the estimate for this sensor and the actual appear to have diverged.
• One decision technique that can be used according to the present invention to determine whether or not to alert on a given sensor estimate is to employ thresholds for the residual for that sensor. Thresholds as used in the prior art are typically used on the gross value of a sensor, and therefore must be set sufficiently wide or high to avoid alerting as the measured parameter moves through its normal dynamic range. A residual threshold is vastly more sensitive and accurate, and is made possible by the use of the sensor value estimate. Since the residual is the difference between the actual observed sensor value and the estimate of that value based on the values of other sensors in the system (using an empirical model like the similarity engine described herein), the residual threshold is set around the expected zero-mean residual, and at a level potentially significantly narrower than the dynamic range of the parameter measured by that sensor. According to the invention, residual thresholds can be set separately for each sensor. The residual thresholds can be determined and fixed prior to entering real-time monitoring mode. A typical residual threshold can be set as a multiple of the empirically determined variance or standard deviation of the residual itself. For example, the threshold for a given residual signal can be set at two times the standard deviation determined for the residual over a window of residual data generated for normal operation. Alternatively, the threshold can be determined “on-the-fly” for each residual, based on a multiplier of the variance or standard deviation determined from a moving window of a selected number of prior samples. Thus, the threshold applied instantly to a given residual can be two times the standard deviation determined from the past hundred residual data values.[0064]
• Another decision technique that can be employed to determine whether or not to alert on a given sensor estimate is called a sequential probability ratio test (SPRT), and is described in the aforementioned U.S. Pat. No. 5,764,509 to Gross et al. It is also known in the art, from the theory of Wald and Wolfowitz, “Optimum Character of the Sequential Probability Ratio Test”, Ann. Math. Stat. 19, 326 (1948). Broadly, for a sequence of estimates for a particular sensor, the test is capable of deciding with preselected missed and false alarm rates whether the estimates and actuals are statistically the same or different, that is, belong to the same or to two different probability distributions.[0065]
• The basic approach of the SPRT technique is to analyze successive observations of a sampled parameter. A sequence of sampled differences between the estimate and the actual for a monitored parameter should be distributed according to some kind of distribution function around a mean of zero. Typically, this will be a Gaussian distribution, but it may be a different distribution, as for example a binomial distribution for a parameter that takes on only two discrete values (this can be common in telecommunications and networking machines and processes). Then, with each observation, a test statistic is calculated and compared to one or more decision limits or thresholds. The SPRT test statistic generally is the likelihood ratio I[0066]_n, which is the ratio of the probability that a hypothesis H₁is true to the probability that a hypothesis H₀is true:$\begin{array}{cc}{l}_n=\frac{\left({\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00000.png,US20020183971A1-20021205-D00004.png"\; state="\{\{state\}\}">y</figure-callout>}}_1,{\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">y</figure-callout>}}_2,\dots ,y_nH_1\right)}{\left({\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00000.png,US20020183971A1-20021205-D00004.png"\; state="\{\{state\}\}">y</figure-callout>}}_1,{\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">y</figure-callout>}}_2,\dots ,y_nH_0\right)}& \left(5\right)\end{array}$

  
• where Y[0067]_nare the individual observations and H_nare the probability distributions for those hypotheses. This general SPRT test ratio can be compared to a decision threshold to reach a decision with any observation. For example, if the outcome is greater than 0.80, then decide H₁is the case, if less than 0.20 then decide H₀is the case, and if in between then make no decision.
• The SPRT test can be applied to various statistical measures of the respective distributions. Thus, for a Gaussian distribution, a first SPRT test can be applied to the mean and a second SPRT test can be applied to the variance. For example, there can be a positive mean test and a negative mean test for data such as residuals that should distribute around zero. The positive mean test involves the ratio of the likelihood that a sequence of values belongs to a distribution H[0068]₀around zero, versus belonging to a distribution H₁around a positive value, typically the one standard deviation above zero. The negative mean test is similar, except H₁is around zero minus one standard deviation. Furthermore, the variance SPRT test can be to test whether the sequence of values belongs to a first distribution H₀having a known variance, or a second distribution H₂having a variance equal to a multiple of the known variance.
• For residuals derived from known normal operation, the mean is zero, and the variance can be determined. Then in run-time monitoring mode, for the mean SPRT test, the likelihood that Ho is true (mean is zero and variance is σ[0069]²) is given by:$\begin{array}{cc}L\left({\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00000.png,US20020183971A1-20021205-D00004.png"\; state="\{\{state\}\}">y</figure-callout>}}_1,{\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">y</figure-callout>}}_2,\dots ,y_nH_0\right)=\frac{1}{{\left(2\mathrm{\pi \sigma }\right)}^{n/2}}{}^{\left[-\frac{1}{2{\mathrm{<figure-callout\; label="\sigma "\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}^2}\sum _{k=1}^ny_k^2\right]}& \left(6\right)\end{array}$

  
• and similarly, for H[0070]₁, where the mean is M (typically one standard deviation below or above zero, using the variance determined for the residuals from normal operation) and the variance is again σ²(variance is assumed the same):$\begin{array}{cc}L\left({\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00000.png,US20020183971A1-20021205-D00004.png"\; state="\{\{state\}\}">y</figure-callout>}}_1,{\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">y</figure-callout>}}_2,\dots ,y_nH_1\right)=\frac{1}{{\left(2\mathrm{\pi \sigma }\right)}^{n/2}}{}^{\left[-\frac{1}{2{\mathrm{<figure-callout\; label="\sigma "\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}^2}\left(\sum _{k=1}^ny_{\mathrm{<figure-callout\; label="k"\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">k</figure-callout>}}^{2}-2\sum _{k=1}^ny_kM+\sum _{k=1}^nM^2\right)\right]}& \left(7\right)\end{array}$

  
• The[0071]ratio1_nfromequations 6 and 7 then becomes:$\begin{array}{cc}{l}_n={}^{\left[-\frac{1}{2{\mathrm{<figure-callout\; label="\sigma "\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}^2}\sum _{k=1}^nM\left(M-2y_k\right)\right]}& \left(8\right)\end{array}$

  
• A SPRT statistic can be defined for the mean test to be the exponent in equation 8:[0072]$\begin{array}{cc}{\mathrm{SPRT}}_{\mathrm{mean}}=-\frac{1}{2{\mathrm{<figure-callout\; label="\sigma "\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}^2}\sum _{k=1}^nM\left(M-2y_k\right)& \left(9\right)\end{array}$

  
• The SPRT test is advantageous because a user-selectable false alarm probability α and a missed alarm probability β can provide thresholds against with SPRT[0073]_meancan be tested to produce a decision:
• 1. If SPRT[0074]_mean≦1n(β/(1−α)), then accept hypothesis H₀as true;
• 2. If SPRT[0075]_mean≧1n((1−β)/α), then accept hypothesis H1 as true; and
• [0076]3. If 1n(β/(1−α))<SPRT_mean<1n((1−β)/α), then make no decision and continue sampling.
• For the variance SPRT test, the problem is to decide between two hypotheses: H[0077]₂where the residual forms a Gaussian probability density function with a mean of zero and a variance of Vσ²; and H₀where the residual forms a Gaussian probability density function with a mean of zero and a variance of σ². The likelihood that H₂is true is given by:$\begin{array}{cc}L\left({\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00000.png,US20020183971A1-20021205-D00004.png"\; state="\{\{state\}\}">y</figure-callout>}}_1,{\mathrm{<figure-callout\; label="y"\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">y</figure-callout>}}_2,\dots ,y_nH_2\right)=\frac{1}{{\left(2\pi {\mathrm{<figure-callout\; label="V"\; filenames="US20020183971A1-20021205-D00000.png,US20020183971A1-20021205-D00004.png"\; state="\{\{state\}\}">V</figure-callout>}}^{1/2}\sigma \right)}^{n/2}}{}^{\left[-\frac{1}{2{\mathrm{V\sigma }}^2}\sum _{k=1}^ny_k^2\right]}& \left(10\right)\end{array}$

  
• The[0078]ratio 1_nis then provided for the variance SPRT test as the ratio ofequation 10 overequation 6, to provide:$\begin{array}{cc}{l}_n=V^{-1/2}{}^{\left[-\frac{1}{2{\mathrm{<figure-callout\; label="\sigma "\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}^2}\sum _{k=1}^ny_{\mathrm{<figure-callout\; label="k"\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">k</figure-callout>}}^2\left(\frac{1-V}{V}\right)\right]}& \left(11\right)\end{array}$

  
• and the SPRT statistic for the variance test is then:[0079]$\begin{array}{cc}{\mathrm{SPRT}}_{\mathrm{variance}}=\frac{1}{2{\mathrm{<figure-callout\; label="\sigma "\; filenames="US20020183971A1-20021205-D00004.png,US20020183971A1-20021205-D00012.png"\; state="\{\{state\}\}">\sigma </figure-callout>}}^2}\left(\frac{V-1}{V}\right)\sum _{k=1}^ny_k^2-\frac{\ln V}{2}& \left(12\right)\end{array}$

  
• Thereafter, the above tests (1) through (3) can be applied as above:[0080]
• 1. If SPRT[0081]_variance≦1n(β/(1−α)), then accept hypothesis H₀as true;
• [0082]2. If SPRT_variance≧1n((1−β)/α), then accept hypothesis H₂as true; and
• [0083]3. If 1n(β/(1−α))<SPRT_variance<1n((1−)/α), then make no decision and continue sampling.
• Each snapshot that is passed to the SPRT test module, can have SPRT test decisions for positive mean, negative mean, and variance for each parameter in the snapshot. In an empirical model-based monitoring system according to the present invention, any such SPRT test on any such parameter that results in an hypothesis other than Ho being accepted as true, is effectively an alert on that parameter. Of course, it lies within the scope of the invention for logic to be inserted between the SPRT tests and the output alerts, such that a combination of a non-H[0084]₀result is required for both the mean and variance SPRT tests in order for the alert to be generated for the parameter, or some other such rule.
• Turning now to the diagnostic function coupled to the model-based monitoring system, depicted in FIG. 9 is the[0085]embodiment902 showing the threealternative avenues906,910 and914 for monitoring data to be passed to the failure signature recognition module916 (dashed lines) for failure mode recognition. Therein is shown a machine or process ofinterest918, instrumented withmultiple sensors920. The sensor data is passed (preferably in real time) to a model922 (preferably empirical, with a reference library or training set923) and also to adifferencing module924. Themodel922 generates estimates that are compared to the actual sensor values in thedifferencing module924 to generate residuals, which are passed to analert test927. Thealert test927 can be the SPRT, or can be residual threshold alerts as described above, or any other alert technique based on the residual. Alerts are generated on detection of deviations from normal, as described above. Alerts may optionally be output from the system in addition to any diagnostic information.Avenue906 shows that actual sensor snapshots can be passed to the failuresignature recognition module916, such that themodule916 compares the actual snapshots to stored snapshots in thefailure mode database930, and upon sufficient match (as described below) the failure mode is output corresponding to that belied by the actual sensor snapshots. Avenue910 represents the alternative embodiment, where residual snapshots (comprising usually near-zero values for each of the monitored sensors) are passed to themodule916, and are compared to stored snapshots of residuals that are known to precede recognized failure modes, and upon a match (as described below), the corresponding failure mode is output. In the third alternative,avenue914 provides for feeding test alerts, more particularly SPRT alerts or residual threshold alerts from thetest927 to themodule916, which compares these, or a sequence of these over time, to SPRT or residual threshold alert patterns (as described below) stored in thedatabase930, and upon a match outputs the corresponding failure mode. As described elsewhere herein, the output of the failure mode can be a display or notification of one or more likely failure modes, investigative action suggestions, and resolution action suggestions, which are all stored in the database with the related failure mode signature. The inventive system also provides for the addition of new failure modes based on actual snapshots, residual snapshots, or alert patterns, by the user in the event none of the failure modes in thedatabase930 sufficiently match the precursor data to the failure. Thus three sources of data can be recognized for failure signatures are presented: 1) Actual sensor data coming from the machine or process of interest; 2) residual data coming from the differencing module; and 3) SPRT or alert test patterns.
• In the generalized model of FIG. 1, a similarity engine may be employed for failure mode signature recognition (regardless of whether a similarity engine is used to do the initial modeling and estimate generation) that operates on either residual or actual signals using the[0086]database140 to identify likely failure modes for automatic feedback control with associated probabilities of the failure modes. Thesignature recognition module140 may be provided with historic data (actuals or residuals) of signatures leading up to historic failures of known mode. Failure mode recognition can execute in parallel with ongoing regular operation of the traditional similarity operator monitoring technology.
• Turning to FIG. 10, an implementation method is shown for populating the[0087]failure mode database930 of FIG. 9 (ordatabase140 of FIG. 1) with precursor data for signature matching, and associated probabilities and action suggestions, for application of the present invention to a production run of identical machines that are designed to have on-board self-diagnostic capabilities. An example of such a machine may be an instrumented electric motor. Instep1010, a plurality of the identical machines are instrumented with sensors as they would be in the field. These machines will be run to failure and ruined, in order to discover the various modes of failure of the machine design. Therefore, a sufficiently large number should be used to provide some statistical measure of the likelihood of each failure mode and to provide sufficient representative precursor data for each failure mode. In step1015, data collection is performed as the instrumented machines are run through routine operational ranges. Instep1020, at least some of the data (preferably from early operation of the machines, before they begin to degrade) is captured for use in building the reference library for the empirical model, if that method of monitoring is to be used. Instep923, the machines are all run to failure, and data is captured from the sensors as they fail.
• In[0088]step1031, the captured data is processed to isolate precursor data for each failure mode. Failure modes are selected by the user of the invention, and are logical groupings of the specific findings from autopsies of each machine failure. The logical groupings of autopsied results into “modes” of failure should be sensible, and should comport with the likelihood that the precursor data leading to that failure mode will be the same or similar each time. However, beyond this requirement, the user is free to group them as seen fit. Thus, for example, a manufacturer of an electric motor may choose to run 50 motors to failure, and upon autopsy, group the results into three major failure modes, related to stator problems, mechanical rotating pieces, and insulation winding breakdown. If these account for a substantial majority of the failure modes of the motor, the manufacturer may choose not to recognize other failure modes, and will accept SPRT or residual threshold alerts from monitoring with no accompanying failure mode recognition as essentially a recognition of some uncommon failure.
• According to another method of the invention, commonly available analysis methods known to those in the art may be used to self-organize the precursor data for each instance of failure into logical groupings according to how similar the precursor data streams are. For example, if the user divines a distinct autopsy result for each of 50 failed motors, but analysis of the alerts shows that 45 of the failures clearly have one of three distinct alert patterns leading to failure (for example 12 failures in one pattern, 19 in another pattern and 14 in the third pattern, with the remaining 5 of the 50 belonging to and defining no recognized pattern), the three distinct patterns may be treated as failure modes. The user then must decide in what way the autopsy results match the failed modes, and what investigative and resolution actions can be suggested for the groups based thereon, and stored with the failure mode signature information.[0089]
• For determining precursor diagnostic data in[0090]step1031, the normal data of1020 should be trained and distilled down to a reference library and used offline to generate estimates, residuals and alerts in response to input of the precursor data streams.
• Finally, in[0091]step1042, the diagnostic precursor signatures, the user input regarding failure mode groupings of those signatures and suggested actions, and the empirical model reference library (if an empirical model will be used) is loaded into the onboard memory store of a computing device accompanying each machine of the production run. Thus, a machine can be provided that may have a display of self-diagnostic results using the experience and empirical data of the autopsied failed machines.
• Turning to FIG. 11, it may be desirable or necessary to begin with an empty failure mode database, and an implementation method for this is shown. For example, in the case of an industrial process having sensors, and to be retrofitted with the diagnostic system of the invention, it may not be feasible to cause the process to run to failure multiple times in order to collect precursor data and failure mode information. Alternatively, it may be desirable to initiate real-time monitoring of the process (or machine) with alerts, and add failure modes as they occur. In[0092]step1153, the process is instrumented with sensors, if they are not already in place. Instep1157, sensor data is collected as before, and the process is operated normally. Instep1161, collected data is used to train a reference library for empirical modeling. Instep1165, the resulting reference library is loaded into the monitoring system, and instep1170 the process is monitored in real time. Upon the occurrence of a failure (or a prevented failure handled due to incipient failure alerts) instep1172, the failure (or prevented failure) is autopsied instep1176. Instep1180, collected data (from a historian or other recording feature for operational data archiving) preceding the failure is retrieved and analyzed (as described below) instep1183 to provide precursor residuals, alerts or actuals of the failure mode. The process operator is also prompted for failure mode information, and associated action suggestions to be stored in the failure mode database. Thus, diagnostic monitoring data on failures is collected and stored in the failure mode database, and becomes better and better with continued monitoring of the process.
• In all cases of populating a failure mode database, the user designates the existence, type, and time stamp of a failure. The designation that a process or machine has failed is subject to the criteria of the user in any case. A failure may be deemed to have occurred at a first time for a user having stringent performance requirements, and may be deemed to have occurred at a later second time for a user willing to expend the machine or process machinery. Alternatively, the designation of a failure may also be accomplished using an automated system. For example, a gross threshold applied to the actual sensor signal as is known in the art, may be used to designate the time of a failure. The alerts of the present invention can also be thresholded or compared to some baseline in order to determine a failure. Thus, according to the invention, the failure time stamp is provided by the user, or by a separate automatic system monitoring a parameter against a failure threshold.[0093]
• Three general possibilities may be provided for failure mode signature analysis, e.g., residual snapshot similarity, actual snapshot similarity or alert pattern correlation. The residual snapshot similarity discussed herein provides for a library of prior residual snapshots, i.e., the difference signals obtained preceding identified failure modes which may be compared using the above-described similarity engine and[0094]equation 4 with a current residual snapshot to determine the development of a known failure mode. Using residual diagnosis, the residual snapshots are identified and stored as precursors to known failure modes. Various criteria may be employed for selecting snapshots representative of the failure mode residuals for use in the library and for determining the defining characteristics of the failure modes, and criteria for determination of the failure modes.
• The actual snapshot similarity used for diagnosis is performed in a manner identical with the residual snapshot similarity. Instead of using residual snapshots, actual snapshots are used as precursor data. Then actual snapshots are compared to the failure mode database of precursor actuals and similarities between them indicate incipient failure modes, as described in further detail below.[0095]
• The alert module output will represent decisions for each monitored sensor decomposed input, as to whether the estimate for it is different or the same. These can in turn be used for diagnosis of the state of the process or equipment being monitored. The occurrence of some difference decisions (alerts on a sensor) in conjunction with other sameness decisions (no alerts on a sensor) can be used as an indicator of likely machine or process states. A diagnostic lookup database can be indexed into by means of the alert decisions to diagnose the condition of the process or equipment being monitored with the inventive system. By way of example, if a machine is monitored with seven sensors, and based on previous autopsy experience, a particular failure mode is evidenced by alerts appearing at first on[0096]sensors #1 and #3, compounded after some generally bounded time by alerts appearing onsensor #4 additionally, then the occurrence of this pattern can be matched to the stored pattern and the failure mode identified. One means for matching the failure modes according to developing sensor alert patterns such as these is the use of Bayesian Belief Networks, which are known to those skilled in the art for use in quantifying the propagation of probabilities through a certain chain of events. However, simpler than that, the matching can be done merely by examining how many alerting sensors correspond to sensor alerts in the database, and outputting the best matches as identified failure mode possibilities. According to yet another method for matching the alert pattern to stored alert patterns, the alerts can be treated as a two-dimensional array of pixels, and the pattern analyzed for likeness to stored patterns using character recognition techniques known in the art.
• Turning to FIGS. 12A, 12B and[0097]12C, several methods are shown for automatically selecting how far prior to a user-designated conventional failure point to go back when incorporating failure mode precursor snapshots into a library for purposes of the residual signature approach and the straight-data signature approach. Shown are the plots for a sensor and model estimate (FIG. 12A), residual (12B) and SPRT alerts (12C). The conventional point of failure as it would be understood in the prior art methods is shown in FIGS. 12A and 12B asline1207 and1209 respectively. Accordingly, the number of snapshots prior to a designated failure to include in “training” or distillation to a representative set that will form a failure mode library for either residual snapshot similarity or actual snapshot similarity can be determined as a fixed number selected by the user, either globally for all failures and failure modes, or specific to each autopsied failure. In other words, the user simply dictates based on his knowledge of the sampling rate of the monitoring of the process or machine, that snapshots are included up to, say, 120 prior to the time of failure. This then determines arange1224 of residual snapshots (or actual snapshots) that are to be distilled.
• According to another method of determining the length of[0098]range1224, the location in FIG. 12C ofline1220 is used to determine the snapshot earliest snapshot in theset1224.Line1220 is determined as the earliest consistent SPRT or residual threshold-alerted snapshot, where “consistent” means that at least a selected number of snapshots in a moving window are alerted for at least a selected number of sensors. Thus, for example in a ten-sensor process, if at least two sensors have had at least three alerts in a seven-snapshot moving window, the beginning (or end) of that window demarks the beginning ofrange1224. However, this would extend back only as far prior to the failure snapshot as there are consistent alerts. In other words, if at least the minimum number of alerts is found in a moving window going back to a time T, and before that the minimum number of alerts is not found until the window is approximately around T-50 (snapshots), the range to extend over for failure mode precursor selection extends back to T, not T-50.
• The[0099]range1224 of residual or actual snapshots, each snapshot comprising a residual value or actual value for each sensor, is then distilled to a representative set for the identified failure mode. This distillation process is essentially the same as the training method described in FIGS. 4 and 5 for developing a reference library for empirical modeling. The training process described in the flowchart of FIG. 5 can be used, as can other training methods known in the art or subsequently developed. In addition, if the instance of failure is of a mode already identified and possessing a library of precursor snapshots, then the library can be augmented. One way of augmenting it is to recombine all of the precursor snapshot sets for that failure mode from all documented instances of the failure, and rerun the training process against the combination. Another way is to add the range ofsnapshots1224 to the existing distilled library, and rerun the training process against that combination.
• This precursor data is processed to provide representative data and the associated failure mode, appropriate to the inventive technique chosen from the three prior mentioned techniques for diagnosing failures. This data is added to any existing data on the failure mode, and the system is set back into monitoring mode. Now, the system has more intelligence on precursor data leading up to the particular failure mode.[0100]
• As with commodity machines, the failure mode granularity is entirely user-selectable. The failure modes can be strictly user defined, where the user must do the autopsy and determine cause. The user must furthermore supply a name and/or ID for the failure mode. The software product of the invention preferably provides an empty data structure for storing:[0101]
• a. Failure mode name or ID.[0102]
• b. Description of what is the cause.[0103]
• C. Possible preventive or curative steps to take.[0104]
• d. Possibly can be linked to automated control response.[0105]
• e. Precursor signature data associated with the failure mode.[0106]
• Turning to FIG. 13, the failure mode[0107]precursor reference library1305 that is included in thefailure mode database140 from FIG. 1 can be seen to comprise groups ofsnapshots1312,1315 and1317 that represent the precursor snapshots (either actual or residual) that are associated with the failure modes A, B and C respectively. Asequence1320 of successive current input snapshots (either actual or residual, depending on the implemented embodiment), depicted as vectors with dots as placeholders for parameter values, is fed into a failure mode similarity engine1324 (comprising the failure modesignature recognition module120 from FIG. 1), disposed to calculate snapshot-to-snapshot similarities as described above with respect to the similarity operators used for modeling andequation 4. Preferably, the snapshots ofsequence1320 all have an identical number of parameters, as do the snapshots in thelibrary1305. Unlike the empirical model described above for generating estimates, theengine1324 does not carry outequation 1 above, and thus does not output estimates of any kind, but instead outputs the snapshot similarity scores of each current snapshot as compared to each stored snapshot for at least some and preferably all modes in thelibrary1305.
• The failure[0108]mode similarity engine1324 of FIG. 13 can better be understood in view of FIG. 14, wherein is shown the results for a comparison of asingle snapshot1407 of either actual data from sensors or residual data from the difference of the actual and estimated data for sensors, when compared using the similarity operator to the failure mode precursors in thelibrary1305. Each snapshot-to-snapshot comparison results in a similarity value, which are charted inchart1415.
• In order to determine one or more failure modes to indicate as output of the diagnostic system of the present invention when employing residual similarity or actual signal similarity, one way of selecting such identified or likely failure mode(s) is shown with respect to FIG. 15.[0109]Reference library1305 contains failure mode signature data (either residual snapshots or actual snapshots) forseveral failure modes1312,1315 and1317. A current snapshot is compared using the similarity operation to generate similarity scores for each comparison to reference library snapshots. The failure mode with a single-snapshot similarity1550 that is highest across all such comparisons in the reference library is designated as the indicated failure mode. In another way of selecting the indicated failure mode, as shown in FIG. 16, the average of all the snapshot similarities for all snapshots in a given failure mode is computed, and theaverages1620,1630 and1640 for each failure mode are compared. Thefailure mode1650 with the highest average similarity is designated as the indicated failure mode for the current snapshot. Either way of designating an indicated failure mode for a given current snapshot, as shown in FIGS. 15 and 16, can be combined with a number of alternative ways of selecting the indicated failure mode over successive snapshots. Accordingly, no failure mode may be displayed to the user based on just one snapshot, but a moving window of snapshots over which a count of elected failure modes according to FIGS.15 or16 is maintained can be used to output to the user an indication of an incipient failure, if the count for any given failure mode over the window exceeds a certain number. For example, the method of electing the failure mode with the highest average similarity (FIG. 16) may be used for each current snapshot, and a moving window of twenty (20) snapshots may be used, and a threshold is employed according to which a failure mode must be elected at least 10 times in that window in order for that failure mode to be indicated as an incipient failure mode to the user. Counts are maintained for all failure modes in the system over the twenty snapshot window, and if one of them achieves a count of greater than 10, it is indicated as an incipient failure to the user.
• Other methods of statistically combining the similarities across the set of all stored residual or actual snapshots in the signature library for a given failure mode may be used to get the “average”, such as using only the middle 2 quartiles and averaging them (thus throwing away extreme matches and extreme mismatches); or only using the top quartile; and so on. Regardless of the test used to determine the one or more indicated “winning” failure modes in each snapshot, “bins” accumulate “votes” for indicated failure modes for each current snapshot, accumulating over a moving window of dozens to hundreds of snapshots, as appropriate. A threshold may also be used such that the failure mode “latches” and gets indicated to the human operator as an exception condition.[0110]
• Alternatively, it is possible to not use any such threshold, but to simply indicate for the moving window which failure mode has the highest count of being designated the indicated failure mode snapshot over snapshot. Another useful output of the system that may be displayed to the user is to indicate the counts for each failure mode, and let the user determine from this information when a particular failure mode seems to be dominating. Under normal operation, it is likely all the failure modes will have approximately equal counts over the window, with some amount of noise. But as a failure mode is properly recognized, the count for that failure mode should rise, and for the other failure modes drop, providing a metric for the user to gauge how likely each failure mode is compared to the others.[0111]
• Turning to FIG. 17, several methods for designating the indicated failure mode, if any, are shown with respect to using alert patterns. Alert test[0112]927 (from FIG. 9) generates alerts onsignal lines1704, at each ofsuccessive snapshots1708, as indicated by the asterisks. According to one method, thepattern1715 of alerts at any given snapshot can be matched to the patterns stored for various failure modes, to determine whether or not a failure mode is indicated. According to another method, thecumulative pattern1720 of alerts can be matched against stored patterns, where alert accumulation occurs over a window of a selected number of snapshots. Yet another way is to match thesequence1730 in which sensors alert to sequences in the database, such that alerts appearing first onsensor1, thensensor4, and thensensor9 would be different from first appearing onsensor4, and thensensors1 and9. Finally, therate1740 of sensor alerting can be matched to stored rates. A combination of these can also be used to provide more sophisticated differentiation of failure mode signatures.
• The pattern match for any of the above alert patterns can be selected from a number of techniques. For example, a complete match may be required, such that a match is not indicated unless each and every alert in the stored pattern is also found in the instant pattern, and no extraneous alerts are found in the instant pattern. Alternatively, a substantial match can be employed, such that at least, say, 75% of the sensors showing alerts in the stored pattern are also found alerting in the instant pattern, and no more than 10% of the instant alerts are not found in the stored pattern. The exact thresholds for matching and extraneous alerts can be set globally, or can be set for each stored pattern, such that one failure mode may tolerate just 65% matching and no more than 10% extraneous alerts, while a second failure mode may be indicated when at least 80% of the stored alerts are matched, and no more than 5% extraneous alerts occurring in the instant pattern are not in the stored pattern. These limits may be set empirically, as is necessary to sufficiently differentiate the failure modes that are desirably recognized, and with sufficient forewarning to provide benefit.[0113]
• According to the invention, it is also permissible to indicate more than one potential failure mode, if pattern matching has these results. Techniques are known in the art for matching patterns and providing probabilities of the likelihood of the match, and any and all of these may be employed within the scope of the present invention.[0114]
• FIG. 18 shows a[0115]physical embodiment1820 for any of the inventive approaches to diagnosis disclosed herein. A process ormachine1822 provides sensor output to aninput bus1824. For example, the process might be a process control system at a chemical processing plant, and the bus is the FieldBus-type architecture commonly used in industry. Aprocessor1826 is disposed to calculate the model estimates of the parameters in response to the input of the actual parameters frombus1824, and further to compare the estimates to the actual sensor values and compute alert tests.Processor1826 is further disposed to execute failure signature recognition, when coupled with amemory1828 for storing program code and loaded with model and signature data. The processor can output control commands back to the process control system for corrective action in the event of a diagnosis of an impending failure. Also, the processor can output the resulting diagnosis and accompanying data to adisplay1832, or can also optionally send it via atransmitter1830 to a remote location; the transmitter could be a web-connected device, or a wireless device, by way of example. The receiver (not shown) could be a pager, another data processing system at a remote location, and the like.
• Generally, the failure mode data store can be in any conventional memory device, such as a hard disk drive, nonvolatile or volatile memory, or on-chip memory. The data store for the empirical modeling data that is used to generate the estimates of parameters in response to actual parameter values can be separate from or the same as the data store which contains failure mode signature information. Further, failure mode action suggestions can also be stored either together with or separately from the other aforementioned data. Such may be the case where the present invention comprises combing a failure mode signature recognition system with an existing maintenance operations resource planning system that automatically generates maintenance requests and schedules them. The computational programs for performing similarity-based residual or actual sensor snapshot failure mode signature recognition; alert pattern-based failure mode signature recognition; process modeling and sensor value estimation; residual generation from actual and estimated values; and alert testing can be carried out on one processor, or distributed as separate tasks across multiple processors that are in synchronous or asynchronous communications with one another. In this way, it is entirely within the inventive scope for the diagnostic system of the present invention to be carried out using a single microprocessor on-board a monitored machine, or using a number of separately located computers communicating over the internet and possibly remotely located from the monitored process or machine. The computational program that comprises the similarity engine that generates estimates in response to live data can also be the same programmed similarity engine that generates similarity scores for use in matching a residual snapshot or actual snapshot to stored snapshots associated with failure modes.[0116]
• It will be appreciated by those skilled in the art, that modifications to the foregoing preferred embodiments may be made in various aspects. Other variations clearly would also work, and are within the scope and spirit of the invention. The present invention is set forth with particularity in the appended claims. It is deemed that the spirit and scope of that invention encompasses such modifications and alterations to the preferred embodiment as would be apparent to one of ordinary skill in the art and familiar with the teachings of the present application.[0117]

Claims (49)

What is claimed is:

1. A monitoring apparatus for diagnosing faults in a system, comprising:

a reference data store containing failure mode identification data and associated parameter data indicative of behavior of said system in the failure mode; and

a similarity engine responsive to monitored parameter data indicative of monitored behavior of said system, for generating at least one similarity value for a comparison of the monitored parameter data to said failure mode associated parameter data, as a diagnostic indication of said failure mode.

2. An apparatus according toclaim 1, wherein said parameter data is residual data.

3. An apparatus according toclaim 2, further comprising:

a model for generating estimates of operational data in response to receiving operational data from said system; and

means for differencing the estimates and the received operational data to generate the residual data.

4. An apparatus according toclaim 1, further comprising failure identification means responsive to similarity values from the similarity engine for determining an indicated failure mode.

5. An apparatus according toclaim 4, wherein said failure mode identification means compares similarity values for a plurality of failure modes in said data store, and identifies at least the failure mode with the highest similarity as an indicated failure mode of said system.

6. An apparatus according toclaim 4, wherein said failure mode identification means compares similarity values for a plurality of failure modes in said data store, and identifies at least the failure mode with the highest average similarity as an indicated failure mode of said system.

7. An apparatus according toclaim 4, wherein said failure mode identification means compares similarity values for a plurality of failure modes in said data store, and identifies as an indicated failure mode of said system at least the failure mode with at least a selected number of highest similarities over a window of successive comparisons.

8. An apparatus according toclaim 4, wherein said failure mode identification means compares similarity values for a plurality of failure modes in said data store, and identifies as an indicated failure mode of said system at least the failure mode with at least a selected number of highest average similarities over a window of successive comparisons.

9. A method for diagnosing faults in a monitored system, comprising the steps of:

acquiring monitored parameter data indicative of monitored behavior of said system; and

comparing for similarity the monitored parameter data to reference parameter data associated with a failure mode to generate a similarity value as a diagnostic indication of said failure mode.

10. A method according toclaim 9, wherein said parameter data is residual data.

11. A method according toclaim 10, further comprising the steps of:

generating estimates of operational data in response to acquiring operational data from said system; and

differencing the estimates and the received operational data to generate the residual data.

12. A method according toclaim 9, further comprising the step of determining an indicated failure mode based on similarity values resulting from the similarity comparisons.

13. A method according toclaim 12, wherein said determining step comprises comparing the similarity values for a plurality of failure modes, and identifying at least the failure mode with the highest similarity as an indicated failure mode of said system.

14. A method according toclaim 12, wherein said determining step comprises comparing the similarity values for a plurality of failure modes, and identifying at least the failure mode with the highest average similarity as an indicated failure mode of said system.

15. A method according toclaim 12, wherein said determining step comprises comparing the similarity values for a plurality of failure modes, and identifying as an indicated failure mode of said system at least the failure mode with at least a selected number of highest similarities over a window of successive comparisons.

16. A method according toclaim 12, wherein said determining step comprises comparing the similarity values for a plurality of failure modes, and identifying as an indicated failure mode of said system at least the failure mode with at least a selected number of highest average similarities over a window of successive comparisons.

17. A monitoring apparatus for diagnosing faults in a system, comprising:

a similarity engine responsive to monitored parameter data from said system for generating estimates of the monitored parameter data;

an alert module disposed to produce parameter alerts in response to a comparison of said estimates to said monitored data; and

means for identifying an impending failure mode in said system by matching said parameter alerts with at least one reference alert pattern associated with said failure mode.

18. An apparatus according toclaim 17, wherein said alert module performs a sequential probability ratio test on at least one parameter.

19. An apparatus according toclaim 17, wherein said alert module produces an alert when the difference of an estimate and corresponding monitored data exceeds a selected threshold.

20. A method for diagnosing faults in a monitored system, comprising the steps of:

comparing for similarity monitored parameter data from said system to reference parameter data characteristic of known behavior of said system;

generating estimates of the monitored parameter data based on the similarity comparison;

generating alerts in response to a comparison of said estimates to said monitored data; and

identifying an impending failure mode in said system by matching said parameter alerts with at least one reference alert pattern associated with said failure mode.

21. A method according toclaim 20, wherein said alert generating step comprises performing a sequential probability ratio test on at least one parameter.

22. A method according toclaim 20, wherein said alert generating step comprises producing an alert when the difference of an estimate and corresponding monitored data exceeds a selected threshold.

23. A monitoring apparatus for diagnosing faults in a system, comprising:

a similarity engine responsive to monitored parameter data from said system for generating estimates of the monitored parameter data;

means for differencing the estimates and the monitored data to generate residual data; and

means for recognizing an impending failure mode by comparing for similarity said residual data to reference residual data associated with the failure mode.

24. An apparatus according toclaim 23, further comprising means for communicating remedial control commands to a control program for said system.

25. An apparatus according toclaim 23, further comprising a control module for operating said system, responsive to recognition of an impending failure of said system from said recognizing means for remedially controlling said system.

26. An apparatus according toclaim 23, further comprising a profile data store for storing and providing reference parameter data characteristic of known behavior of said system, to said similarity engine for generation of the estimates.

27. An apparatus according toclaim 23, further comprising a failure mode data store for storing said reference residual data and associated failure mode identification data.

28. An apparatus according toclaim 27 wherein said failure mode data store also stores information about remedial steps specific to a failure mode.

29. A method for diagnosing faults in a monitored system, comprising the steps of:

comparing for similarity monitored parameter data from said system to reference parameter data characteristic of known behavior of said system;

generating estimates of the monitored parameter data based on the similarity comparison;

differencing the estimates and the monitored data to generate residual data; and

comparing for similarity the residual data to reference residual data associated with a failure mode, as a diagnostic indication of said failure mode.

30. A method according toclaim 29, further comprising the step of communicating a remedial control command to a control program for said system.

31. A method according toclaim 29, further comprising the steps of:

recognizing an impending failure mode based on the residual similarity comparison step; and

controlling remedially said system responsive to recognition of an impending failure of said system.

32. A computer program product for diagnosing faults in a monitored system, comprising:

a residual testing program module for generating alerts in response to residual signals characterizing behavior of said system; and

a fault pattern detection program module disposed to receive said alerts and indicate a fault in said system upon matching the alerts to a reference pattern associated with said fault.

33. A program product according toclaim 32, wherein the residual testing program module performs a sequential probability ratio test on said residuals.

34. A program product according toclaim 32, wherein the residual testing program module compares a residual to a threshold to generate an alert when the residual exceeds the threshold.

35. A program product according toclaim 32, further comprising an empirical model program module disposed to generate estimates of parameter signals from said system, and difference the parameter signals and the estimates to generate the residual signals.

36. A program product according toclaim 35, wherein said empirical model program module employs a similarity operator to generate said estimates.

37. A method for forming a diagnostic monitoring system for a machine, comprising the steps of:

instrumenting a plurality of like machines with sensors;

operating said plurality of machines and collecting operational sensor data characterizing operation;

autopsying each of said plurality of machines upon failure to determine a failure mode, and storing the failure modes for use in the diagnostic monitoring system;

selecting a failure precursor portion of said operational sensor data collected for a selected interval prior to failure of each machine; and

distilling each said failure precursor data portion into at least one failure signature data set and associating each set with a corresponding failure mode.

38. A method according toclaim 37 wherein said distilling step comprises selecting time-correlated snapshots of operational sensor data.

39. A method according toclaim 37 wherein said distilling step comprises generating time-correlated snapshots of estimated sensor data corresponding to at least a selected failure precursor data portion, differencing the estimates with the selected failure precursor data to form residual snapshots, and selecting at least one residual snapshot to be associated with the failure mode.

40. A method according toclaim 39 wherein the estimated sensor data is generated by an empirical model based on the collected operational sensor data.

41. A diagnostic monitoring apparatus for a sensor-instrumented system selected from a process and a machine, comprising:

an operational model module for modeling said system and generating estimates for said sensors in response to receiving actual values of said sensors;

a differencing module for generating residual signals from said estimates and said actual values;

a reference library for storing failure modes and associated residual data values; and

a failure mode recognition engine disposed to compare generated residual signals with said residual data values to select and output a recognized failure mode for said system.

42. A diagnostic monitoring apparatus according toclaim 41, wherein said failure mode recognition engine comprises a similarity operation module for generating a similarity score for a comparison of said generated residual signals with said residual data values in said reference library.

43. A diagnostic monitoring apparatus according toclaim 42, wherein said failure mode recognition engine further comprises a failure mode decision module responsive to the similarity scores generated by the similarity operation module to select at least one failure mode to output.

44. A diagnostic monitoring apparatus for a sensor-instrumented system selected from a process and a machine, comprising:

an operational model module for modeling said system and generating estimates for said sensors in response to receiving actual values of said sensors;

a differencing module for generating residual signals from said estimates and said actual values;

a testing module disposed to receive said residual signals and generate alerts in response thereto;

a reference library for storing failure modes and associated alert signatures; and

a failure mode recognition engine disposed to compare generated alerts with said alert signatures to select and output a recognized failure mode for said system.

45. A diagnostic monitoring apparatus for a sensor-instrumented system selected from a process and a machine, comprising:

sensor input means for receiving current sensor data from said system in operation;

a reference library for storing failure modes and associated snapshots of sensor data from said system;

a failure mode recognition engine disposed to compare current sensor data with said snapshots to select and output a recognized failure mode for said system.

46. A diagnostic monitoring apparatus according toclaim 45, wherein said failure mode recognition engine comprises a similarity operation module for generating a similarity score for a comparison of said current sensor data with a snapshot of sensor data stored in said reference library.

47. A diagnostic monitoring apparatus according toclaim 46, wherein said failure mode recognition engine further comprises a failure mode decision module responsive to the similarity scores generated by the similarity operation module to select at least one failure mode to output.

48. A diagnostic monitoring apparatus according toclaim 47, wherein said failure decision module selects the at least one failure mode based on a highest similarity score for any one snapshot in said reference library.

49. A diagnostic monitoring apparatus according toclaim 47, wherein said failure decision module selects the at least one failure mode based on a highest average similarity score across sets of snapshots associated with each failure mode in said reference library.

Images