Adaptive slow feature analysis based industrial process fault detection method

Abstract

The application discloses an industrial process fault detection method based on adaptive slow feature analysis, which comprises the following steps: firstly, collecting industrial process sensor measurement data; then, determining process state variables with high correlation with key variables and forming a required data set; dividing the data set into a training set X train and a test set X test in a 6:4 ratio, and optimizing the training set data to obtain a fault detection model; establishing a calculation formula of a fault detection index; finally, judging whether the working condition of the industrial process changes according to whether the fault detection index exceeds a threshold value, if the working condition changes, resampling and updating the fault detection model, and detecting whether the test set has a fault. The application solves the problems that the existing slow feature analysis fault detection technology cannot adapt to the multi-working condition operation of an industrial system and cannot distinguish the abnormality caused by faults and working condition switching.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault diagnosis, and particularly relates to an industrial process fault detection method based on adaptive slow feature analysis. BACKGROUND

[0002] At present, the automation degree of industrial processes is higher and higher, and the complexity of industrial systems is increasing, which makes it more and more difficult to detect faults in real time. For example, multi-working condition switching and device operating parameter changes over time in industrial processes all bring greater challenges to fault detection technology. The intelligentization of industrial production requires more efficient fault detection technology, which not only can find faults in the operation of production equipment in real time, but also can cope with the difficulties brought by multi-working condition operation.

[0003] Industrial process fault detection methods mainly include principal component analysis, independent component analysis and slow feature analysis. Among them, slow feature analysis can extract slowly changing features in industrial processes from time series, and is very suitable for fault detection and diagnosis of slowly changing industrial processes. However, the traditional slow feature analysis method uses a single detection model, which cannot distinguish whether the abnormality is caused by a fault or a working condition switch when detecting an abnormality, limiting its application in multi-working condition industrial scenarios. Therefore, the application designs a method of adaptive slow feature analysis to realize fault detection of multi-working condition industrial processes. First, for the key variables in the industrial process, the maximum mutual information coefficient method is used to select appropriate variables. Second, a fault detection model is established by using the slow feature analysis method. Finally, whether the detection model needs to be updated is determined according to the designed update strategy. If yes, the resampling method is used for adaptive updating. If no, it is determined whether the system has a fault. In this way, the problem that the traditional slow feature analysis method cannot adapt to the multi-working condition operation of industrial systems and cannot distinguish between faults and working condition switching caused abnormalities is solved. SUMMARY

[0004] The purpose of the application is to provide an industrial process fault detection method based on adaptive slow feature analysis, which solves the problem that the existing slow feature analysis fault detection technology cannot adapt to the multi-working condition operation of industrial systems and cannot distinguish between faults and working condition switching caused abnormalities.

[0005] The technical solution adopted by the application is an industrial process fault detection method based on adaptive slow feature analysis, which is implemented according to the following steps:

[0006] Step 1, collecting industrial process sensor measurement data;

[0007] Step 2, determining process state variables with high correlation with key variables and forming a required data set;

[0008] Step 3, dividing the data set obtained in step 2 into a training set and a test set in a ratio of 6:4 and test set , obtaining a fault detection model by optimizing the training set data;

[0009] Step 4, establishing a calculation formula of the fault detection index;

[0010] Step 5, judging whether the working condition of the industrial process changes according to whether the fault detection index exceeds a threshold value, if the working condition changes, resampling and updating the fault detection model;

[0011] Step 6, detecting whether the test set obtained in step 3 has a fault.

[0012] The application is characterized in that,

[0013] Step 1 is implemented according to the following steps:

[0014] obtaining industrial process monitoring data by using sensors, N which are divided into process state variables m for variable number, a key variable representing product quality, N and variable length.

[0015] Step 2 is implemented according to the following steps:

[0016] Step 2.1, taking the first process state variable i as the horizontal coordinate, taking the key variable as the vertical coordinate, drawing a scatter plot, and giving grid parameters and according to the specific range size of and a to perform b row a column grid on the scatter plot; b

[0017] Step 2.2, calculating the maximum mutual information value of the key variable and the process state variable wherein, and are respectively a point of the key variable and the variable to be tested, is the joint probability of and , , are respectively the conditional probability of y and x , and the maximum mutual information value is normalized to obtain ; ​​

[0018] Step 2.3: By changing the meshing parameters in Step 2.1 a and b To obtain different a and b The mutual information values ​​under the given values ​​are selected, and the maximum value is taken as the maximum mutual information coefficient (MIC), denoted as: :

[0019] (1)

[0020] Among them, the meshing parameters a and b The selection range is subject to the threshold B limit;

[0021] Step 2.4: Calculate the remaining process state variables and key variables respectively. Y The maximum mutual information value is calculated using the same method as steps 2.1, 2.2, and 2.3, yielding the maximum mutual information coefficients between all process variables and key variables. , , ..., ;

[0022] Step 2.5: Sort the largest mutual information coefficients from largest to smallest, and select the process state variables that are more correlated with the key variables as the training model. The number of selected variables is 60% of the total number of variables.

[0023] Step 3 is implemented in the following steps:

[0024] Step 3.1: Process the training set Normalization yields , It has the property of having zero mean and its covariance matrix being an identity matrix;

[0025] Step 3.2, for Perform singular value decomposition, i.e. ,in, for The diagonal matrix obtained by decomposition for The characteristic matrix obtained by decomposition for The transpose of the matrix is ​​used to determine the whitening data matrix. Data after whitening The covariance matrix is ​​the identity matrix, that is, it has the property... ;

[0026] Step 3.3, for Differentiation yields ,right Perform singular value decomposition to obtain , for The diagonal matrix obtained by decomposition for The characteristic matrix obtained by decomposition for The transpose, according to , , The transformation matrix is ​​calculated. Transformation matrix That is, the detection model that needs to be obtained.

[0027] Step 4 is implemented in the following steps:

[0028] Step 4.1: Use the transformation matrix obtained in Step 3. Calculate the detection indicators. For Perform matrix decomposition , Transformation matrix transpose, for The diagonal matrix obtained by decomposition for The characteristic matrix obtained by decomposition for transpose;

[0029] Step 4.2: Calculate the squared prediction error SPE and Hotelling statistics :

[0030] (2)

[0031] (3)

[0032] in, I As a unit array, For the test set Data at a certain moment, for transpose, SPE and These are the detection indicators we hope to obtain.

[0033] Step 5 is implemented in the following steps:

[0034] Step 5.1: Using the data in the test set, calculate the current data. Value and Value, that is and Set detection indicators The threshold is Detection indicators the change amount of the threshold value of the detection index threshold value of the detection index the change amount of the threshold value of the determination , , , If any of the above conditions is met, the detection model is updated, i.e., step 5.2; if not, it is determined whether a fault has occurred, i.e., step 6.

[0035] Step 5.2, resample the obtained in step 3, calculate the distance between the abnormal points of the test set and the th sample of the training set i , , wherein, is the th column vector in the test set sample j , is the th column vector in the training set sample j , the weight i of the th sample in the training set is calculated, and the sampling probability i of the th sample in the training set is calculated according to the weight;

[0036] Step 5.3, calculate the sampling probability of all training set samples obtained in step 3 according to the method of step 5.2, resample the obtained in step 3, obtain a new training set , and input the into the detection model to retrain the detection model, i.e., update the fault detection model.

[0037] Step 6 is implemented according to the following steps:

[0038] Step 6.1, apply the kernel density estimation method commonly used in fault detection technology to calculate the control limits SPE and of the detection index and the detection index, respectively, which are referred to as fault threshold values.

[0039] Step 6.2, determine whether the detection index exceeds the fault threshold value. If it exceeds the fault threshold value, it is determined that a fault has occurred. If it does not exceed the fault threshold value, it is determined that the system is normal.

[0040] The beneficial effect of the present application is that the industrial process fault detection method of adaptive slow feature analysis realizes real-time detection of the fault of the multi-condition industrial process. The maximum mutual information coefficient method effectively selects the process state variable, the slow feature analysis method well extracts the features in the monitoring data, and the adaptive updating strategy and the resampling scheme guarantee the high adaptability of the proposed method to the multi-condition operation of the industrial system, that is, in the case of condition change, the method can distinguish whether the process state change is caused by fault or condition switching. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is the overall flow chart of the industrial process fault detection method of adaptive slow feature analysis;

[0042] Figure 2 is the detection effect diagram of the adaptive slow feature analysis fault detection method applied to the Tennessee Eastman (TE) simulation platform;

[0043] Figure 3 is the detection effect diagram of the traditional slow feature analysis fault detection method applied to the Tennessee Eastman (TE) simulation platform. DETAILED DESCRIPTION

[0044] The present application will be described in detail below in combination with the drawings and specific embodiments.

[0045] The industrial process fault detection method of adaptive slow feature analysis of the present application has a flow chart as shown in Figure 1 , and is specifically implemented according to the following steps:

[0046] Step 1, collecting industrial process sensor measurement data;

[0047] Step 1 is specifically implemented according to the following steps:

[0048] The industrial process monitoring data obtained by using the sensor includes, for example, reactor pressure, cooling water outlet temperature, feed flow, etc. The data is divided into process state variables N m The number of variables is the key variable representing the product quality, N and the variable length.

[0049] Step 2, determining the process state variable with high correlation degree with the key variable and forming the required data set;

[0050] Step 2 is specifically implemented according to the following steps:

[0051] Step 2.1, taking the first process state variable i ​​For the abscissa, the key variable For the ordinate, a scatter plot is drawn according to and The grid parameters are given according to the specific range size of a and b The scatter plot is a row b column gridded;

[0052] Step 2.2, the maximum mutual information value of the key variable and the process state variable is calculated wherein, and are the key variable and the variable to be measured at a certain point, is the joint probability of and , , are the conditional probabilities of y and x , respectively, and the maximum mutual information value is normalized to obtain ;

[0053] Step 2.3, in order to more accurately evaluate the correlation between the key variable and the process state variable , the mutual information value in step 2.2 is optimized. Specifically, by changing the gridding parameters a and b in step 2.1, the mutual information values under different a and b values are obtained, and the maximum value is selected as the maximum mutual information coefficient MIC value (Maximal information coefficient), denoted as :

[0054] (1)

[0055] wherein the selection range of the gridding parameters a and b is limited by the threshold value B , and the size of B is empirically set to the 0.6th power of the data amount.

[0056] Step 2.4, the maximum mutual information values of the remaining process state variables and the key variable Y are calculated respectively, and the calculation method is the same as steps 2.1, 2.2 and 2.3, to obtain the maximum mutual information coefficients of all process variables and the key variable, i.e. , , …, ;

[0057] Step 2.5, sort the maximum mutual information coefficients from large to small, select the process state variables with larger correlation with the key variables as the training model, and the number of selected variables is 60% of the number of variables.

[0058] Step 3, divide the data set obtained in step 2 into a training set and a test set in a ratio of 6:4;

[0059] Step 3 is implemented according to the following steps:

[0060] Step 3.1, normalize the training set to obtain , with zero mean and covariance matrix as the identity matrix;

[0061] Step 3.2, singular value decomposition of , that is, , where is the diagonal matrix obtained by decomposition, is the feature matrix obtained by decomposition, is the transpose of , and the whitened data matrix is determined by using the matrix obtained by decomposition , and the covariance matrix of the whitened data is the identity matrix, that is, it has the property ;

[0062] Step 3.3, derive to obtain , and singular value decomposition of to obtain , is the diagonal matrix obtained by decomposition, is the feature matrix obtained by decomposition, is the transpose of , and the transformation matrix is calculated according to , , , , and the transformation matrix is the detection model to be obtained.

[0063] Step 4, establish the calculation formula of the fault detection index;

[0064] Step 4 is implemented according to the following steps: ​​

[0065] Step 4.1, using the transformation matrix obtained in step 3 Calculate the detection index. For Perform matrix decomposition , is the transpose of the transformation matrix , is the diagonal matrix obtained by decomposition, is the feature matrix obtained by decomposition, is the transpose of

[0066] Step 4.2, calculate the squared prediction error SPE (Squared prediction error) and Hotelling's T2 (Hotelling's T2):

[0067] (2)

[0068] (3)

[0069] wherein, I is the identity matrix, is the data of the test set at a certain time, is the transpose of SPE and is the detection index that we hope to obtain.

[0070] Step 5, according to whether the fault detection index exceeds the threshold value, judge whether the working condition of the industrial process changes, if it changes, resample and update the fault detection model;

[0071] Step 5 is implemented according to the following steps:

[0072] Step 5.1, using the data in the test set, calculate the value and value of the current data, i.e. and , set the threshold value of the detection index to , the threshold value of the change amount of the detection index to , the threshold value of the detection index to , the threshold value of the change amount of the detection index to , judge , 、 、 If any one of the above conditions is met, the detection model is updated, i.e., step 5.2; if not, it is determined whether a fault has occurred, i.e., step 6.

[0073] Step 5.2, resampling the training set obtained in step 3, calculating the distance between the abnormal point of the test set and the first sample of the training set , , , i , , , , , j , , , j , i , , i , ,

[0074] Step 5.3, calculating the sampling probability of all training set samples obtained in step 3 according to the method of step 5.2, resampling the training set obtained in step 3, obtaining a new training set , , , ,

[0075] Step 6, detecting whether a fault has occurred in the test set obtained in step 3.

[0076] Step 6 is implemented according to the following steps:

[0077] Step 6.1, applying the kernel density estimation method commonly used in fault detection technology to calculate the control limits of the detection indexes SPE , , , , ,

[0078] Step 6.2, determining whether the detection index exceeds the fault threshold, if it exceeds the fault threshold, it is determined that a fault has occurred, if it does not exceed, it is determined that the system is normal.

[0079] Embodiment

[0080] In order to better verify the effect of the present application in industrial process fault detection, the data set obtained from the Tennessee-Eastman (TE) simulation platform is selected to verify the effect of the present application. The entire experimental process is shown in Figure 1 .

[0081] TE dataset is a dataset for fault detection and diagnosis. It consists of training set and test set, a total of 22 samples, each sample contains 52 observation variables. TE process includes reactor, condenser, compressor, separator and stripper five main units, there are 53 variables in TE process, of which 12 are operation variables and 41 are measured variables. The data in the TE dataset is composed of 22 different simulation run data, of which d00.dat and d00_te.dat are samples under normal working condition, and the remaining samples are used to train and test error conditions. TE dataset is widely used in fault detection, and can be used to test and compare the performance of different fault detection algorithms.

[0082] The experimental parameters are set to mode 1: G / H ratio 50:50; mode 2: G / H ratio 10:90; fault 1: material ABC composition change (random); fault 2: reactor kinetic property change (random); fault occurs at 500-1000, 1500-2000 sampling points; condition 1 is 0-500 sampling points, condition 2 is 1000-1500 sampling points; it can be understood that the 0-500 sampling points are in normal state, one kind of gap fault occurs at 500-1000 sampling points, which is adjusted to normal state by the system, and is switched to the normal state of condition 2 at 1000-1500 sampling points, and then a fault occurs.

[0083] Figure 2 The results obtained by the present application can be seen that the detection model is retrained at the 501, 1001, 1501 sampling points, respectively, and the faults and normal states are detected, the detection indexes are higher than the control limit when the faults occur, and the faults are judged, and the detection indexes are lower than the control limit when the normal states, and no alarm is given. The present application can detect faults in real time, and will not alarm the normal state as a fault.

[0084] Figure 3 The results obtained by the traditional slow feature analysis can be seen from the figure that the traditional method judges the time of normal state as fault, and cannot distinguish the normal state and fault state. It is calculated that 84.5% of the time is judged as fault by the detection model. It can be shown that the traditional method is difficult to correctly detect faults under the condition of multiple working conditions of industrial system and monitoring data changing with time.

Claims

1. An industrial process fault detection method based on adaptive slow feature analysis, characterized in that, The specific steps are as follows: Step 1: Collect sensor measurement data for industrial processes; Step 2: Identify process state variables that are highly correlated with key variables and construct the required dataset; Step 3: Divide the dataset obtained in Step 2 into a training set in a 6:4 ratio. and test set The fault detection model is obtained by optimizing the training set data; Step 3 is implemented in the following steps: Step 3.1: Process the training set Normalization yields , It has the property of having zero mean and its covariance matrix being an identity matrix; Step 3.2, for Perform singular value decomposition, i.e. ,in, for The diagonal matrix obtained by decomposition for The characteristic matrix obtained by decomposition for The transpose of the matrix is ​​used to determine the whitening data matrix. Data after whitening The covariance matrix is ​​the identity matrix, that is, it has the property... ; Step 3.3, for Differentiation yields ,right Perform singular value decomposition to obtain , for The diagonal matrix obtained by decomposition for The characteristic matrix obtained by decomposition for The transpose, according to , , The transformation matrix is ​​calculated. Transformation matrix This refers to the detection model that needs to be obtained; Step 4: Establish the calculation formula for fault detection indicators; Step 5: Determine whether the operating conditions of the industrial process have changed based on whether the fault detection indicators exceed the threshold. If they have changed, resample and update the fault detection model. Step 6: Check if the test set obtained in Step 3 is faulty.

2. The industrial process fault detection method based on adaptive slow feature analysis according to claim 1, characterized in that, Step 1 is implemented in the following steps: Using sensors to collect data from industrial processes N Group monitoring data, divided into process state variables m For the number of variables, Key variables characterizing product quality N The length of the variable.

3. The industrial process fault detection method based on adaptive slow feature analysis according to claim 2, characterized in that, Step 2 is implemented in the following steps: Step 2.1, with the first i process state variables Using the horizontal axis as the key variable Plot a scatter plot with the vertical axis as the ordinate, based on... and The specific range size is given by the meshing parameters. a and b Perform scatter plot a OK b Column gridding; Step 2.2: Calculate key variables and process state variables Maximum mutual information value ,in, and These are a point on the key variable and the variable to be tested, respectively. yes and The joint probability, , They are y and x The conditional probability is obtained by normalizing the maximum mutual information value. ; Step 2.3: By changing the meshing parameters in Step 2.1 a and b To obtain different a and b The mutual information values ​​under the given values ​​are selected, and the maximum value is taken as the maximum mutual information coefficient (MIC), denoted as: : (1) Among them, the meshing parameters a and b The selection range is subject to the threshold B limit; Step 2.4: Calculate the remaining process state variables and key variables respectively. Y The maximum mutual information value is calculated using the same method as steps 2.1, 2.2, and 2.3, yielding the maximum mutual information coefficients between all process variables and key variables. , , ..., ; Step 2.5: Sort the largest mutual information coefficients from largest to smallest, and select the process state variables that are more correlated with the key variables as the training model. The number of selected variables is 60% of the total number of variables.

4. The industrial process fault detection method based on adaptive slow feature analysis according to claim 3, characterized in that, Step 4 is implemented in the following steps: Step 4.1: Use the transformation matrix obtained in Step 3. Calculate the detection indicators, for Perform matrix decomposition , Transformation matrix transpose, for The diagonal matrix obtained by decomposition for The characteristic matrix obtained by decomposition for Transpose of; Step 4.2: Calculate the squared prediction error SPE and Hotelling statistics : (2) (3) in, I As a unit array, For the test set Data at a certain moment, for transpose, SPE and These are the detection indicators we hope to obtain.

5. The industrial process fault detection method based on adaptive slow feature analysis according to claim 4, characterized in that, Step 5 is implemented in the following steps: Step 5.1: Using the data in the test set, calculate the current data. Value and Value, that is and Set detection indicators The threshold is Detection indicators Change The threshold is Detection indicators The threshold is Detection indicators Change The threshold is ,judge , , , If any one of these conditions is met, the detection model is updated (step 5.2); otherwise, it is determined whether a fault has occurred (step 6). Step 5.2, regarding the results obtained in Step 3 Perform resampling and calculate the test set. Anomalies and training set The i Distance of each sample , ,in, For test set samples The first in j Column vector, For training set samples The first in j Column vector, calculate the column vector in the training set. i Weights of each sample Then, calculate the weight of the training set. i Sampling probability of each sample ; Step 5.3: Calculate the sampling probability of all training set samples obtained in Step 3 according to the method in Step 5.2, and re-collect the samples obtained in Step 3. To obtain a new training set ,Will The data is input into the detection model, and the detection model is retrained, which is equivalent to updating the fault detection model.

6. The industrial process fault detection method based on adaptive slow feature analysis according to claim 5, characterized in that, Step 6 is implemented in the following steps: Step 6.1: Apply the kernel density estimation method commonly used in fault detection technology to calculate... SPE and Control limits of detection indicators and This is called the fault threshold; Step 6.2: Determine whether the detection index exceeds the fault threshold. If it exceeds the fault threshold, it is determined that a fault has occurred. If it does not exceed the threshold, the system is determined to be normal.

Images