Sensor fault robust multi-output soft measurement method and system based on adversarial learning

Abstract

The invention discloses a robust multi-output soft measurement method and system for sensor faults based on adversarial learning, and relates to the robust multi-output soft measurement method and system for the sensor faults. The method aims at solving the problems that most existing soft measurement models are single-output soft measurement models, and prediction performance is prone to being interfered by sensor faults. The system comprises an adversarial sample generation module, a spatial feature extraction module, a linear attention mechanism module, a time feature extraction module and a prediction module. The invention belongs to the technical field of soft measurement.

Description

Technical Field

[0001] The present invention relates to a sensor fault robust multi-output soft sensor method and system, belonging to the technical field of soft sensors. Background Art

[0002] In the industrial production process, key variables usually refer to variables that are closely related to product quality but difficult to directly measure. To improve product quality, optimize the production process, and ensure the safety of operators, it is particularly important to accurately measure these key variables. Soft sensor technology realizes the prediction of key variables by constructing a mathematical model, taking easily measurable process variables as inputs and difficult-to-measure key variables as outputs. According to different construction methods, soft sensor models can be divided into two categories: knowledge-driven soft sensor models and data-driven soft sensor models. Knowledge-driven methods rely on a deep understanding of the production process mechanism and are suitable for production environments with clear mechanisms. However, in complex production processes where the mechanism has not been fully understood, the applicability of knowledge-driven methods is limited. In contrast, due to the convenience of data collection and the low requirement for knowledge of the production process mechanism, data-driven soft sensor models have been widely used in industrial production. From the perspective of the number of predicted variables, soft sensor models can be divided into single-output soft sensor models and multi-output soft sensor models. Current research mainly focuses on single-output models, that is, models that only predict one key variable at a time. However, in many industrial production processes, multiple key variables need to be monitored simultaneously, and single-output models are difficult to meet this demand. Therefore, multi-output soft sensor models become particularly important. On the other hand, data-driven soft sensor models usually rely on data collected by hardware sensors. When some sensors fail, the input data may deviate, resulting in errors in the prediction of key variables by the soft sensor model, which in turn affects the production process. Although the robustness of the model can be enhanced by training the model with sensor fault data, due to the variety and unpredictability of sensor faults, it is extremely difficult to collect sufficient and applicable fault data. Therefore, how to construct a multi-output soft sensor method and system that is robust to sensor faults without using sensor fault data has become an urgent problem to be solved.

[0003] Since most existing soft sensor models are single-output soft sensor models and their prediction performance is easily interfered by sensor faults, the following defects exist:

[0004] (1) It is impossible to simultaneously predict multiple variables

[0005] Most current data-driven soft sensor models mainly focus on single-output problems, that is, only one key variable can be predicted each time. This cannot meet the actual needs in many industrial scenarios. For example, in the penicillin production process, multiple key variables (such as penicillin concentration, biomass concentration, etc.) need to be monitored simultaneously, while a single-output model cannot predict multiple variables at the same time, resulting in an inability to comprehensively and accurately evaluate the product quality at the current moment. Therefore, the existing technology has the defect of being unable to meet the demand for simultaneous prediction of multiple variables.

[0006] (2) The prediction accuracy of the soft sensor model decreases when the hardware sensor fails

[0007] In actual industrial production, the occurrence of hardware sensor failures is common, and the types of failures are complex and diverse. Current data-driven soft sensor models usually assume that the sensor data is intact, ignoring the data deviation that may be brought about by sensor failures. Sensor failures will cause deviations in the model input data, thus affecting the prediction results and reducing the stability and safety of the production process. Although some methods attempt to solve this problem, due to the inability to predict the type and occurrence time of the failure, the existing technology has not provided an effective and generally applicable solution for sensor failure robustness.

[0008] (3) The cost of collecting a suitable sensor failure training set is relatively high

[0009] Using sensor failure data to train the model can enhance the robustness of the model to sensor failures. However, due to the wide variety and unpredictability of sensor failure types, it requires a relatively high cost to collect sufficient and applicable failure data. Summary of the Invention

[0010] The present invention aims to solve the problem that most existing soft sensor models are single-output soft sensor models and their prediction performance is easily interfered by sensor failures, and further proposes a sensor failure robust multi-output soft measurement method and system based on adversarial learning.

[0011] The technical solution adopted by the present invention to solve the above problems is as follows: The steps of a sensor failure robust multi-output soft measurement method based on adversarial learning according to the present invention include:

[0012] Step 1: Obtain the historical data set D = {(X, Y)}, and split it into the training set D tr = {(X tr , Y tr )} and the fault-free test set D te = {(X te , Y te )}, and perform data processing on the training set and the fault-free test set;

[0013] Step 2: Use the processed X trInput into the untrained soft sensor model to obtain the temporary predicted value Y of the key variable 1 , generate adversarial sample X using the fast gradient sign method adv , and use (X adv , Y tr ) to train the model;

[0014] Step 3. Inject faults into the process variable data X in the processed fault-free test set according to the sensor fault mathematical model to obtain the sensor bias fault test data X te , the sensor stuck fault test data X te_bias , the sensor Gaussian noise fault test data X te_stuck , the sensor spike fault test data X te_gauss_noise ; te_spike ;

[0015] Step 4. Input the fault-free test data X te and the bias fault test data X te_bias , the stuck fault test data X te_stuck , the Gaussian noise fault test data X te_gauss_noise , the spike fault test data X te_spike into the trained soft sensor model to obtain the predicted values Y pr , Y pr_bias , Y pr_stuck , Y pr_gauss_noise , Y pr_spike of the key variable respectively.

[0016] Furthermore, Step 101. Process the obtained data set to obtain the training set D tr ={(X tr , Y tr )} and the fault-free test set D te ={(X te , Y te )}, where X represents the process variable, Y represents the process variable, t represents the time step, p represents the number of process variables, and k represents the number of key variables;

[0017] Step 102. After obtaining the training set D tr ={(X tr , Y tr )} and the fault-free test set D te ={(X te , Y te )}, perform "maximum-minimum" normalization processing on them respectively. The maximum-minimum normalization processing process is where represents the variable value after normalization, x represents the variable value without processing, min represents the minimum value in the variable, and max represents the maximum value in the variable.

[0018] Further, step 2 specifically includes:

[0019] Step 201: Input the processed fault-free training data X tr into the untrained soft sensor model to obtain the predicted value Y of the key variable 1 ;

[0020] Step 202: Obtain the predicted value Y of the key variable 1 , and use the fast gradient sign method to generate the adversarial sample X adv ;

[0021] Step 203: Use (X adv , Y tr ) to train the soft sensor model, and the loss function used during the training process is the root mean square error loss function.

[0022] Further, the specific steps for using the fast gradient sign method to generate the adversarial sample X adv in step 202 are as follows:

[0023] Step 20201: Calculate the loss function L between the predicted value Y 1 and the true value Y tr , and the mean square error loss function is used;

[0024] Step 20202: Calculate the partial derivative of the loss function L with respect to X tr ;

[0025] Step 20203: Use the fast gradient sign method to generate the adversarial sample X adv ,

[0026] Further, step 3 specifically includes:

[0027] Step 301: Use the fault-free test data X te and the sensor fault mathematical model to construct the deviation sensor fault test data X te_bias , randomly select n process variables to add faults, and add 0.3 to the normalized data of these n process variables;

[0028] Step 302: Use the fault-free test data X te and the sensor fault mathematical model to construct the stuck sensor fault test data X te_stuck , randomly select n process variables to add faults, and add 0.3 to the normalized data of these n process variables;

[0029] Step 303: Use the fault-free test data X te and the sensor fault mathematical model to construct Gaussian noise sensor fault test data X te_gauss_noise , randomly select n process variables to add faults, and add Gaussian noise with a mean of 0.3 and a variance of 0.1 to the original data;

[0030] Step 304: Use the fault-free test data X te and the sensor fault mathematical model to construct spike sensor fault test data Y pr_spike , randomly select n process variables to add faults, randomly select some moments in these n process variables, and add 10 to the original data.

[0031] Furthermore, step 4 specifically includes:

[0032] Step 401: Input the fault-free test data X te into the trained soft sensor model to obtain the predicted value Y pr of the key variable;

[0033] Step 402: Input the deviation fault test data X te_bias into the trained soft sensor model to obtain the predicted value Y pr_bias of the key variable;

[0034] Step 403: Input the stuck fault test data X te_stuck into the trained soft sensor model to obtain the predicted value Y pr_stuck of the key variable;

[0035] Step 404: Input the Gaussian noise fault test data X te_gauss_noise into the trained soft sensor model to obtain the predicted value Y pr_gauss_noise of the key variable;

[0036] Step 405: Input the spike fault test data X te_spike into the trained soft sensor model to obtain the predicted value Y pr_spike .

[0037] The sensor fault robust multi-output soft measurement system based on adversarial learning described in the present invention includes an adversarial sample generation module, a spatial feature extraction module, a linear attention mechanism module, a temporal feature extraction module, and a prediction module;

[0038] The adversarial sample generation module uses the fault-free data set D tr ={(X tr , Y tr )} as the training set, inputs the training set into the adversarial sample generation module composed of a gated recurrent unit and a fully connected layer, and generates adversarial samples X through the fast gradient sign method.adv ;

[0039] Spatial feature extraction module, which inputs the adversarial sample X generated by the adversarial sample generation module adv into the spatial feature extraction module composed of a one-dimensional convolutional neural network to obtain the feature H containing the spatial information in the process variables s ;

[0040] Linear attention mechanism module, which inputs the feature H containing the spatial information in the process variables s into the linear attention mechanism module, so as to obtain the selection vector R with the same size as X adv The element size of this vector represents the contribution of the corresponding position process variable to the output. The larger the element value, the more important the corresponding process variable is to the prediction result. On the contrary, it means that the variable is not important to the prediction result or the sensor collecting this variable fails;

[0041] Temporal feature extraction module, which is reused with the prediction module and the adversarial sample generation module. It inputs the adversarial sample X generated by the adversarial sample generation module adv and the selection vector R generated by the linear attention mechanism module through Hadamard product into the temporal feature extraction module, so as to obtain the feature H containing spatio-temporal information s-t and reduce the influence of sensor fault data on the prediction result;

[0042] Prediction module, which contains two sub-modules, and respectively receives the feature H containing spatio-temporal information extracted by the temporal feature extraction module s-t to obtain the predicted values Y pr1 、Y pr2 .

[0043] The beneficial effects of the present invention are as follows: The present invention consists of three parts: an adversarial sample generation stage, a training stage, and a multi-variable prediction stage; compared with the traditional single-output soft sensor method, the method of the present invention adds an adversarial sample generation module, which generates adversarial samples by using easily collectable fault-free data to replace the sensor fault data for training the model, thus solving the problem of difficult collection of sensor fault data and effectively improving the robustness of the model to sensor faults; by adding a spatial feature extraction module, a linear attention mechanism module, and a temporal feature extraction module to the proposed method, the present invention effectively mines the spatio-temporal information in the process variables and reduces the adverse effects of sensor fault data on the prediction result, thereby improving the prediction performance of the model for key variables and the robustness to sensor faults; by adding a prediction module containing two sub-modules to the proposed method, the present invention has the ability to simultaneously predict multiple key variables; finally, by adjusting the hyperparameters and loss function of the model, the model achieves good results. Description of the Drawings

[0044] Figure 1 It is the process schematic of a sensor fault robust multi-output soft measurement method based on adversarial learning according to the present invention Figure 1 ;

[0045] Figure 2 It is the process schematic of a sensor fault robust multi-output soft measurement method based on adversarial learning according to the present invention Figure 2 ;

[0046] Figure 3 It is the process schematic of a sensor fault robust multi-output soft measurement method based on adversarial learning according to the present invention Figure 3 ;

[0047] Figure 4 It is the structure schematic of a sensor fault robust multi-output soft measurement system based on adversarial learning according to the present invention Figure 1 ;

[0048] Figure 5 It is the structure schematic of a sensor fault robust multi-output soft measurement system based on adversarial learning according to the present invention Figure 2 。 Specific embodiments

[0049] Specific embodiment 1: As shown in Figures 1 to 3 , a sensor fault robust multi-output soft measurement method based on adversarial learning, the specific steps include:

[0050] Step 1, obtain the historical data set D = {(X, Y)}, and split it into the training set D tr = {(X tr , Y tr )} and the fault-free test set D te = {(X te , Y te )}, and perform data processing on the training set and the fault-free test set; specifically including:

[0051] Step 101, process the obtained data set to obtain the training set D tr = {(X tr , Y tr )} and the fault-free test set D te = {(X te , Y te )}, where X represents the process variable, Y represents the process variable, t represents the time step, p represents the number of process variables, and k represents the number of key variables;

[0052] Step 102, obtain the training set D tr = {(Xtr , Y tr )} and the fault-free test set D te = {(X te , Y te )}, after that, perform "max-min" normalization on them respectively. The max-min normalization process is where represents the variable value after normalization, x represents the unprocessed variable value, min represents the minimum value in the variable, and max represents the maximum value in the variable;

[0053] Step 2: Input the processed X tr into the untrained soft sensor model to obtain the temporary predicted value Y 1 of the key variable, use the fast gradient sign method to generate the adversarial sample X adv , and use (X adv , Y tr ) to train the model; specifically including:

[0054] Step 201: Input the processed fault-free training data X tr into the untrained soft sensor model to obtain the predicted value Y 1 of the key variable;

[0055] Step 202: Obtain the predicted value Y 1 of the key variable, use the fast gradient sign method to generate the adversarial sample X adv ; the specific steps are:

[0056] Step 20201: Calculate the loss function L between the predicted value Y 1 and the true value Y tr , and the mean square error loss function is used;

[0057] Step 20202: Calculate the partial derivative of the loss function L with respect to X tr ;

[0058] Step 20203: Use the fast gradient sign method to generate the adversarial sample X adv ,

[0059] Step 203: Use (X adv , Y tr ) to train the soft sensor model, and the root mean square error loss function is used during the training process;

[0060] Step 3: Inject faults into the process variable data X te in the processed fault-free test set according to the sensor fault mathematical model to obtain the sensor bias fault test data X te_bias, Sensor Stuck Fault Test Data X te_stuck , Sensor Gaussian Noise Fault Test Data X te_gauss_noise , Sensor Spike Fault Test Data X te_spike ; Specifically including:

[0061] Step 301, Use the fault-free test data X te and the sensor fault mathematical model to construct the deviation sensor fault test data X te_bias , Randomly select n process variables to add faults, and add 0.3 to the normalized data of these n process variables;

[0062] Step 302, Use the fault-free test data X te and the sensor fault mathematical model to construct the stuck sensor fault test data X te_stuck , Randomly select n process variables to add faults, and add 0.3 to the normalized data of these n process variables;

[0063] Step 303, Use the fault-free test data X te and the sensor fault mathematical model to construct the Gaussian noise sensor fault test data X te_gauss_noise , Randomly select n process variables to add faults, and add Gaussian noise with a mean of 0.3 and a variance of 0.1 to the original data;

[0064] Step 304, Use the fault-free test data X te and the sensor fault mathematical model to construct the spike sensor fault test data Y pr_spike , Randomly select n process variables to add faults, randomly select some moments in these n process variables, and add 10 to the original data;

[0065] Step 4, Input the fault-free test data X te and the deviation fault test data X te_bias , the stuck fault test data X te_stuck , the Gaussian noise fault test data X te_gauss_noise , the spike fault test data X te_spike into the trained soft sensor model to obtain the predicted values Y pr 、Y pr_bias 、Y pr_stuck 、Y pr_gauss_noise 、Y pr_spike ; Specifically including:

[0066] Step 401, Input the fault-free test data X te into the trained soft sensor model to obtain the predicted value Y pr ;

[0067] Step 402: Input the deviation fault test data X te_bias into the trained soft sensor model to obtain the predicted value Y pr_bias of the key variable;

[0068] Step 403: Input the stuck fault test data X te_stuck into the trained soft sensor model to obtain the predicted value Y pr_stuck of the key variable;

[0069] Step 404: Input the Gaussian noise fault test data X te_gauss_noise into the trained soft sensor model to obtain the predicted value Y pr_gauss_noise of the key variable;

[0070] Step 405: Input the spike fault test data X te_spike into the trained soft sensor model to obtain the predicted value Y pr_spike .

[0071] Specific implementation method 2: As shown in Figure 4 and Figure 5 , a sensor fault robust multi-output soft measurement system based on adversarial learning includes an adversarial sample generation module, a spatial feature extraction module, a linear attention mechanism module, a temporal feature extraction module, and a prediction module;

[0072] The adversarial sample generation module uses the fault-free dataset D tr ={(X tr , Y tr )} as the training set, inputs the training set into the adversarial sample generation module composed of a gated recurrent unit and a fully connected layer, and generates adversarial samples X adv through the fast gradient sign method;

[0073] The spatial feature extraction module inputs the adversarial samples X adv generated by the adversarial sample generation module into the spatial feature extraction module composed of a one-dimensional convolutional neural network to obtain the feature H s containing the spatial information in the process variables;

[0074] The linear attention mechanism module inputs the feature H s containing the spatial information in the process variables into the linear attention mechanism module, so as to obtain a selection vector R of the same size as X adv . The element size of this vector represents the contribution of the corresponding process variable to the output. The larger the element value, the more important the corresponding process variable is to the prediction result. On the contrary, it means that the variable is not important to the prediction result or the sensor collecting this variable has a fault;

[0075] The time feature extraction module, which is reused with the prediction module and the adversarial sample generation module, takes the adversarial sample X generated by the adversarial sample generation module adv and performs a Hadamard product with the selection vector R generated by the linear attention mechanism module and inputs it into the time feature extraction module, thereby obtaining the feature H containing spatio-temporal information s-t and reducing the impact of sensor fault data on the prediction result;

[0076] The prediction module, which includes two sub-modules, respectively receives the feature H containing spatio-temporal information extracted by the time feature extraction module s-t , and respectively obtains the predicted values Y pr1 、Y pr2 .

[0077] Working principle

[0078] The described sensor fault robust multi-output soft measurement system based on adversarial learning consists of an adversarial sample generation module 1, a spatial feature extraction module 2, a linear attention mechanism module 3, a time feature extraction module 11, and a prediction module 12. First, in the adversarial sample generation module 1, the training set D tr ={(X tr ,Y tr )} is input into the adversarial sample generation module 1 to obtain the adversarial sample X adv ; Secondly, the adversarial sample X generated by the adversarial sample generation module is input into the spatial feature extraction module composed of a one-dimensional convolutional neural network to obtain the feature H containing the spatial information in the process variables adv ; Thirdly, the feature H containing the spatial information in the process variables is input into the linear attention mechanism module, thereby obtaining a selection vector R of the same size as X s , and the element size of this vector represents the contribution of the corresponding position process variable to the output. The larger the element value, the more important the corresponding process variable is to the prediction result. On the contrary, it means that the variable is not important to the prediction result or the sensor collecting the variable has a fault; Then, the adversarial sample X generated by the adversarial sample generation module s and the selection vector R generated by the linear attention mechanism module perform a Hadamard product and are input into the time feature extraction module, thereby obtaining the feature H containing spatio-temporal information adv and reducing the impact of sensor fault data on the prediction result; Finally, the feature H containing spatio-temporal information extracted by the time feature extraction module adv is input into the prediction module to obtain the predicted values of two key variables s-t . s-t As shown in

[0079] Figure 5 shown, the adversarial sample generation module 1 includes:

[0080] ​The time feature extraction module 11, in the adversarial sample generation stage, is used to receive the training set D tr ={(X tr , Y tr )}, so as to extract the useful information contained in the process variable X tr .

[0081] The prediction module 12, in the adversarial sample generation stage, is used to receive the features extracted by the time feature extraction module 11, so as to obtain the temporary predicted values of the key variables.

[0082] Specifically, the useful information in the process variable is extracted by the time feature extraction module 11, then the temporary predicted values of the key variables are obtained by using the prediction module 12, and then the adversarial sample X adv is generated according to the fast gradient sign method. The formula is as follows: X adv =X tr +ε·M·sgn(▽ Xadv L(Y 1 , Y tr ).

[0083] In the described sensor fault robust multi-output soft measurement system based on adversarial learning, multiple modules work together to improve the prediction accuracy in the case of sensor faults. First, the adversarial sample generation module uses the fault-free data set and the fast gradient sign method to generate adversarial samples; then, the spatial feature extraction module extracts the features containing spatial information, and then generates a selection vector through the linear attention mechanism module to evaluate the contribution of each process variable to the prediction result. Then, the adversarial sample and the selection vector are combined and input into the time feature extraction module to capture spatio-temporal information and reduce the impact of sensor faults on the prediction. Finally, the prediction module generates the predicted values of the key variables based on the spatio-temporal features. The overall system effectively improves the robustness to sensor faults through adversarial learning technology, ensuring the accuracy of multi-output soft measurement.

[0084] The above is only a preferred embodiment of the present invention, and does not impose any form of limitation on the present invention. Although the present invention has been disclosed above with a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to make equivalent changes within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention and is based on the technical essence of the present invention, any simple modification, equivalent replacement and improvement of the above embodiments within the spirit and principle of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A sensor fault robust multi-output soft-sensing method based on adversarial learning, characterized in that: The specific steps include: Step 1: Get the historical data set D = {(X, Y)} and split it into the model training set D tr ={(X tr ,Y tr )} and the fault-free test set D te ={(X te ,Y te )}, perform data processing on the training set and the fault-free test set; Step 2: The processed X tr Input into the untrained soft sensor model to obtain the temporary predicted value Y1 of the key variable, and use the fast gradient symbol method to generate adversarial samples X adv , use (X adv , Y tr ) Train the model; Step 3: The processed process variable data X in the fault-free test set te According to the sensor fault mathematical model, the fault is injected into it to obtain the sensor deviation fault test data X te_bias , Sensor stuck fault test data X te_stuck , Sensor Gaussian noise fault test data X te_gauss_noise , sensor spike fault test data X te_spike ; Step 4: Transform the fault-free test data X te and deviation fault test data X te_bias , stuck fault test data X te_stuck , Gaussian noise fault test data X te_gauss_noise , peak fault test data X te_spike Input into the trained soft sensor model to obtain the predicted value Y of the key variable pr , Y pr_bias , Y pr_stuck , Y pr_gauss_noise , Y pr_spike .

2. According to claim 1, a sensor fault robust multi-output soft-sensing method based on adversarial learning is characterized in that: Step 1 specifically includes: Step 101: Process the acquired data set to obtain the training set D of the model. tr ={(X tr ,Y tr )} and the fault-free test set D te ={(X te ,Y te )}, where X∈R t*p ,Y∈R t*k , X represents process variables, Y represents process variables, t represents time step, p represents the number of process variables, and k represents the number of key variables; Step 102: Get the training set D of the model tr ={(X tr ,Y tr )} and the fault-free test set D te ={(X te ,Y te )}, perform "maximum-minimum" normalization processing on them respectively. The maximum and minimum normalization processing process is as follows: in represents the normalized variable value, x represents the unprocessed variable value, min represents the minimum value in the variable, and max represents the maximum value in the variable.

3. According to claim 1, a sensor fault robust multi-output soft-sensing method based on adversarial learning is characterized in that: Step 2 specifically includes: Step 201: Process the fault-free training data X tr Input into the untrained soft sensor model to obtain the predicted value Y1 of the key variable; Step 202: Get the predicted value Y1 of the key variable and use the fast gradient sign method to generate adversarial samples X adv ; Step 203, use (X adv , Y tr ) is used to train the soft sensor model, and the loss function used in the training process is the root mean square error loss function.

4. The sensor fault robust multi-output soft sensing method based on adversarial learning according to claim 3 is characterized in that: In step 202, the adversarial sample X is generated using the fast gradient sign method. adv The specific steps are: Step 20201, calculate the predicted value Y1 and the true value Y tr The loss function L between them uses the mean square error loss function; Step 20202, calculate the loss function L versus X tr The partial derivative of Step 2203: Generate adversarial sample X using fast gradient sign method adv , X adv =X adv +ε·M·sgn(▽ Xadv L(Y1,Y tr )).

5. The sensor fault robust multi-output soft sensing method based on adversarial learning according to claim 1 is characterized in that: Step 3 specifically includes: Step 301: Using the fault-free test data X te and sensor fault mathematical model to build deviation sensor fault test data X te_bias , randomly select n process variables to add faults, and add 0.3 to the normalized data of these n process variables; Step 302: Using the fault-free test data X te and sensor fault mathematical model construction stuck sensor fault test data X te_stuck , randomly select n process variables to add faults, and add 0.3 to the normalized data of these n process variables; Step 303: Using the fault-free test data X te And the mathematical model of sensor failure constructs Gaussian noise sensor failure test data X te_gauss_noise , randomly select n process variables to add faults, and add Gaussian noise with a mean of 0.3 and a variance of 0.1 to the original data; Step 304: Using the fault-free test data X te and sensor fault mathematical model to construct peak sensor fault test data Y pr_spike , randomly select n process variables to add faults, randomly select some moments in these n process variables, and add 10 to the original data.

6. The sensor fault robust multi-output soft sensing method based on adversarial learning according to claim 1 is characterized in that: Step 4 specifically includes: Step 401: The fault-free test data X te Input into the trained soft sensor model to obtain the predicted value Y of the key variable pr ; Step 402: The deviation fault test data X tebias Input into the trained soft sensor model to obtain the predicted value Y of the key variable pr_bias ; Step 403: The stuck fault test data X te_stuck Input into the trained soft sensor model to obtain the predicted value Y of the key variable pr_stuck ; Step 404: Gaussian noise fault test data X te_gauss_noise Input into the trained soft sensor model to obtain the predicted value Y of the key variable pr_gauss_noise ; Step 405: The peak fault test data X te_spike Input into the trained soft sensor model to obtain the predicted value Y of the key variable pr_spike .

7. A sensor fault robust multi-output soft measurement system based on adversarial learning, characterized in that: It includes adversarial sample generation module, spatial feature extraction module, linear attention mechanism module, temporal feature extraction module and prediction module; Adversarial sample generation module, using the fault-free dataset D tr ={(X tr ,Y tr )} as the training set, and input the training set into the adversarial sample generation module composed of a gated recurrent unit and a fully connected layer, and generate adversarial samples X by the fast gradient symbol method adv ; The spatial feature extraction module extracts the adversarial sample X generated by the adversarial sample generation module. adv Input into the spatial feature extraction module composed of a one-dimensional convolutional neural network to obtain the feature H containing the spatial information of the process variable s ; The linear attention mechanism module takes the feature H containing the spatial information of the process variable s Input into the linear attention mechanism module to obtain adv The same size selection vector R, the element size of this vector represents the contribution of the process variable at the corresponding position to the output. The larger the element value, the more important the corresponding process variable is to the prediction result. Conversely, it means that the variable is not important to the prediction result or the sensor collecting the variable is faulty. The temporal feature extraction module is reused with the prediction module and the adversarial sample generation module to generate the adversarial sample X adv The selection vector R generated by the linear attention mechanism module is input into the temporal feature extraction module to obtain the feature H containing spatiotemporal information. s-t and reduce the impact of sensor fault data on prediction results; The prediction module consists of two submodules, which receive the feature H containing spatiotemporal information extracted by the time feature extraction module. s-t , and obtain the predicted values ​​Y of the two key variables respectively pr1 , Y pr2 .

Images