An industrial process fault detection method based on prior information assisted neural network modeling

Abstract

The application relates to an industrial process fault detection method based on prior information assisted neural network modeling, and the steps are as follows: after standardization processing of training data, the standardized training data are used to train a 1DRGANomaly network, prior data are combined to construct a prior information assisted network, and in the online stage, the statistical quantity output by the 1DRGANomaly network and the prior information assisted network is fused to calculate the statistical quantity and SPE 12‑PIA , and the control limit is used for monitoring. The above-mentioned fault detection method introduces a prior information assisted strategy, and the detection performance of faults in industrial process data is improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial process fault detection technology, and relates to an industrial process fault detection method based on prior information-assisted neural network modeling. Background Technology

[0002] With the continuous development of science and technology, industrial systems are becoming increasingly integrated and complex. However, this also significantly increases the likelihood of failures in industrial processes. If these failures are not handled promptly, they can lead to serious accidents, even causing personal injury and property damage. Therefore, industrial process fault detection technology is crucial in this context. With the continuous development of industrial control systems and the widespread application of sensor technology, large amounts of data can be acquired in real time, promoting the widespread application of deep learning-based fault detection technology.

[0003] 1DRGANomaly networks are a novel algorithm in the field of industrial process monitoring, proposed by Deng in 2023. However, as an unsupervised learning method, 1DRGANomaly networks utilize only normal data during the modeling process. In real-world industrial scenarios, however, known fault data often exists, which plays a crucial role in improving fault detection performance. Therefore, how to integrate fault information into 1DRGANomaly network modeling is a problem that requires in-depth research. Summary of the Invention

[0004] This invention addresses the issue that while 1DRGANomaly networks, as an unsupervised learning method, utilize normal data in the modeling process, real-world industrial scenarios often involve known fault data. This known fault data is crucial for improving fault detection performance. The invention provides a priori information-assisted neural network modeling method for industrial process fault detection. This method utilizes prior fault data to build an auxiliary model, extracts feature information from historical data, and fuses this auxiliary model with the 1DRGANomaly network to establish an overall monitoring framework. The proposed industrial process fault detection method based on priori information-assisted neural network modeling includes the following steps:

[0005] (i) Collect process data under normal operating conditions from the historical database as training data X. train and verification data X val And using training data X train mean and standard deviation For training data X train and verification data X val The data is standardized to obtain the standardized training data X.train ′ and verification data X val ′;

[0006] (ii) Using training data X train A 1DRGANomaly network was constructed using gradient descent as the main model, and the validation data X was used. val Input the 1DRGANomaly network to obtain deep features z. val and With residual characteristic e 1,val and e 2,val Calculate the corresponding statistic T1 2 , Given the confidence level α, the SPE1 and SPE2 statistics are calculated using the kernel density estimation method, resulting in statistic T1. 2 Corresponding control limits Corresponding control limits The control limit SPE corresponding to the statistic SPE1 1,lim The control limit SPE corresponding to the statistic SPE2 2,lim ;

[0007] (III) Collect available prior fault data X priori Using training data X train mean and standard deviation For prior fault data X priori The standardized prior fault data X′ is obtained by performing standardization. priori At the same time, the training data X train The label of ' is set to 0, and the prior fault X is... train The labels of ' are set to 1 to form the prior dataset. Where x1 to Represents training data, arrive Let X1 represent the prior data, n1 represent the number of training data, and n2 represent the number of prior data. The prior dataset X1 is used as input, and a fully connected neural network is constructed using the gradient descent method as an auxiliary model for prior information.

[0008] (iv) Collecting test data x test Using training data X train mean and standard deviation For test data x test Standardization is performed to obtain standardized test data x. test ′;

[0009] (v) Test data x test Input the 1DRGANomaly network to obtain deep features z.test and With residual characteristic e 1,test and e 2,test Calculate the corresponding statistic SPE 1,t and SPE 2,t A Bayesian fusion strategy was employed to obtain the fused statistics. and SPE 12,t With control limits and SPE lim ; test data x test Input a fully connected neural network and obtain the output.

[0010] (vi) Statistical measures and SPE 12,t With output Weighted fusion yields new statistics and SPE 12-PIA,t If any statistic exceeds the limit, then the test data x is judged to be... test A malfunction has occurred.

[0011] Furthermore, in step (a), the training data X train The standardization process is as follows:

[0012] Using training data X train mean and standard deviation Using formula (1) to analyze the training data X train After standardization, the expression for formula (1) is:

[0013]

[0014] Training data X train After standardization using the above formula (1), the standardized training data X can be obtained. train ′.

[0015] Using training data X train mean and standard deviation The verification data X is tested using formula (2). val After standardization, the expression for formula (2) is:

[0016]

[0017] Validation data X val After standardization using the above formula (2), the standardized verification data X can be obtained. val ′.

[0018] Furthermore, in step (ii), the generator loss function L in the 1DRGANomaly network is calculated using formulas (3) and (4). D-Net and the loss function L for discrimination G-Net After the loss functions of the two networks have fully converged, their respective training weights are saved. The specific process is as follows:

[0019] minL D-Net =L adv (3)

[0020] minL G-Net =λ1L con +λ2L enc +λ3L adv (4)

[0021] In the formula, λ1, λ2, and λ3 are the input reconstruction loss L con Encoder reconstruction loss L enc and combat losses L adv The coefficient of L. con L enc and L adv As shown in formulas (5)-(7):

[0022]

[0023] In the formula, z is the output of encoder 1. x is the output of encoder 2, and x is the input. f is the output of the decoder, and f(·) is the output of the discriminator.

[0024] Verification data X train By inputting the 1DRGANomaly network, deep features z can be obtained. train and Verification data X val By inputting the 1DRGANomaly network, deep features z can be obtained. val and With residual characteristic e 1,val and e 2,val ,in Based on feature z val , e 1,val and e 2,val The statistics T1 are constructed using formulas (8)-(11) respectively. 2 , The SPE1 and SPE2 statistics, expressed by formulas (8)-(11), are as follows:

[0025]

[0026]

[0027] SPE2=||e 1,val ||1 (11)

[0028] In the formula, Σ1 and Σ2 are z train and The covariance matrix is ​​given by ||·||1 and ||·||2, which represent the Manhattan distance and Euclidean distance, respectively.

[0029] Given a confidence level α, calculate the statistic T1 using the kernel density estimation method. 2 Corresponding control limits Corresponding control limits The control limit SPE corresponding to the statistic SPE1 1,lim The control limit SPE corresponding to the statistic SPE2 2,lim .

[0030] Furthermore, in step (iii), available prior fault data X is collected. priori Using training data X train mean and standard deviation For prior fault data X priori The standardized prior fault data X′ is obtained by performing standardization. priori The specific process is as follows:

[0031] Using training data X train mean and standard deviation Using formula (12) to analyze the prior fault data X priori After standardization, the expression for formula (12) is:

[0032]

[0033] Prior fault data X priori After standardization using the above formula (12), the standardized prior fault data X′ can be obtained. priori .

[0034] Use the training data X train The label of ' is set to 0, and the prior fault X is... train The labels of ' are set to 1 to form the prior dataset. Where x1 to Represents training data, arrive Let n1 represent the number of training data points and n2 represent the number of prior data points. Using the prior dataset X1, the loss function L in the prior information-assisted discrimination network is calculated using formula (13).PIA-Net After the network's loss function fully converges, its respective training weights are preserved.

[0035]

[0036] In the formula, y is the real label. It is the predicted value of a fully connected neural network.

[0037] Furthermore, in step (iv), the training data X is used. train mean and standard deviation The test data x is analyzed using formula (14). test The standardization process is as follows:

[0038]

[0039] Test data x test After standardization using the above formula (14), the standardized test data x can be obtained. test ′.

[0040] Furthermore, in step (five), the test data x test Inputting into the 1DRGANomaly network allows us to obtain deep features z. test and With residual characteristic e 1,test and e 2,test X′ priori The specific process is as follows:

[0041] Calculate z test,t Corresponding statistics Corresponding statistics e 1,test,t The corresponding statistic SPE 1,t and e 2,test,t The corresponding statistic SPE 2,t The calculation formulas are shown in formulas (15) to (18):

[0042]

[0043] SPE 2,t =||e 2,test,t ||1 (18)

[0044] In the formula, Σ1 and Σ2 are z train and The covariance matrix is ​​given by ||·||1 and ||·||2, which represent the Manhattan distance and Euclidean distance, respectively.

[0045] For each substatistic, for a given test vector x, the posterior failure probability and As shown in formulas (19) and (20).

[0046]

[0047] In the formula, and Let be the prior probability of the fault. and This represents the conditional probability under fault conditions. and The probability of a sample appearing can be calculated using equations (21) and (22).

[0048]

[0049] In the formula, and This represents the prior normal probability. and This represents the probability of occurrence under normal operating conditions. The prior normal probability and failure probability can be determined from process knowledge, and are usually set as the confidence level γ and the significance level 1-γ, respectively.

[0050] After converting all statistics into failure probabilities, global monitoring statistics are constructed through Bayesian inference mechanism, as shown in equations (23) and (24).

[0051]

[0052] Control Limits and SPE lim This is the confidence level γ.

[0053] Test data x test Input prior information to assist the network, and the model output can be obtained. The specific process is as follows:

[0054] o′ fault,t The output of the prior information-assisted network after passing through the softmax function is shown in formula (25).

[0055]

[0056] In the formula, e is a natural number. The output of neuron 0, This is the output of neuron 1.

[0057] Set the output of the neural network to 0 and 1. The output is shown in formula (26).

[0058]

[0059] Furthermore, in step (six), the statistical measure... and SPE 12,t With output Weighted fusion to obtain statistics and SPE 12-PIA,t If any statistic exceeds the limit, then the test data x is judged to be... test The malfunction occurred, and the specific process is as follows:

[0060] Statistic and SPE 12-PIA,t The weighted fusion formulas are shown in equations (27) and (28):

[0061]

[0062] If any statistic or SPE 12-PIA,t Exceeding the limit, i.e. or SPE 12-PIA,t ≥SPE lim Then determine the test data x test A malfunction has occurred. Attached Figure Description

[0063] Figure 1 This is a flowchart of the industrial process fault detection method based on prior information-assisted neural network modeling according to the present invention.

[0064] Figure 2a The structure of this invention Figure 1 DRGANomaly network structure diagram;

[0065] Figure 2b This is a structural diagram of a fully connected neural network, as shown in the present invention.

[0066] Figure 2c This is the network structure diagram of 1DRGANomaly based on prior information assisted by the present invention;

[0067] Figure 3 This is a structural diagram of the Tennessee-Eastman (TE) process described in an embodiment of the present invention;

[0068] Figure 4a This is a schematic diagram illustrating the monitoring results of TE process fault 19 using a 1DRGANomaly network in an embodiment of the present invention.

[0069] Figure 4b This is a schematic diagram illustrating the monitoring results of TE process fault 19 using the prior information-assisted 1DRGANomaly network industrial process fault detection method described in this invention, as per an embodiment of the present invention. Detailed Implementation

[0070] The present invention will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0071] See Figure 1 This invention discloses an industrial process fault detection method based on prior information-assisted 1DRGANomaly networks, comprising the following steps:

[0072] (a) Collect process data under normal operating conditions from the historical database as training data X train and verification data X val And using training data X train mean and standard deviation For training data X train and verification data X val The data is standardized to obtain the standardized training data X. train ′ and verification data X val The specific steps are as follows:

[0073] Using training data X train mean and standard deviation Using formula (1) to analyze the training data X train After standardization, the expression for formula (1) is:

[0074]

[0075] Training data X train After standardization using the above formula (1), the standardized training data X can be obtained. train ′.

[0076] Using training data X train mean and standard deviation The verification data X is tested using formula (2). val After standardization, the expression for formula (2) is:

[0077]

[0078] Validation data X val After standardization using the above formula (2), the standardized verification data X can be obtained. val ′.

[0079] (II) See Figure 2a Using training data X trainA 1DRGANomaly network was constructed using gradient descent as the main model, and the validation data X was used. val Input the 1DRGANomaly network to obtain deep features z. val and With residual characteristic e 1,val and e 2,val Calculate the corresponding statistic T1 2 , Given the confidence level α, the SPE1 and SPE2 statistics are calculated using the kernel density estimation method, resulting in statistic T1. 2 Corresponding control limits Corresponding control limits The control limit SPE corresponding to the statistic SPE1 1,lim The control limit SPE corresponding to the statistic SPE2 2,lim The specific steps are as follows:

[0080] The generator loss function L in the 1DRGANomaly network is calculated using formulas (3) and (4). D-Net and the loss function L for discrimination G-Net After the loss functions of the two networks have fully converged, their respective training weights are saved. The specific process is as follows:

[0081] minL D-Net =L adv (3)

[0082] minL G-Net =λ1L con +λ2L enc +λ3L adv (4)

[0083] In the formula, λ1, λ2, and λ3 are the input reconstruction loss L con Encoder reconstruction loss L enc and combat losses L adv The coefficient of L. con L enc and L adv As shown in formulas (5)-(7):

[0084]

[0085] In the formula, z is the output of encoder 1. x is the output of encoder 2, and x is the input. f is the output of the decoder, and f(·) is the output of the discriminator.

[0086] Verification data X train By inputting the 1DRGANomaly network, deep features z can be obtained. trainand Verification data X val By inputting the 1DRGANomaly network, deep features z can be obtained. val and With residual characteristic e 1,val and e 2,val ,in Based on feature z val , e 1,val and e 2,val The statistics T1 are constructed using formulas (8)-(11) respectively. 2 , The SPE1 and SPE2 statistics, expressed by formulas (8)-(11), are as follows:

[0087]

[0088] SPE2=||e 1,val ||1 (11)

[0089] In the formula, Σ1 and Σ2 are z train and The covariance matrix is ​​given by ||·||1 and ||·||2, which represent the Manhattan distance and Euclidean distance, respectively.

[0090] Given a confidence level α, calculate the statistic T1 using the kernel density estimation method. 2 Corresponding control limits Corresponding control limits The control limit SPE corresponding to the statistic SPE1 1,lim The control limit SPE corresponding to the statistic SPE2 2,lim .

[0091] (III) Collect available prior fault data X priori Using training data X train mean and standard deviation For prior fault data X priori The standardized prior fault data X′ is obtained by performing standardization. priori At the same time, the training data X train The label of ' is set to 0, and the prior fault X is... train The labels of ' are set to 1 to form the prior dataset. Where x1 to Represents training data, arrive Let n1 represent the number of training data points and n2 represent the number of prior data points. See also... Figure 2bUsing the prior dataset X1, a fully connected neural network is constructed using gradient descent as a prior information-assisted model. The specific steps are as follows:

[0092] Using training data X train mean and standard deviation Using formula (12) to analyze the prior fault data X priori After standardization, the expression for formula (12) is:

[0093]

[0094] Prior fault data X priori After standardization using the above formula (12), the standardized prior fault data X′ can be obtained. priori .

[0095] Use the training data X train The label of ' is set to 0, and the prior fault X is... train The labels of ' are set to 1 to form the prior dataset. Where x1 to Represents training data, arrive Let n1 represent the number of training data points and n2 represent the number of prior data points. Using the prior dataset X1, the loss function L in the prior information-assisted discrimination network is calculated using formula (13). PIA-Net After the network's loss function fully converges, its respective training weights are preserved.

[0096]

[0097] In the formula, y is the real label. It is the predicted value of a fully connected neural network.

[0098] See Figure 2c This refers to the 1DRGANomaly network model, which is constructed using prior information assistance by combining the 1DRGANomaly network main model with a fully connected neural network auxiliary model.

[0099] (iv) Collect test data x test Using training data X train mean and standard deviation For test data x test Standardization is performed to obtain standardized test data x. test The specific steps are as follows:

[0100]

[0101] Test data x testAfter standardization using the above formula (14), the standardized test data x can be obtained. test ′.

[0102] (v) Test data x test Input the 1DRGANomaly network to obtain deep features z. test and With residual characteristic e 1,test and e 2,test Calculate the corresponding statistic SPE 1,t and SPE 2,t A Bayesian fusion strategy was employed to obtain the fused statistics. and SPE 12,t With control limits and SPE lim The specific steps are as follows:

[0103] Test data x test Inputting into the 1DRGANomaly network allows us to obtain deep features z. test and With residual characteristic e 1,test and e 2,test X′ priori The specific process is as follows:

[0104] Calculate z test,t Corresponding statistics Corresponding statistics e 1,test,t The corresponding statistic SPE 1,t and e 2,test,t The corresponding statistic SPE 2,t The calculation formulas are shown in formulas (15) to (18):

[0105]

[0106] SPE 2,t =||e 2,test,t ||1 (18)

[0107] In the formula, Σ1 and Σ2 are z train and The covariance matrix is ​​given by ||·||1 and ||·||2, which represent the Manhattan distance and Euclidean distance, respectively.

[0108] For each substatistic, for a given test vector x, the posterior failure probability and As shown in formulas (19) and (20).

[0109]

[0110] In the formula, and Let be the prior probability of the fault. and This represents the conditional probability under fault conditions. and The probability of a sample appearing can be calculated using equations (21) and (22).

[0111]

[0112]

[0113] In the formula, and This represents the prior normal probability. and This represents the probability of occurrence under normal operating conditions. The prior normal probability and failure probability can be determined from process knowledge, and are usually set as the confidence level γ and the significance level 1-γ, respectively.

[0114] After converting all statistics into failure probabilities, global monitoring statistics are constructed through Bayesian inference mechanism, as shown in equations (23) and (24).

[0115]

[0116] Control Limits and SPE lim This is the confidence level γ.

[0117] Test data x test Input a fully connected neural network and obtain the output. The specific steps are as follows:

[0118] o′ fault,t The output of the prior information-assisted network after passing through the softmax function is shown in formula (25).

[0119]

[0120] In the formula, e is a natural number. The output of neuron 0, This is the output of neuron 1.

[0121] Set the output of the neural network to 0 and 1. The output is shown in formula (26).

[0122]

[0123] (vi) Statistical measures and SPE 12,t With output Weighted fusion yields new statistics and SPE 12-PIA,t If any statistic exceeds the limit, then the test data x is judged to be... test The following steps are required to resolve the malfunction:

[0124] Statistic and SPE 12-PIA,t The weighted fusion formulas are shown in equations (27) and (28):

[0125]

[0126] In the formula, λ is the weighting parameter.

[0127] If any statistic or SPE 12-PIA,t Exceeding the limit, i.e. or SPE 12-PIA,t ≥SPE lim Then determine the test data x test A malfunction has occurred.

[0128] In the above method, steps (i) to (iii) are the offline modeling stage, and steps (iv) to (vi) are the online monitoring stage.

[0129] The fault detection method of the present invention standardizes the training data, trains the 1DRGANomaly network using the standardized training data, and trains a prior information auxiliary network by combining prior information. In the online stage, the statistical quantities output by the 1DRGANomaly network and the prior information auxiliary network are obtained using the test set, and the statistical quantities of the two models are fused to establish an overall monitoring framework, thereby improving the fault detection performance of the model.

[0130] To more clearly illustrate the beneficial effects of the fault detection method of the present invention, the following description, in conjunction with embodiments, further explains the fault detection method of the present invention.

[0131] Example:

[0132] The Tennessee-Eastman (TE) process is an experimental platform developed by Downs and Vogel of Eastman Chemical Company based on a real chemical process. It is now widely used to verify the merits of control algorithms and process monitoring methods. See also... Figure 3 The TE process mainly consists of five units: a reactor, a product condenser, a gas-liquid separator, a circulating compressor, and a stripping tower. The TE process has a total of 52 variables, including 41 measured variables and 11 operational variables. In this embodiment, as shown in Table 1, there are 21 faults.

[0133] Table 1

[0134]

[0135]

[0136] In this embodiment, two methods are used for simulation comparison: the 1DRGANomaly network and the prior information-assisted 1DRGANomaly network of this invention. The fault detection rate (FDR) is used as the performance metric, defined as the percentage of fault samples exceeding the control limit compared to all actual fault samples. This invention sets the confidence level to 99%, the weighting parameter λ to 0.05, and the first 160 samples in the test data are considered normal samples, with subsequent samples being fault samples.

[0137] Taking fault 19 as an example, the fault detection effect is illustrated. Fault 19 is an unknown fault. A simulation comparison is performed using the traditional 1DRGANomaly network and the 1DRGANomaly network assisted by the prior information of this invention. The monitoring graph for fault 19 can be found in [link to monitoring graph]. Figure 4a and Figure 4b See also Figure 4a Traditional 1DRGANomaly networks have some detection effectiveness, but the overall detection performance is poor, and their statistical measures are limited. The detection rate was 81.13%, and the statistical measure SPE was... 12 The detection rate was 69.63%. (See also...) Figure 4b The prior information-assisted 1DRGANomaly network provided by this invention provides statistical statistics. The detection rate was 94.50%, and the statistical measure SPE was [missing information]. 12-PIA The detection rate was 88.63%, which is the highest compared to the 1DRGANomaly network. Therefore, the 1DRGANomaly network provided by this invention significantly improves the detection performance of TE process fault 19.

[0138] Table 2 shows the fault detection rates of the traditional 1DRGANomaly network and the prior information-assisted 1DRGANomaly network for TE process faults 1-21.

[0139] Table 2

[0140]

[0141]

[0142] As shown in Table 2, the prior information-assisted 1DRGANomaly network provided by this invention achieved the best overall monitoring results and demonstrated the best fault detection performance. Furthermore, the improvement in monitoring performance for faults 9, 11, 15, 16, 19, 20, and 21 was particularly significant.

[0143] Based on the above analysis, the prior information-assisted 1DRGANomaly network provided by this invention, by using prior information to establish a prior information-assisted discrimination network and fusing the above network statistics to obtain new statistics, has a better fault detection effect than the 1DRGANomaly network.

[0144] The embodiments described above are merely illustrative of the present invention and are not intended to limit the scope of protection of the present invention. Any simple modifications and variations made by those skilled in the art within the scope of the technical solutions described in the present invention should be included within the scope of the above-mentioned patent applications.

Claims

1. A method for industrial process fault detection based on prior information-assisted neural network modeling, characterized in that, Includes the following steps: （ 1) Collect process data under normal operating conditions from the historical database as training data. and verification data and using training data mean and standard deviation training data and verification data Standardization is performed to obtain standardized training data. and verification data ; (ii) Using training data A 1DRGANomaly network was constructed using gradient descent as the main model, and validation data was used. Input the 1DRGANomaly network to obtain deep features and With residual characteristics and Calculate the corresponding statistics , , and Statistic, given confidence level The statistic is calculated using the kernel density estimation method. Corresponding control limits , Corresponding control limits Statistics Corresponding control limits and statistics Corresponding control limits ; (iii) Collect available prior fault data Using training data mean and standard deviation Prior fault data Standardization is performed to obtain standardized prior fault data. At the same time, the training data The label is set to 0, indicating a priori faults. The labels are set to 1 to form the prior dataset. ;in arrive Represents training data, arrive Represents prior data, Indicates the number of training data. Indicates the number of prior data points, using the prior dataset. A fully connected neural network is constructed using gradient descent as a prior information-assisted model. (iv) Collect test data Using training data mean and standard deviation For test data Standardization is performed to obtain standardized test data. ; (v) Test data Input the 1DRGANomaly network to obtain deep features and With residual characteristics and Calculate the corresponding statistic , , and A Bayesian fusion strategy was employed to obtain the fused statistics. and With control limits and Test data Input is fed into a fully connected neural network to obtain the output. ; (vi) Statistical measures and With output Weighted fusion yields new statistics and If any statistic exceeds the limit, then the test data is judged to be invalid. A malfunction has occurred.

2. The method as described in claim 1, characterized in that, In step (i), the training data Standardization is performed, specifically by using training data. mean and standard deviation Using formula (1) to process the training data After standardization, the expression for formula (1) is: （1） Training data After standardization using the above formula (1), the standardized training data can be obtained. ; Using training data mean and standard deviation The verification data is tested using formula (2). After standardization, the expression for formula (2) is: （2） Validation data After standardization using the above formula (2), the standardized verification data can be obtained. .

3. The method as described in claim 2, characterized in that, In step (ii), the generator loss function in the 1DRGANomaly network is calculated using formulas (3) and (4). and the loss function of discrimination Once the loss functions of the two networks have fully converged, their respective training weights are saved. （3） （4） In the formula , and These are the input reconstruction losses. Encoder reconstruction loss and combat losses The coefficient; , and As shown in formulas (5)-(7): （5） （6） （7） In the formula For the output of encoder 1, For the output of encoder 2, For input, This is the output of the decoder. For the output of the discriminator; Verification data By inputting the 1DRGANomaly network, deep features can be obtained. and , will verify data By inputting the 1DRGANomaly network, deep features can be obtained. and With residual characteristics and ,in , ; Based on characteristics , , and The statistics are constructed using formulas (8)-(11) respectively. , , and The statistics, expressed by formulas (8)-(11), are as follows: （8） （9） （10） （11） In the formula, and They are respectively and The covariance matrix, and These represent Manhattan distance and Euclidean distance, respectively. Given confidence level The statistic is calculated using the kernel density estimation method. Corresponding control limits , Corresponding control limits Statistics Corresponding control limits and statistics Corresponding control limits .

4. The method as described in claim 3, characterized in that, In step (iii), available prior fault data is collected. Using training data mean and standard deviation Prior fault data Standardization is performed to obtain standardized prior fault data. The specific process is as follows: Using training data mean and standard deviation Using formula (12) to analyze prior fault data After standardization, the expression for formula (12) is: （12） Prior fault data After standardization using the above formula (12), the standardized prior fault data can be obtained. ; training data The label is set to 0, indicating a priori faults. The labels are set to 1 to form the prior dataset. ;in arrive Represents training data, arrive Represents prior data, Indicates the number of training data. Indicates the number of prior data points, using the prior dataset. The loss function in the prior information-assisted discrimination network is calculated using formula (13). After the network's loss function fully converges, its respective training weights are preserved. （13） In the formula, It's a real label. It is the predicted value of a fully connected neural network.

5. The method as described in claim 4, characterized in that, In step (iv), training data is used. mean and standard deviation The test data is analyzed using formula (14). The standardization process is as follows: （14） Test data After standardization using the above formula (14), the standardized test data can be obtained. .

6. The method as described in claim 5, characterized in that, In step (v), the test data Inputting into the 1DRGANomaly network allows for the acquisition of deep features. and With residual characteristics and The specific process is as follows: calculate Corresponding statistics , Corresponding statistics , Corresponding statistics and Corresponding statistics The calculation formulas are shown in formulas (15) to (18): （15） （16） （17） （18） In the formula, and They are respectively and The covariance matrix, and These represent Manhattan distance and Euclidean distance, respectively. For each substatistic, for a given test vector Posterior probability of failure and As shown in formulas (19) and (20); （19） （20） In the formula, and This represents the prior probability of the fault. and This represents the conditional probability under fault conditions. and The probability of a sample appearing can be calculated using equations (21) and (22); （21） （22） In the formula, and This represents the prior normal probability. and This represents the probability of occurrence under normal operating conditions; the prior normal probability and failure probability can be determined from process knowledge, and are usually set as confidence levels respectively. and significance level 1- ; After converting all statistics into failure probabilities, global monitoring statistics are constructed through Bayesian inference mechanism, as shown in equations (23) and (24); （23） （24） Control Limits and That is, the confidence level. ; Test data By inputting prior information into the auxiliary network, the model output can be obtained. The specific process is as follows: The output of the prior information-assisted network after passing through the softmax function is shown in formula (25). （25） In the formula, For natural numbers, The output of neuron 0, This is the output of neuron 1; Set the output of the neural network to 0 and 1. The output is shown in formula (26). （26）。 7. The method as described in claim 6, characterized in that, In step (vi), the statistical quantity is... and With output Weighted fusion to obtain statistics and If any statistic exceeds the limit, then the test data is judged to be invalid. The malfunction occurred, and the specific process is as follows: Statistic and The weighted fusion formulas are shown in equations (27) and (28): （27） （28） In the formula, Weighted parameters; If any statistic or Exceeding the limit, i.e. or Then judge the test data A malfunction has occurred.

Images