Quality-related fault detection method based on deep decomposition echo state network

Abstract

The invention relates to a quality-related fault detection method based on a deep decomposition echo state network, which comprises the following steps of: dividing into process variable data and quality variable data, and standardizing by using a mean value and a variance of the process variable data and the quality variable data; the method comprises the following steps: constructing a first decomposition network to extract residual information, decomposing a process variable matrix U by using a principal component regression method, sending the decomposed process variable matrix U into a deep echo state network to extract potential residual information, calculating residual vectors # imgabs0 # and # imgabs1 # to construct a second decomposition network to extract dynamic information, and extracting dynamic characteristics X by using an echo state network method. Decomposing the X by using a principal component regression method, further sending the X into a deep echo state network to extract dynamic information in the X, and respectively calculating score matrixes Tre and Tun of the X; fusing and constructing a quality-related statistic # imgabs2 # and a quality-independent statistic # imgabs3 #, and calculating the control limits of the quality-related statistic # imgabs2 # and the quality-independent statistic # imgabs3 And collecting real-time process data, performing standardization processing, calculating a residual vector and a score vector of online data, and judging whether a fault affects a quality variable or not according to a statistical magnitude detection result. According to the fault detection method, the echo state network and the principal component regression method are combined, and the detection performance of industrial process data quality related faults is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of quality-related fault detection, and relates to a quality-related fault detection method based on a Deep Decomposition Echo State Network (DDESN for short). Background Art

[0002] With the development of automation and artificial intelligence technologies, industrial process fault detection technologies have received increasing attention. Complex industrial processes generally contain a large number of measured variables and have characteristics such as non-linearity and dynamics. In modern industrial systems, faults unrelated to quality usually do not cause property losses or pose threats to life safety. If production is immediately stopped for maintenance once a fault is detected, not only will a large amount of manpower and material resources be wasted, but also a series of adverse effects such as equipment hardware aging and shortened service life may occur, which is instead unfavorable for industrial production. Therefore, further determining whether a fault is quality-related based on fault detection has become an essential feature of an industrial process fault detection algorithm with application value.

[0003] Traditional echo state network fault detection models do not distinguish whether a fault is quality-related. At the same time, statistical analysis methods are mostly used in the field of quality-related fault detection. These methods often have problems of insufficient feature extraction and are difficult to effectively process industrial data. Therefore, inventing a quality-related fault detection method based on a deep decomposition echo state network can improve the deficiencies of the existing technology, and can further determine whether a fault affects quality variables on the basis of detecting faults in real time and accurately, reducing fault losses. Summary of the Invention

[0004] Aiming at the problems that the traditional echo state method cannot distinguish quality-related faults and the degree of decomposition of quality-related information is low, the present invention provides a quality-related fault detection method based on a deep decomposition echo state network. This method introduces Principal Component Regression (PCR for short) into the echo state network to construct two deep decomposition networks. One is to decompose process variables using the principal component regression method to construct a first decomposition network to extract residual information, and the other is to use the echo state network to extract dynamic features to construct a second decomposition network to extract dynamic information, and then establish a fusion statistic, thereby improving the quality-related fault detection results of industrial processes.

[0005] To achieve the above object, a quality-related fault detection method based on a deep decomposition echo state network provided by the present invention includes the following steps:

[0006] S1: Collect data and standardize: Obtain industrial process data, divide it into process variable data and quality variable data, and standardize it using their means and variances. After standardization, a process variable matrix and a quality variable matrix are obtained, where n u is the number of samples, m u is the number of process variables, and m y is the number of quality variables;

[0007] S2: Construct the first decomposition network to extract residual information: First, decompose the process variable matrix U using the principal component regression method, dividing it into a quality-related space and a quality-unrelated space Then, send U1 and U2 into a deep echo state network respectively to extract potential residual information, and calculate the residual vectors and

[0008] S3: Construct the second decomposition network to extract dynamic information: Use the echo state network method to extract the dynamic features of the process variable matrix U Decompose X using the principal component regression method, dividing it into a quality-unrelated component and a quality-related component Further send them into a deep echo state network to extract the dynamic information X re and X un , and calculate their score matrices T re and T un respectively;

[0009] S4: Establish a fusion statistic: Comprehensively consider the residual information and dynamic information obtained from the two decomposition paths, and fuse them to construct a quality-related statistic and a quality-unrelated statistic , and calculate their control limits;

[0010] S5: Detect quality-related faults: Collect real-time process data, perform standardization processing, calculate the residual vector and score vector of the online data, and judge whether the fault affects the quality variables according to the statistical detection results.

[0011] 2. The quality-related fault detection method based on a deep decomposition echo state network involved in the present invention, wherein step S1 specifically includes the following steps:

[0012] S101: Collect normal operation data samples of the industrial process, divide them into process variable data U0 and quality variable data Y0, and calculate the means and variances of U0 and Y0 respectively;

[0013] S102: Standardize using the mean and variance, and calculate the process variable matrix U and the quality variable matrix Y according to formulas (1) and (2) as the input and output of the model:

[0014]

[0015] 3. The quality-related fault detection method based on the deep decomposition echo state network involved in the present invention, wherein step S2 specifically includes the following steps:

[0016] S201: Decompose the process variable matrix U using the principal component regression method, and calculate the covariance matrix Perform eigenvalue decomposition on the matrix U0 to obtain the eigenvector matrix where k is the number of principal components in the principal component regression;

[0017] S202: Further calculate the score matrix according to formula (3)

[0018] T = UP (3)

[0019] Calculate the load matrix according to the least squares regression of the quality variable Y and the score matrix T

[0020] Q T = (T T T) -1 T T Y (4)

[0021] Calculate the correlation coefficient matrix between U and the online predicted value of the quality variable according to formula (5) between

[0022] B = PQ T (5)

[0023] S203: Further perform singular value decomposition on the correlation coefficient matrix B through formula (6):

[0024]

[0025] Partition the matrix into the quality-related part and the quality-unrelated part Finally, decompose the process variable U into the quality-related space and the quality-unrelated space

[0026] U1 = UP B1 P B1 T (7)

[0027] U2 = UP B2 P B2 T (8)

[0028] S204: Take the decomposed quality-related process variable U1 as the input and the quality variable matrix Y as the output, train the deep echo state network, and obtain the output matrix of the model quality-related space Take the decomposed quality-unrelated process variable U2 as the input and the quality variable matrix Y as the output, train the deep echo state network, and obtain the output matrix of the model quality-unrelated space

[0029] S205: Predict the output vector according to the quality-related part and the quality-unrelated part prediction output vector The prediction residual vector can be calculated according to formulas (9) and (10):

[0030]

[0031] where, y t is the true quality variable vector

[0032] 3. The quality-related fault detection method based on the deep decomposition echo state network involved in the present invention, wherein step S3 specifically includes the following steps:

[0033] S301: Send the process variable matrix U into the echo state network to extract dynamic features and obtain the dynamic feature matrix Calculate the input matrix X according to formula (11) P :

[0034]

[0035] S302: Use the principal component regression method to perform eigenvalue decomposition on the input matrix X P to obtain the eigenvector matrix Furthermore, the score matrix T can be calculated according to formula (12) as follows:

[0036] T = XP (12)

[0037] Calculate the loading matrix Q and the correlation coefficient matrix B according to formulas (4) to (6), perform singular value decomposition on the B matrix, and divide the matrix into the quality-related part and the quality-unrelated part Then use formulas (13) and (14) to decompose the dynamic feature X into the quality-related dynamic feature matrix and the quality-unrelated dynamic feature matrix

[0038] X1 = XP B1 P B1 T (13)

[0039] X2 = XP B2 P B2 T (14)

[0040] S303: Feed the quality-related dynamic feature matrix X1 into the deep echo state network to extract dynamic features and decompose them according to the method described in S302 to further obtain the deep quality-related dynamic feature matrix X re1 and the coefficient matrix P B11 and the deep quality-unrelated dynamic feature matrix X un1 and the coefficient matrix P B21 , and calculate the score matrices of their respective parts according to formulas (15) and (16) and

[0041] T re1 = X re1 P B11 (15)

[0042] T un1 = X un1 P B21 (16)

[0043] S304: Feed the quality-unrelated dynamic feature matrix X2 into the deep echo state network to extract dynamic features and decompose them according to the method described in S302 to obtain the deep quality-related dynamic feature matrix X re2 and the coefficient matrix P B12 and the deep quality-unrelated dynamic feature matrix X un2 and the coefficient matrix P B22 , and then calculate their respective score matrices and

[0044] 4. The quality-related fault detection method based on the deep decomposition echo state network involved in the present invention, wherein step S4 specifically includes the following steps:

[0045] S401: Use the residual information vector obtained during the decomposition of the process variable matrix U, and further obtain the quality-related statistic SPE re and the quality-unrelated statistic SPE un :

[0046]

[0047] S402: Based on the score matrix of the quality-related principal component part obtained during the decomposition of the dynamic feature X, further construct the corresponding statistics according to formulas (19) and (20) as follows:

[0048]

[0049] wherein, and are the score vectors in the score matrix;

[0050] Based on the score matrix of the quality-unrelated principal component part obtained during the decomposition of the dynamic feature, further construct the corresponding statistics according to formulas (21) and (22) as follows:

[0051]

[0052] wherein, and are the score vectors in the score matrix;

[0053] S403: This method extracts quality-related information from two paths: internal feature decomposition and process variable decomposition. Since there are many information modules generated during the deep decomposition process, there are three quality-related statistics and three quality-unrelated statistics respectively. Therefore, considering all the statistics comprehensively, when any statistic alarms, it is considered that a fault has occurred, and all the statistics are finally fused into the quality-related statistic and the quality-unrelated statistic

[0054]

[0055] where fuse(·) represents taking the detection result of the statistic when any statistic exceeds the control limit.

[0056] 5. The quality-related fault detection method based on the deep decomposition echo state network involved in the present invention, wherein step S5 specifically includes the following steps:

[0057] S501: Collect real-time data in the industrial process, and standardize the data using the mean and variance of the training set to obtain the standardized process variable matrix U new ;

[0058] S502: Input the standardized industrial process data U new into the deep decomposition echo state network for process variable decomposition and dynamic feature decomposition;

[0059] S503: Process the online data using the orthogonal projection matrix obtained by decomposition in the offline stage, and calculate the quality-related statistic and the quality-unrelated statistic

[0060] S504: Judge respectively and whether it exceeds the limit. If neither exceeds the limit, there is no fault.

[0061] S505: If exceeds the limit, it indicates that this fault is a quality-related fault, which affects the quality variable.

[0062] S506: If does not exceed the limit while exceeds the limit, it indicates that this fault is a quality-unrelated fault and cannot affect the change of the quality variable.

[0063] Compared with the prior art, the present invention has the following advantages: (1) The quality-related fault detection method provided by the present invention introduces the principal component regression method into the echo state network model, enabling the traditional echo state network to judge whether a fault affects the quality variable while realizing fault detection, and can meet the requirements of process safety monitoring; (2) Using the echo state network reservoir to extract features, it is applicable to continuous industrial processes with the characteristics of non-linearity, dynamics, and multi-variables. By constructing the first and second deep decomposition networks, the simultaneous extraction of residual information and dynamic information is realized from two decomposition paths, which can more effectively decompose the quality-related information and improve the accuracy of quality-related fault detection; (3) By fusing statistics, the quality-related information obtained by decomposition is comprehensively considered. Its overall process is simple, the principle is reliable, the judgment of whether a fault affects the quality variable is accurate, the real-time safety monitoring effect is good, the applicable range is wide, the logic is strong, and it is environmentally friendly. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is the working principle flow chart of the quality-related fault detection method involved in the present invention.

[0065] Figure 2 is the structural diagram of the Tennessee-Eastman process of the chemical model involved in the present invention.

[0066] Figure 3 is the quality variable diagram of TE process fault 5 involved in the embodiment of the present invention.

[0067] Figure 4 a is the schematic diagram of the monitoring result of the quality-related statistic of TE process fault 5 by using the principal component regression method in the embodiment of the present invention.

[0068] Figure 4 b is the schematic diagram of the monitoring result of the quality-related statistic of TE process fault 5 by using the deep decomposition echo state network method described in the present invention in the embodiment of the present invention.

[0069] Figure 5This is a schematic diagram of the monitoring results of the quality-independent statistic for the fault 5 of the TE process using the deep decomposition echo state network method described in the embodiments of the present invention. Detailed implementation manners

[0070] Next, the present invention will be specifically described through exemplary implementation manners. However, it should be understood that without further narration, the elements, structures, and features in one implementation manner can also be beneficially combined into other implementation manners.

[0071] Refer to Figure 1 , the present invention discloses a quality-related fault detection method based on a deep decomposition echo state network, which includes the following steps:

[0072] S1: Collect data and standardize it: Obtain industrial process data, divide it into process variable data and quality variable data, and standardize it using its mean and variance. After the standardization process, a process variable matrix and a quality variable matrix are obtained, where n u is the number of samples, m u is the number of process variables, and m y is the number of quality variables;

[0073] S2: Construct a first decomposition network to extract residual information: First, decompose the process variable matrix U using the principal component regression method, and divide it into a quality-related space and a quality-independent space Then, send U1 and U2 into a deep echo state network respectively to extract potential residual information, and calculate the residual vectors and

[0074] S3: Construct a second decomposition network to extract dynamic information: Use the echo state network method to extract the dynamic characteristics of the process variable matrix U Decompose X using the principal component regression method, and divide it into a quality-independent component and a quality-related component Further send them into a deep echo state network to extract the dynamic information X re and X un , and calculate their score matrices T re and T un respectively;

[0075] S4: Establish a fusion statistic: Comprehensively consider the residual information and dynamic information obtained from the two decomposition paths, and fuse and construct a quality-related statistic and a quality-independent statistic and calculate their control limits;

[0076] S5: Detect quality-related faults: Collect real-time process data, perform standardization processing, calculate the residual vector and score vector of the online data, and determine whether the faults affect the quality variables according to the statistical test results.

[0077] 2. The quality-related fault detection method based on the deep decomposition echo state network involved in this embodiment, wherein step S1 specifically includes the following steps:

[0078] S101: Collect normal operation data samples of the industrial process, divide them into process variable data U0 and quality variable data Y0, and calculate the mean and variance of U0 and Y0 respectively;

[0079] S102: Perform standardization processing using the mean and variance, and calculate the process variable matrix U and the quality variable matrix Y according to formulas (1) and (2) as the input and output of the model:

[0080]

[0081] 3. The quality-related fault detection method based on the deep decomposition echo state network involved in this embodiment, wherein step S2 specifically includes the following steps:

[0082] S201: Decompose the process variable matrix U using the principal component regression method, and calculate the covariance matrix Perform eigenvalue decomposition on the matrix U0 to obtain the eigenvector matrix where k is the number of principal components in the principal component regression;

[0083] S202: Further calculate the score matrix according to formula (3)

[0084] T = UP (3)

[0085] Calculate the load matrix according to the least squares regression of the quality variable Y and the score matrix T using formula (4)

[0086] Q T =(T T T) -1 T T Y (4)

[0087] Calculate the correlation coefficient matrix between U and the online predicted value of the quality variable according to formula (5) between

[0088] B = PQ T (5)

[0089] S203: Further perform singular value decomposition on the correlation coefficient matrix B through formula (6):

[0090]

[0091] Partition the matrix into a quality-related part and a quality-unrelated part Finally, use equations (7) and (8) to decompose the process variable U into a quality-related space and a quality-unrelated space

[0092] U1 = UP B1 P B1 T (7)

[0093] U2 = UP B2 P B2 T (8)

[0094] S204: Use the decomposed quality-related process variable U1 as the input and the quality variable matrix Y as the output to train a deep echo state network to obtain the output matrix of the model's quality-related space Use the decomposed quality-unrelated process variable U2 as the input and the quality variable matrix Y as the output to train a deep echo state network to obtain the output matrix of the model's quality-unrelated space

[0095] S205: Predict the output vector according to the quality-related part and the quality-unrelated part's predicted output vector The prediction residual vector can be calculated according to equations (9) and (10):

[0096]

[0097] where y t is the true quality variable vector.

[0098] 3. The quality-related fault detection method based on the deep decomposition echo state network involved in this embodiment, where step S3 specifically includes the following steps:

[0099] S301: Feed the process variable matrix U into the echo state network to extract dynamic features and obtain the dynamic feature matrix Calculate the input matrix X according to equation (11) P :

[0100]

[0101] S302: Use the principal component regression method to perform eigenvalue decomposition on the input matrix X P to obtain the eigenvector matrix Furthermore, the score matrix T can be calculated according to formula (12) as follows:

[0102] T = XP (12)

[0103] Calculate the loading matrix Q and the correlation coefficient matrix B according to formulas (4) to (6), perform singular value decomposition on the B matrix, and divide the matrix into a quality-related part and a quality-unrelated part Then, decompose the dynamic feature X into a quality-related dynamic feature matrix and a quality-unrelated dynamic feature matrix

[0104] X1 = XP B1 P B1 T (13)

[0105] X2 = XP B2 P B2 T (14)

[0106] S303: Send the quality-related dynamic feature matrix X1 into the deep echo state network to extract dynamic features and decompose them according to the method described in S302 to further obtain a deep quality-related dynamic feature matrix X re1 and coefficient matrix P B11 and a deep quality-unrelated dynamic feature matrix X un1 and coefficient matrix P B21 , and calculate the score matrices of their respective parts according to formulas (15) and (16) and

[0107] T re1 = X re1 P B11 (15)

[0108] T un1 = X un1 P B21 (16)

[0109] S304: Send the quality-unrelated dynamic feature matrix X2 into the deep echo state network to extract dynamic features and decompose them according to the method described in S302 to obtain a deep quality-related dynamic feature matrix X re2 and coefficient matrix P B12 and a deep quality-unrelated dynamic feature matrix X un2 and coefficient matrix P B22 , and then calculate their respective score matrices and

[0110] 4. The quality-related fault detection method based on the deep decomposition echo state network involved in this embodiment, wherein step S4 specifically includes the following steps:

[0111] S401: Use the residual information vector obtained during the decomposition of the process variable matrix U, and further obtain the quality-related statistic SPE according to formulas (17) and (18) re and the quality-unrelated statistic SPE un :

[0112]

[0113] S402: According to the score matrix of the quality-related principal component part obtained during the decomposition of the dynamic feature X, further construct the corresponding statistics according to formulas (19) and (20) as follows:

[0114]

[0115] Wherein, and are the score vectors in the score matrix;

[0116] According to the score matrix of the quality-unrelated principal component part obtained during the decomposition of the dynamic feature, further construct the corresponding statistics according to formulas (21) and (22) as follows:

[0117]

[0118] Wherein, and are the score vectors in the score matrix;

[0119] S403: This method extracts quality-related information from two paths: internal feature decomposition and process variable decomposition. Since there are many information modules generated during the deep decomposition process, there are three quality-related statistics and three quality-unrelated statistics each. Therefore, considering all statistics comprehensively, when any statistic alarms, it is considered that a fault has occurred, and all statistics are finally fused into the quality-related statistic and the quality-unrelated statistic

[0120]

[0121] where fuse(·) represents taking the detection result of the statistic when any statistic exceeds the control limit.

[0122] 5. The quality-related fault detection method based on the deep decomposition echo state network involved in this embodiment, wherein step S5 specifically includes the following steps:

[0123] S501: Collect real-time data in the industrial process, and standardize the data using the mean and variance of the training set to obtain the standardized process variable matrix U new ;

[0124] S502: Input the standardized industrial process data U new into the deep decomposition echo state network for process variable decomposition and dynamic feature decomposition;

[0125] S503: Process the online data using the orthogonal projection matrix obtained from the decomposition in the offline stage, and calculate the quality-related statistic and the statistic unrelated to quality

[0126] S504: Judge respectively and whether they exceed the limit. If neither exceeds the limit, there is no fault;

[0127] S505: If exceeds the limit, it indicates that the fault is a quality-related fault, which affects the quality variable;

[0128] S506: If does not exceed the limit while exceeds the limit, it indicates that the fault is a quality-unrelated fault, which cannot affect the change of the quality variable.

[0129] In the above method, steps S1 to S4 are the offline modeling stage, and step S5 is the online detection stage. To more clearly illustrate the beneficial effects of the above quality-related fault detection method of the present invention, the following further illustrates the above quality-related fault detection method of the present invention in conjunction with embodiments.

[0130] Embodiment:

[0131] In this embodiment, the quality-related fault detection method based on the deep decomposition echo state network is applied to the Tennessee - Eastman process. Downs and Vogel of an American chemical company developed the Tennessee - Eastman process based on a real chemical process. This process is widely used to test and evaluate process control methods and fault diagnosis methods. The Tennessee - Eastman process mainly includes five typical chemical plants, as Figure 4 shown, namely the reactor, condenser, stripper, compressor, and separator. The Tennessee - Eastman process involves 8 components from A to H; Figure 3The digital labels 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 respectively represent a kind of fluid, which is simply referred to as: Flow 1, Flow 2, Flow 3, Flow 4, Flow 5, Flow 6, Flow 7, Flow 8, Flow 9, Flow 10, Flow 11, Flow 12, Flow 13 in the following text. The Tennessee - Eastman process in this embodiment contains 52 measurement variables (as shown in Table 1), among which variables 1 - 11 are control variables, variables 12 - 33 are process variables, and 34 - 52 are quality variables. The Tennessee - Eastman process in this embodiment simulates 15 kinds of faults (as shown in Table 2), among which, faults 1 - 7 are step changes of variables, faults 8 - 12 are enhanced volatility of variables, fault 13 is a slow shift of reaction kinetics, and faults 14 - 15 are valve stickiness. Usually, the 15 faults are divided into quality - related faults and quality - unrelated faults. The influence of fault 5 on quality variables is relatively special. The quality variables are abnormal in the early stage of the fault and then gradually return to the normal state. Therefore, fault 5 is separately divided into quality semi - related faults. The specific fault classification is shown in Table 3.

[0132] Table 1 Measurement Variables of Tennessee - Eastman Process

[0133]

[0134]

[0135] Table 2 Fault Descriptions of Tennessee - Eastman Process

[0136] Serial number Fault description Type 1 A / C feed ratio, B component unchanged Step 2 B component, A / C feed ratio unchanged Step 3 D feed temperature Step 4 Inlet temperature of reactor cooling water Step 5 Inlet temperature of condenser cooling water Step 6 A feed loss Step 7 Pressure loss in C presence Step 8 A, B, C feed components Random variable 9 D feed temperature Random variable 10 C feed temperature Random variable 11 Inlet temperature of reactor cooling water Random variable 12 Inlet temperature of condenser cooling water Random variable 13 Reaction dynamics Slow offset 14 Reactor cooling water regulating valve Viscous 15 Condenser cooling water regulating valve Viscous

[0137] Table 3 Fault Classification of TE Process

[0138]

[0139] In this embodiment, three methods, namely the principal component regression method (PCR), the echo state network - principal component regression method (ESN - PCR), and the deep decomposition echo state network method (DDESN), are used for simulation comparison. For the quality - related faults of the TE process, simulation analysis is carried out. Table 4 gives the detection results of the quality - related statistics of the three methods under the conditions of quality - related faults (1, 2, 6, 7, 8, 10, 12, 13). When these faults occur, the quality - related statistics of the model should issue an alarm in time. It can be seen from Table 4 that the detection rate of the DDESN model is the highest for most quality - related faults. For example, for fault 8, the detection rate of DDESN is increased to 97% compared with the other two methods. The final average detection rate of quality - related faults is 88.22%, which is the highest among the three methods.

[0140] Table 4 Quality - Related Statistics under the Conditions of Quality - Related Faults Detection Rate (%)

[0141]

[0142] Perform simulation analysis on the quality - independent faults of the TE process. Table 5 presents the detection results of the quality - related statistics of three methods under the conditions of quality - independent faults (3, 4, 9, 11, 14, 15). Since these faults do not affect the quality variables, the quality - related statistics should issue alarms as few as possible. Since faults 3, 9, and 15 are minor faults and it is difficult to distinguish this type of special fault itself, deeply mining the information of this type of fault will cause a certain degree of decline in the model performance. However, it will have very significant advantages for other faults. As can be seen from Table 5, under the conditions of quality - independent faults, the average alarm rate of the quality - related statistics of the DDESN method in this embodiment is much lower than that of other methods. For example, fault 4 is a quality - independent fault, and the quality - related statistics of the DDESN model is only 0.37%, far lower than the other two methods, indicating that this method effectively identifies that this fault does not affect the quality variables. Similarly, for fault 14, the quality - related statistics of the DDESN model drops to 1.25%. The final average detection rate of the quality - related statistics of the DDESN model under the conditions of quality - independent faults is 3.14%, which is the lowest among the three methods. This result shows that the proposed DDESN method is the most effective in distinguishing whether the occurring faults affect the quality variables, and when a fault occurs but does not affect the quality variables, the false alarm situation of the quality - related statistics is the least.

[0143] Table 5 Quality - related statistics under the conditions of quality - independent faults Detection rate (%)

[0144]

[0145] Perform simulation analysis on the quality semi - related faults of the TE process. The change of quality variable 36 in the case of fault 5 is as Figure 3 shown. After the fault occurs, the quality will fluctuate, but it will return to normal soon. However, the fault is still occurring. Therefore, fault 5 is separately classified as a quality semi - related fault, which means that the influence of this fault will disappear after a period of time in the quality - related space, while the fault still exists in the quality - independent space. Some quality - related fault detection methods generally classify fault 5 as a quality - related fault, and the detection rate of the quality - related statistics is very high. However, this does not conform to the actual situation of the influence of fault 5 on the quality variables, and these detection methods cannot effectively distinguish whether fault 5 affects the quality variables.

[0146] Comparison of the detection results of the quality - related statistics of the three methods is as Figure 4As shown, where (a) is the detection result of the quality-related statistic of the PCR method, and (b) is the detection result of the quality-related statistic of the DDESN method. The quality-related statistic of the traditional PCR method for fault 5 has been alarming, which does not conform to the actual situation. The DDESN method combined with the echo state network issues an alarm at the initial stage of the fault, but then the situation of sample overrun is greatly reduced, which well illustrates that the time for fault 5 to affect the quality variable is short, and subsequent faults occur in the quality-unrelated space. The quality-unrelated statistics of the DDESN method such as Figure 5 As shown, it can be seen that after the alarm stops in the quality-related space, the quality-unrelated statistics still indicate the occurrence of a fault, which very much conforms to the actual impact of fault 5 on the quality variable.

[0147] Based on the above analysis, the DDESN method provided by the present invention, by introducing the principal component regression method, deeply decomposes the quality-related information in industrial process data, enables the traditional echo state network method to judge whether a fault affects the quality variable, and significantly improves the quality-related fault detection effect compared with the principal component regression method.

Claims

1. A quality-related fault detection method based on a deep decomposed echo state network, characterized in that Including the following steps: S1: Collect data and standardize: Obtain industrial process data, divide it into process variable data and quality variable data, and standardize it using its mean and variance. After standardization, a process variable matrix and a quality variable matrix are obtained, where n u is the number of samples, m u is the number of process variables, and m y is the number of quality variables; S2: Construct the first decomposition network to extract residual information: First, decompose the process variable matrix U using the principal component regression method and divide it into the quality-related space and the quality-unrelated space Then, send U1 and U2 into the deep echo state network respectively to extract potential residual information and calculate the residual vector and S3: Construct the second decomposition network to extract dynamic information: Use the echo state network method to extract the dynamic features of the process variable matrix U Decompose X using the principal component regression method and divide it into quality-independent components and quality-related components Further send them into the deep echo state network to extract the dynamic information X re and X un , and calculate their score matrices T respectively re and T un ; S4: Establish a fusion statistic: Comprehensively consider the residual information and dynamic information obtained from the two decomposition paths, and fuse and construct quality-related statistics statistics unrelated to quality and calculate its control limits; S5: Detect quality-related faults: Collect real-time process data, perform standardization processing, calculate the residual vector and score vector of the online data, and determine whether the faults affect the quality variables according to the statistical test results.

2. The quality-related fault detection method based on the deep decomposition echo state network according to claim 1, wherein: The specific steps of step S1 include the following steps: S101: Collect normal operation data samples of the industrial process, divide them into process variable data U0 and quality variable data Y0, and calculate the mean and variance of U0 and Y0 respectively; S102: Perform standardization processing using the mean and variance, and calculate the process variable matrix U and the quality variable matrix Y according to formulas (1) and (2) as the input and output of the model:

3. The quality-related fault detection method based on the deep decomposed echo state network according to claim 1, characterized in that: The specific steps of step S2 include the following steps: S201: Decompose the process variable matrix U using the principal component regression method and calculate the covariance matrix Perform eigenvalue decomposition on the matrix U0 to obtain the eigenvector matrix where k is the number of principal components in the principal component regression; S202: Further calculate the score matrix according to formula (3) : T = UP (3) The load matrix is calculated by the least squares regression of the mass variable Y and the score matrix T according to formula (4). : Q T = (T T T) -1 T T Y(4) Calculate the correlation coefficient matrix between U and the online predicted value of the mass variable according to formula (5). :​ B = PQ T (5) S203: Further perform singular value decomposition on the correlation coefficient matrix B through formula (6): Partition the matrix into a quality-related part and a quality-unrelated part Finally, decompose the process variable U into a quality-related space and a quality-unrelated space : U1 = UP B1 P B1 T (7) U2 = UP B2 P B2 T (8) S204: Use the decomposed quality-related process variable U1 as the input and the quality variable matrix Y as the output to train a deep echo state network to obtain the output matrix of the model's quality-related space Use the decomposed quality-unrelated process variable U2 as the input and the quality variable matrix Y as the output to train a deep echo state network to obtain the output matrix of the model's quality-unrelated space S205: Predict the output vector based on the quality-related part and the output vector predicted by the part unrelated to quality The prediction residual vector can be calculated according to Formulas (9) and (10): where y t is the true quality variable vector.

4. The quality-related fault detection method based on the deep decomposed echo state network according to claim 1, wherein: The specific steps of step S3 include the following steps: S301: Send the process variable matrix U into the echo state network to extract dynamic features and obtain the dynamic feature matrix Calculate the input matrix X according to formula (11) P : S302: Using the principal component regression method, input matrix X P is subjected to eigenvalue decomposition to obtain the eigenvector matrix Furthermore, the score matrix T can be calculated according to formula (12) as follows: T = XP (12) Calculate the loading matrix Q and the correlation coefficient matrix B according to formulas (4) to (6), perform singular value decomposition on the B matrix, and divide the matrix into a mass-related part and a mass-unrelated part Then, use formulas (13) and (14) to decompose the dynamic feature X into a mass-related dynamic feature matrix and a mass-unrelated dynamic feature matrix : X1 = XP B1 P B1 T (13) X2 = XP B2 P B2 T (14) S303: Feed the quality-related dynamic feature matrix X1 into the deep echo state network to extract dynamic features and decompose them according to the method described in S302, further obtaining the deep quality-related dynamic feature matrix X re1 and the coefficient matrix P B11 the deep quality-unrelated dynamic feature matrix X un1 and the coefficient matrix P B21 , and calculate the score matrices of their respective parts according to formulas (15) and (16) and : T re1 = X re1 P B11 (15) T un1 = X un1 P B21 (16) S304: Feed the quality-independent dynamic feature matrix X2 into the deep echo state network to extract dynamic features and decompose them according to the method described in S302 to obtain the deep quality-related dynamic feature matrix X re2 and the coefficient matrix P B12 with the deep quality-independent dynamic feature matrix X un2 and the coefficient matrix P B22 , and then calculate their respective score matrices and 5. The quality-related fault detection method based on the deep decomposed echo state network according to claim 1, characterized in that: The specific steps of step S4 include the following steps: S401: Use the residual information vector obtained during the decomposition of the process variable matrix U to further obtain the quality-related statistic SPE according to formulas (17) and (18), re and the quality-unrelated statistic SPE un : S402: According to the score matrix of the quality-related principal component part obtained during the dynamic feature X decomposition process, further construct the corresponding statistics according to formulas (19) and (20) as follows: Among them, and are score vectors in the score matrix; According to the score matrix of the quality-unrelated principal component part obtained during the dynamic feature decomposition process, further construct the corresponding statistics according to formulas (21) and (22) as follows: Among them, and are score vectors in the score matrix; S403: This method extracts quality-related information from two paths: internal feature decomposition and process variable decomposition. Since there are many information modules generated in the in-depth decomposition process, there are three quality-related statistics and three quality-unrelated statistics respectively. Therefore, considering all statistics comprehensively, when any statistic alarms, it is considered that a failure has occurred, and all statistics are finally fused into quality-related statistics through formulas (23) and (24). and quality-unrelated statistics Where fuse(·) represents taking the test result of the statistic when any statistic exceeds the control limit.

6. A quality-related fault detection method based on a deep decomposition echo state network according to claim 1: The specific steps of step S5 include the following steps: S501: Collect real-time data in the industrial process, and standardize the data using the mean and variance of the training set to obtain the standardized process variable matrix U new ; S502: Input the standardized industrial process data U new into the deep decomposition echo state network for process variable decomposition and dynamic feature decomposition; S503: Process the online data using the orthogonal projection matrix obtained from the offline decomposition stage, and calculate the quality-related statistics Statistics independent of quality S504: Determine respectively whether and exceed the limit. If neither exceeds the limit, no fault occurs; S505: If it is out of limit, it indicates that this fault is a quality-related fault, affecting the quality variable; S506: If is not exceeded while is exceeded, it indicates that this fault is a quality-independent fault and cannot affect the change of quality variables.

Images