MDUNN-based diesel engine fault diagnosis method under variable loadTechnical Field
The application relates to the technical field of diesel engine fault diagnosis, in particular to a MDUNN-based diesel engine fault diagnosis method under variable load.
Background
The diesel engine is a universal power machine with high heat efficiency, wide power range and good operation maneuverability, and has wide application in various fields of heavy commercial vehicles, engineering and agricultural machines, ships, power stations and the like. Most of the fields served by the diesel engine are important in national economy, if timely fault detection and maintenance cannot be realized, serious accidents caused by fault accumulation can cause economic loss, and the life safety of related personnel is threatened, so that timely and accurate diagnosis of the internal combustion engine is very important.
The fault diagnosis technology based on deep learning has the advantages of high diagnosis precision, high running speed and self-adaptive feature extraction. For a diesel engine, the cylinder cover vibration signal has remarkable characteristics on various faults of the diesel engine, and is suitable for being used as a data base of a deep learning diagnosis model. However, the working environment of the diesel engine is complex, the running load is adjusted in real time along with the actual demand, and the power stability and the stability of the deep learning fault training set and the test set are difficult to ensure. The actual fault data set of the diesel engine and vibration signals acquired by real-time operation and maintenance have serious non-independent same distribution characteristics, so that the fault diagnosis technology of the diesel engine based on deep learning has difficulty in application of real variable load scenes.
In the field of fault diagnosis, a learner develops a deep learning model for solving similar problems, and the model has certain robustness and generalization. Some classical robust deep convolution networks have achieved considerable effects on fault diagnosis of relatively simple mechanical structures such as bearings, gear boxes and wind turbines, however, vibration signals of a cylinder cover of a diesel engine have the following two characteristics of long transmission paths, multiple related parts, multiple transmission paths and transmission of the same fault excitation from multiple paths to the same receiving point. Meanwhile, due to its complex working environment, the diesel engine vibration signal is coupled with strong background noise. Therefore, the diesel engine needs a model with stronger signal analysis capability and better robustness to realize fault diagnosis under variable load, and the method has important significance in constructing a diagnosis model of the diesel engine under variable load.
At present, the following problems exist in the diagnosis of variable load faults of a diesel engine:
1. The first-layer convolution determines the feature entering quantity of CNN, and the common single-path 'fusiform' CNN network architecture with narrow first bit and wide middle cannot fully extract the features.
2. The multi-domain mixing enables the total feature set to be rapidly expanded, the feature intersection set to be rapidly reduced, a non-independent co-distribution scene is formed, and a common deep learning model cannot adapt to a multi-domain environment based on independent co-distribution assumption.
3. The common attention mechanism can lead the network to intensively learn intersection characteristics, and can play an obvious role on simple machinery with highly overlapped different domain characteristic sets, however, the intersection characteristics which are rapidly reduced cannot meet the diagnosis requirement due to the huge difference of signals caused by the change of the working state of the diesel engine.
Because of the above-mentioned shortcomings of the prior art, there is a need for a variable load diesel engine diagnostic method that can adequately capture signal features and form accurate decisions in a multi-domain feature space.
The information disclosed in the background section of the invention is only for enhancement of understanding of the general background of the invention and should not be taken as an acknowledgement or any form of suggestion that this information forms the prior art already known to a person skilled in the art.
Disclosure of Invention
In order to solve the technical problems mentioned in the background art or at least partially solve the technical problems, the application provides a diesel engine fault diagnosis method under MDUNN-based variable load, which comprises the following steps:
S100, based on an acceleration sensor arranged on a cylinder cover of a diesel engine, collecting vibration signals of different faults under different loads;
S200, dividing vibration signals of different faults under different loads according to fixed signal lengths, and establishing fault labels in one-to-one correspondence with fault classes to obtain a fault data set, wherein the fault data set comprises all fault labels;
S300, constructing MDUNN a model, and training MDUNN in stages by using a fault dataset;
S400, inputting MDUNN cylinder cover vibration signals acquired in real time, and outputting the fault type of the diesel engine.
Step S300 specifically includes the following steps:
S310, constructing a multi-depth convolutional neural network;
S320, carrying out consolidation training on the feature matrix weights;
s330, carrying out consolidation training on the depth weight matrix.
Step S310 includes the steps of:
S311, constructing a convolutional neural network, increasing the number of first-layer kernels of the convolutional to 128, keeping the number of subsequent-layer kernels to 64, forming a feature extractor by a first-layer convolution of the 128 kernels and a hidden layer of the 64 kernels, adding a layer of convolution kernel with the size of 1 after the feature extractor, and adjusting the number of channels to a preset category number, wherein the category number is selected according to the fault category in the fault data set, and the convolution output feature can be directly used for decision.
S312, reducing the convolution layers containing 64 convolution kernels in the convolution structure, thereby reducing the abstraction degree, forming convolution paths with different depths, directly taking global average pooling of the results of the convolution paths, namely GAP (GAP) for short, taking cross entropy with fault labels through a SoftMax classifier respectively, then descending in a gradient manner, completing initial training, and discarding the GAP and the SoftMax classifier after the training is finished.
Step S320 includes the steps of:
S321, carrying out backward inner average on matrix dot products of output features and shapes embedded by different depth convolutions, namely adding a layer of selectively connected unbiased full-connection layer, enabling the output to still keep category independence, taking cross entropy between the output of the module and a fault label through a softMax classifier, carrying out gradient descent, discarding the softMax classifier after training is finished, and taking the output as the feature input of the next repeated module.
S322, after the stage training is finished, assigning thresholds for different output features of the convolution, wherein the features higher than the thresholds directly determine the output class advantages, and the features lower than the thresholds play a role in constructing thresholds for other features, and the feature thresholds are calculated by the following formula:
The number of rows of FWM is the category number, and the number of columns n is the feature number in a single channel;
FWMab is the weight of the corresponding position of row b of FWM weight matrix a;
FWMij is the weight of the corresponding position of the i row and j column except the a row and b column of the FWM weight matrix;
Oij is the feature or scoring value of the i row and j column output except the FWM weight matrix a row and b column;
Thresoldab is the feature threshold of the corresponding position of row b of FWM weight matrix a.
Wherein the FWM weight matrix is constrained by a sigmoid function, the formula is as follows:
wherein x is an additionally registered parameter;
Lambda is a constraint coefficient of a value range;
sigma (x) is the weight of the sigmoid constraint.
Step S330 specifically comprises the steps of adding a depth weight matrix in the same form as the feature weight layer after the feature weight layer, summing the depth weight matrix in the same line after the output feature dot product of the feature weight layer, repeating step S320 to obtain a depth weight layer, wherein the output of the depth weight layer is identical to the shape of a fault label, and can be used as the final output adopted in actual fault diagnosis to finish MDUNN training.
The step S400 specifically includes:
s100, acquiring a cylinder cover vibration signal based on an acceleration sensor arranged on a cylinder cover of a diesel engine;
s200, inputting cylinder cover vibration signals MDUNN at intervals of preset time;
S300, MDUNN, carrying out multi-depth feature extraction on the cylinder cover vibration signal, and regulating the dimension reduction in two weight layers to obtain a fault type prediction result;
s400, outputting the fault type of the diesel engine.
The acceleration sensor is a unidirectional acceleration sensor, the unidirectional acceleration sensor is electrically connected with the multichannel data acquisition system, and the multichannel data acquisition system is electrically connected with the multichannel data acquisition system and can acquire cylinder cover vibration signals through the unidirectional acceleration sensor.
Compared with the prior art, the technical scheme provided by the embodiment of the application has the advantages that the diesel engine fault diagnosis method based on MDUNN under variable load enables a network to memorize and process multi-domain characteristics through the multi-depth convolution embedded structure and the consolidation learning method simulating a subconscious mechanism, and avoids network degradation caused by common characteristic reduction of variable load data sets.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description, serve to explain the principles of the invention.
In order to more clearly illustrate the embodiments of the invention or the technical solutions of the prior art, the drawings which are used in the description of the embodiments or the prior art will be briefly described, and it will be obvious to a person skilled in the art that other drawings can be obtained from these drawings without inventive effort.
FIG. 1 is a schematic diagram of steps of a method for diagnosing a diesel engine fault under a plurality of variable loads based on MDUNN according to an embodiment of the present application;
Fig. 2 is a schematic structural diagram of a diesel engine and an acceleration sensor according to a method for diagnosing a diesel engine fault under MDUNN-based variable load according to an embodiment of the present application;
FIG. 3 is a schematic diagram of a multi-depth convolution network structure of a diesel engine fault diagnosis method under MDUNN-based variable load according to an embodiment of the present application;
fig. 4 is a schematic diagram of a MDUNN structure of a diesel engine fault diagnosis method under MDUNN-based variable load according to an embodiment of the present application.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present application more apparent, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application, and it is apparent that the described embodiments are some embodiments of the present application, but not all embodiments of the present application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.
In order to facilitate understanding, the following describes a diesel engine fault diagnosis method under a multiple variable load based on MDUNN in detail, where MDUNN is called Multi-depth Unconscious Neural Network, and refers to a first-layer Multi-convolution kernel Multi-depth subconscious neural network diagnosis model, as shown in fig. 1 and fig. 4, and includes the following steps:
s100, based on an acceleration sensor arranged on a cylinder cover of the diesel engine, collecting vibration signals of different faults under different loads.
As shown in fig. 2, the acceleration sensor is a unidirectional acceleration sensor, and the unidirectional acceleration sensor is electrically connected with the multi-channel data acquisition system, and the multi-channel data acquisition system is electrically connected with the multi-channel data acquisition system and can acquire the cylinder cover vibration signal through the unidirectional acceleration sensor.
S200, dividing vibration signals of different faults under different loads according to fixed signal lengths, and establishing fault labels in one-to-one correspondence with fault classes to obtain a fault data set, wherein the fault data set comprises all fault labels.
S300, constructing MDUNN a model, and training MDUNN in stages by using the fault dataset.
Specifically, as shown in fig. 3, step S300 includes the following steps:
S310, constructing a multi-depth convolutional neural network.
Specifically, step S310 includes the steps of:
S311, constructing a convolutional neural network, increasing the number of first-layer kernels of the convolutional to 128, keeping the number of subsequent-layer kernels to 64, forming a feature extractor by a first-layer convolution of the 128 kernels and a hidden layer of the 64 kernels, adding a layer of convolution kernel with the size of 1 after the feature extractor, and adjusting the number of channels to a preset category number, wherein the category number is selected according to the fault category in the fault data set, so that the convolution output feature can be directly used for decision.
S312, reducing the convolution layers containing 64 convolution kernels in the convolution structure, thereby reducing the abstraction degree, forming convolution passages with different depths, directly taking global average pooling of results of the convolution passages, namely GAP (GAP) for short, taking cross entropy with fault labels through a SoftMax classifier respectively, performing gradient descent, completing initial training, maximizing the characterization capability of each convolution depth, simultaneously avoiding interference of the parameter-containing classifier on non-intersection feature extraction, and discarding GAP and SoftMax classifier after training is finished, wherein the learning effect on the features only depends on the occurrence frequency and the significance of the features.
S320, carrying out consolidation training on the feature matrix weights.
Specifically, step S320 includes the steps of:
S321, carrying out backward internal average on matrix dot products of output characteristics embedded by convolution with different depths and shapes, namely adding a layer of selectively connected unbiased full-connection layer, enabling the output to still keep category independence, taking cross entropy between the output of the module and a fault label through a softMax classifier, carrying out gradient descent, discarding the softMax classifier after training is finished, inputting the output as the characteristic of the next repeated module, learning connection weights under the structure plays a role similar to feedback consolidation, generating weights under full-load domain sample input, and adjusting the expression intensity of convolution neurons, so that the model is suitable for a multi-load average environment.
S322, after the stage training is finished, assigning thresholds for different output features of the convolution, wherein the features higher than the thresholds directly determine the output class advantages, and the features lower than the thresholds play a role in constructing thresholds for other features, and the thresholds of the final results of the feature values are calculated by the following formula:
The number of rows of FWM is the category number, and the number of columns n is the feature number in a single channel;
FWMab is the weight of the corresponding position of row b of FWM weight matrix a;
FWMij is the weight of the corresponding position of the i row and j column except the a row and b column of the FWM weight matrix;
Oij is the characteristic or scoring value of the FWM weight matrix output by the i rows and j columns except the a rows and b columns;
Thresoldab is the feature threshold of the corresponding position of row b of FWM weight matrix a.
Specifically, the FWM weight matrix is constrained by the following formula:
wherein x is an additionally registered parameter;
Lambda is a constraint coefficient of a value range;
sigma (x) is the weight of the sigmoid constraint.
S330, carrying out consolidation training on the depth weight matrix.
Specifically, step S430 includes the steps of:
And adding a depth weight matrix in the same form as the feature weight layer after the feature weight layer, summing the depth weight matrix with the output feature dot product of the feature weight layer in a post-row, repeating the step S320 to obtain the depth weight layer, wherein the output of the depth weight layer is identical to the shape of the fault label, and the depth weight layer can be used as the final output adopted in actual fault diagnosis to finish MDUNN training.
S400, inputting MDUNN cylinder cover vibration signals acquired in real time, and outputting the fault type of the diesel engine.
Specifically, the S400 specifically includes:
and S410, acquiring a cylinder cover vibration signal based on an acceleration sensor arranged on a cylinder cover of the diesel engine.
S420, inputting cylinder cover vibration signals MDUNN at intervals of preset time.
S430, MDUNN, carrying out multi-depth feature extraction on the cylinder cover vibration signal, and regulating the dimension reduction in the two weight layers to obtain a fault type prediction result.
S440, outputting the fault type of the diesel engine.
It should be noted that in this document, relational terms such as "first" and "second" and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises an element.
The foregoing is only a specific embodiment of the invention to enable those skilled in the art to understand or practice the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.