Early warning method for tower flushing fault of primary distillation tower in crude oil distillation processTechnical Field
The invention relates to industrial process fault detection, in particular to a primary distillation tower washing fault early warning method for a crude oil distillation process based on data driving.
Background
Crude oil distillation is a key link of a refining enterprise, and a typical process flow comprises the following steps: primary distillation, atmospheric distillation and vacuum distillation. The primary distillation is the first process in the crude oil distillation process, and whether the primary distillation is stable or not has great influence on subsequent processing.
The tower flushing is a common fault in the primary distillation process, and means that the phenomenon that the gas phase flow is too large and the liquid phase is directly brought into an upper tray in the distillation process can cause the fractionation effect to be poor or the fractionation to be damaged, thereby seriously influencing the subsequent process flow. The tower flushing fault can be caused by high water content, low density, excessive oil inlet amount and the like of crude oil. In recent years, with more and more types of crude oil imported in China, the properties of the crude oil are difficult to grasp, and the failure of the initial distillation tower washing sometimes occurs. Enterprises urgently need a proper method to carry out early warning on tower rushing faults, so that timely operation and remediation are realized, and loss is reduced.
Compared with other fault detection methods based on mechanism models and the like, the fault detection method based on data driving is developed rapidly in recent years. However, the existing method usually analyzes all measurable variables, so that generally only fault detection is possible, and accurate diagnosis of faults is difficult to realize. If effective variables are selected for analyzing the tower-rushing fault, the fault can be accurately diagnosed, and the accuracy of fault early warning can be increased.
In addition, Principal Component Analysis (PCA), a data analysis method that uses only the square prediction error SPE or Hotelling's T, is widely used in data-based fault detection2Single comment of (1)Price index is easy to report by mistake, and the problem of high rate of report by mistake is the most headache of enterprises.
Finally, the existing fault early warning method generally comprises the steps of once modeling and continuous forecasting after data in the production process is collected, and a constant model is adopted, so that false alarm or missed alarm is easy to occur.
Disclosure of Invention
Aiming at the problems, the invention discloses a method for early warning of tower flushing faults of a primary distillation tower in a crude oil distillation process, which can perform rolling modeling aiming at the tower flushing faults of the primary distillation tower and perform fault early warning by adopting comprehensive evaluation indexes.
The method comprises the following steps:
(1) analyzing the technological process and the sensor distribution condition of the primary tower, and determining the initial variable range related to the tower flushing fault;
(2) further screening in the initial variable ranges: if the variable is the controlled variable of the closed-loop control system, the variable is removed, and meanwhile, the manipulated variable of the closed-loop control system is selected to enter a final variable range; if the variable open loop is not controlled, directly selecting to enter a final variable range;
(3) selecting top t aiming at the data to be measured obtained every day1Taking data of days as a training set;
(4) preprocessing a training set, including smoothing filtering and standardization;
(5) performing principal component analysis on the preprocessed sample, and calculating the control limit of the square prediction error SPE and Hotelling's T2A control limit for the statistic;
(6) preprocessing the data to be detected, and calculating SPE and T at each moment2And (5) statistics, and early warning by integrating two indexes.
In the present method, the variables are selected within the following ranges: the system comprises a feed pressure, a feed temperature, a tower top pressure, a tower top temperature, an initial top cold reflux amount, an initial top circulating reflux amount, an initial top oil gas aftercooler outlet temperature, an initial top reflux tank pressure and an initial top reflux tank liquid level.
In this process, t1The value range of (1) is 5-10 days, and confirmed fault data are removed.
Preferably, the smoothing filtering algorithm uses a Savitzky-Golay (S-G) convolution smoothing algorithm.
In the method, a sample matrix X (d) is reconstructed according to the following formula, wherein d is a time lag factor:
in the method, training data is standardized, and the data to be tested is standardized by using the mean value and the variance of a training set.
Has the advantages that:
(1) the method provided by the invention aims at the tower-flushing accident of the primary distillation tower, the relevant variables are selected in a targeted manner for modeling, and the time sequence of industrial process data is considered in the modeling process, so that the prediction result is more accurate.
(2) The method provided by the invention can synthesize SPE and T2The two evaluation indexes reduce false alarm.
(3) The method provided by the invention can automatically update the model every day, roll calculation, effectively reduce false alarm and missed alarm, and has important application value for finding faults in advance, reducing loss and stabilizing production.
Drawings
FIG. 1 is a flow chart of an implementation of a method for early warning of a tower-flushing fault of a primary distillation tower in a crude oil distillation process
FIG. 2 is a schematic diagram of the process and instrument of a primary tower No. 5 of a refinery
FIG. 3 shows the fault warning results of a certain oil refining enterprise from 11 months 15 to 11 months 19
FIG. 4 shows the fault warning results of a certain oil refining enterprise from 11 months 28 days to 12 months 5 days
Detailed description of the preferred embodiment
The following detailed computing process and specific operation flow are given in conjunction with the accompanying drawings and specific examples to further explain the present invention. The present embodiment is implemented on the premise of the technical solution of the present invention, but the scope of the present invention is not limited to the following embodiment.
In this case, a certain oil refining enterprise is taken as an example, the enterprise has a plurality of crude oil distillation units, wherein the crude oil processing capacity of the No. 5 atmospheric and vacuum distillation unit is 800 ten thousand tons/year. Because of various types of crude oil, the enterprise has repeatedly occurred the accident of tower flushing of the primary distillation tower. The effectiveness and implementation process of the method are described by the tower-rushing accident which occurs recently by the device, and then the case of successfully pre-warning and avoiding the tower-rushing by using the method is introduced. First, a tower crash failure occurring around 17 o' clock at 11/18/2018 in the preliminary tower will be described.
The implementation flow of this case is shown in fig. 1, and the specific implementation steps are as follows:
(1) and analyzing the working process of the primary tower and the distribution condition of the sensors, and determining the initial variable range related to the tower flushing fault. The process and the instrument of the enterprise No. 5 primary distillation tower are shown in figure 2. Through process analysis, the variables related to the tower-rushing accident are as follows: the system comprises a feeding pressure (6), a feeding temperature (10), a tower top pressure (5), tower top temperatures (46 and 8), an initial top cooling reflux amount (2), an initial top circulating reflux amount (19), an initial top oil gas after-cooler outlet temperature (47), an initial top reflux tank pressure (4) and an initial top reflux tank liquid level (56).
(2) Further screening in the initial variable ranges: wherein the tower top temperature (8) is controllable through the primary top cooling reflux amount (2); the pressure (4) of the primary top reflux tank is controllable through two valves; the initial top reflux tank liquid level (56) is controllable by the initial top circulating reflux amount. The 3 variables are effectively controlled in a Distributed Control System (DCS), and the variable change range is small, so that the fault early warning is basically not favorable, and the variables can be removed in early warning modeling. In order to stabilize the 3 variables, the corresponding manipulated variables have large variation amplitude, which is very beneficial to timely finding the abnormity. Thus, the variables retained are: the system comprises a feeding pressure (6), a feeding temperature (10), a tower top pressure (5), an initial top cooling reflux amount (2), an initial top circulating reflux amount (19) and an initial top oil gas after-cooler outlet temperature (47).
(3) Selecting front t aiming at the obtained data to be detected1The data of the day is modeled as a training set, where t1Take 5 and the corresponding training data for each day as shown in table 1.
TABLE 1 data to be tested and corresponding training set ranges
| Range of data to be measured | Training set data range |
| … | … |
| 11 month and 15 days | 11 months and 10 days to 11 months and 14 days |
| 11 month and 16 days | 11 months and 11 days to 11 months and 15 days |
| 11 month and 17 days | 12 days in 11 months to 16 days in 11 months |
| 11 month and 18 days | 13 days in 11 months to 17 days in 11 months |
| 11 month and 19 days | 11 months and 14 days to 11 months and 18 days |
| … | … |
(5) And performing S-G smoothing on the training set and the data to be measured, wherein the window is 11, and the order is 3.
(6) Calculating a hysteresis factor, recombining the sample matrix, and calculating the SPE control limit and the T2And (5) controlling the limit.
Taking a model of 11 months and 15 days as an example, data of 11 months and 10 days to 11 months and 14 days are taken to form a sample matrix X. Starting from d ═ 0, recombining a sample matrix X (d) according to the formula (2), standardizing, and then carrying out PCA analysis on the sample matrix, and respectively calculating the number of principal elements, the static correlation coefficient r (d) and the new correlation coefficient rnew(d) The formulas are shown as (3) and (4), and the calculation process is shown as table 2. When d is 2, the new relation coefficient rnew(2) ≦ 0, so the hysteresis factor takes 2, at which time the normalized sample matrix X (2) has also been obtained.
r(d)＝m-k-r(d-1) (3)
Wherein m is the number of variables and k is the number of principal elements.
TABLE 2 hysteresis factor calculation Process
| Value of d | Number of variables | Number of principal elements | Static coefficient of relation | Number of new relations |
| 0 | 6 | 3 | 3 | 3 |
| 1 | 12 | 3 | 9 | 3 |
| 2 | 18 | 3 | 15 | 0 |
Performing PCA analysis on the sample matrix X (2) to obtain a load matrix P and an eigenvalue matrix lambda as follows:
let the detection level α be 0.99, and the SPE control limit calculated from equations (5) and (6) be 3.0679, T2The control limit is 11.3578.
Wherein,λjis the j-th eigenvalue, c, of the covariance matrix of the sample matrix XαIs the standard normal deviation corresponding to the upper limit (1- α). times.100%.
Where n is the number of samples, k is the number of pivot elements, and F (p, n-1, α) is a critical value with an upper limit of α X100% of the F distribution with degrees of freedom k and n-1.
(7) Forming a data matrix X to be measured by using the data of 11 months and 15 days according to d-2testAfter normalizing the values, SPE and T at each time are calculated based on equations (7) and (8)2Statistics are obtained.
SPE＝xT(I-PPT)x (7)
Where P is a load matrix of dimension m x k.
Wherein λ is a diagonal matrix composed of eigenvalues corresponding to the first k principal elements.
TABLE 3 variable controlled event values
| Controlled condition of variable | Value taking |
| Closed loop controlled | 1 |
| Manually or open-loop uncontrolled | 0 |
The results of the rolling model prediction for days 11/month 15 to 11/month 19 are shown in fig. 3. SPE and T2The moment when the statistics are simultaneously over-limited is indicated by the vertical line. It can be seen that the SPE statistics overrun from 11 months, 18 days, 16: 04. Before that, T2The statistic amount is over-limited because the feed temperature and feed pressure of the primary tower are increased, the gas phase in the tower is increased, and the control system automatically increases the cold reflux amount, so T2Increasing; however, at this time, the relationship between the variables is not changed, and the SPE statistics are always normal, and the failure does not necessarily occur. However, as the cold reflux amount approaches the set upper limit, the gas phase cannot be further increased, the final output result is that tower flushing early warning is sent, the time is advanced by nearly 1 hour compared with the actual manual intervention time, and the method is verified to be capable of successfully early warning the tower flushing fault of the primary distillation tower.
FIG. 4 shows the early warning results of the oil refining enterprise No. 5 tower from 2018, 11.28 days in month to 12.5 days in month, from 11.28 days in month, 22:36, SPE and T of the data to be measured2And if the statistic exceeds the limit at the same time, the method sends out early warning through calculation. After receiving the early warning, the enterprise immediately takes manual intervention measures, so that accidents are successfully avoided from happening and expanded, and the occurrence of a tower rushing fault is effectively prevented.