Sensor fault robust multi-output soft-sensing method and system based on adversarial learning
By generating adversarial examples and a feature extraction module through adversarial learning, a multi-output soft measurement system is constructed, which solves the problem of multivariate prediction under the influence of sensor failure and improves the robustness and prediction accuracy of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-01-21
- Publication Date
- 2026-06-02
AI Technical Summary
Existing soft measurement models are mostly single-output models, which cannot meet the needs of simultaneous prediction of multiple variables. Furthermore, the prediction accuracy decreases when the sensor fails, and the cost of collecting sensor failure data is high.
An adversarial learning-based approach is adopted to generate adversarial examples from fault-free data. Combined with spatial and temporal feature extraction modules, a multi-output soft measurement system is constructed. Adversarial examples are used to replace sensor fault data for model training, thereby enhancing the robustness of the model.
It enables multivariate prediction in the event of sensor failure, improves prediction accuracy and robustness, and reduces the impact of sensor failure on prediction results.
Smart Images
Figure CN120067531B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a sensor fault robust multi-output soft measurement method and system, belonging to the field of soft measurement technology. Background Technology
[0002] In industrial production, key variables typically refer to those closely related to product quality but difficult to measure directly. Accurate measurement of these key variables is crucial for improving product quality, optimizing production processes, and ensuring operator safety. Soft measurement technology constructs mathematical models, using easily measurable process variables as input and difficult-to-measure key variables as output, to predict these key variables. Based on their construction methods, soft measurement models can be divided into two categories: knowledge-driven soft measurement models and data-driven soft measurement models. Knowledge-driven methods rely on a deep understanding of the mechanisms of the production process and are suitable for production environments with relatively clear mechanisms. However, in complex production processes where the mechanisms are not yet fully understood, the applicability of knowledge-driven methods is limited. In contrast, data-driven soft measurement models are widely used in industrial production due to their ease of data acquisition and lower requirements for knowledge of the production process mechanisms. From the perspective of the number of predictable variables, soft measurement models can be divided into single-output soft measurement models and multi-output soft measurement models. Current research focuses primarily on single-output models, i.e., models that predict only one key variable at a time. However, many industrial production processes require the simultaneous monitoring of multiple key variables, a requirement that single-output models struggle to meet. Therefore, multi-output soft measurement models become particularly important. On the other hand, data-driven soft measurement models typically rely on data collected by hardware sensors. When some sensors fail, the input data may become biased, leading to errors in the soft measurement model's predictions of key variables and consequently impacting the production process. While the robustness of models can be enhanced by training them with sensor failure data, collecting sufficient and applicable failure data is extremely difficult due to the diverse and unpredictable nature of sensor failures. Therefore, developing a multi-output soft measurement method and system that is robust to sensor failures without requiring sensor failure data is a pressing problem.
[0003] Since most existing soft measurement models are single-output soft measurement models, and their predictive performance is easily affected by sensor failures, they have the following drawbacks:
[0004] (1) Unable to predict multiple variables simultaneously
[0005] Most current data-driven soft measurement models focus on single-output problems, meaning they can only predict one key variable at a time. This fails to meet practical needs in many industrial scenarios. For example, in penicillin production, multiple key variables (such as penicillin concentration and biomass concentration) need to be monitored simultaneously. Single-output models cannot predict multiple variables at the same time, resulting in an inability to comprehensively and accurately assess product quality at any given moment. Therefore, existing technologies have a deficiency in meeting the need for simultaneous prediction of multiple variables.
[0006] (2) When hardware sensors malfunction, the prediction accuracy of soft measurement models decreases.
[0007] In actual industrial production, hardware sensor failures are common and varied in type. Current data-driven soft measurement models typically assume that sensor data is intact, neglecting the data deviations that sensor failures may introduce. Sensor failures can cause deviations in the model's input data, thereby affecting prediction results and reducing the stability and safety of the production process. Although some methods have attempted to address this issue, due to the unpredictability of the type and timing of failures, existing technologies have not yet provided an effective and universally applicable robust solution for sensor failures.
[0008] (3) Collecting suitable sensor fault training sets is costly.
[0009] Using sensor fault data to train a model can enhance its robustness to sensor faults, but due to the wide variety and unpredictability of sensor faults, collecting sufficient and applicable fault data requires high costs. Summary of the Invention
[0010] To address the problem that most existing soft measurement models are single-output soft measurement models and their predictive performance is easily affected by sensor faults, this invention proposes a sensor fault 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 the sensor fault robust multi-output soft measurement method based on adversarial learning described in the present invention include:
[0012] Step 1: Obtain the historical dataset D = {(X,Y)} and split it into the model's training set D. tr ={(X tr ,Y tr )} and fault-free test set D te ={(X te ,Y te Data processing is performed on the training set and the fault-free test set.
[0013] Step 2: Process the X trInputting the data into an untrained soft sensor model yields provisional predicted values Y1 for key variables. Adversarial samples X are then generated using the fast gradient sign method. adv , using (X adv Y tr Train the model;
[0014] Step 3: Transfer the process variable data X from the processed fault-free test set. te Based on the mathematical model of sensor faults, a fault is injected to obtain 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 ;
[0015] Step 4: Transfer 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 The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr Y pr_bias Y pr_stuck Y pr_gauss_noise Y pr_spike .
[0016] Further, in step 101, the acquired dataset is processed to obtain the training set D of the model. tr ={(X tr ,Y tr )} and fault-free test set D te ={(X te ,Y te )},in X represents process variables, Y represents process variables, t represents time steps, p represents the number of process variables, and k represents the number of key variables;
[0017] Step 102: Obtain the training set D of the model. tr ={(X tr ,Y tr )} and fault-free test set D te ={(X te ,Y te After that, perform "maximum-min" normalization on each of them. The maximum-min normalization process is as follows: in represents the normalized variable value, x represents the unprocessed variable value, min represents the minimum value of the variable, and max represents the maximum value of the variable.
[0018] Furthermore, step 2 specifically includes:
[0019] Step 201: Process the fault-free training data X tr Input into an untrained soft sensor model to obtain the predicted value Y1 of the key variable;
[0020] Step 202: Obtain the predicted value Y1 of the key variable, and use the fast gradient sign method to generate adversarial samples X. adv ;
[0021] Step 203, use (X) adv Y tr The soft measurement model is trained using the root mean square error loss function.
[0022] Furthermore, in step 202, the fast gradient sign method is used to generate adversarial samples X. adv The specific steps are as follows:
[0023] Step 20201: Calculate the predicted value Y1 and the true value Y. tr The loss function L used is the mean squared error loss function.
[0024] Step 20202: Calculate the loss function L against X. tr The partial derivatives;
[0025] Step 20203: Generate adversarial examples X using the fast gradient sign method. adv ,
[0026] Furthermore, step 3 specifically includes:
[0027] Step 301: Utilize fault-free test data X te Mathematical model of sensor faults and 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: Utilize fault-free test data X te The construction of a mathematical model for sensor faults is stuck, hindering the testing of sensor fault 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: Utilize fault-free test data X teConstructing Gaussian noise sensor fault test data X using a mathematical model of sensor faults 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: Utilize fault-free test data X te Constructing a mathematical model of sensor faults and peak sensor fault test data Y pr_spike Randomly select n process variables to add faults, and randomly select some time points from each of these n process variables, adding 10 to the original data.
[0031] Furthermore, step 4 specifically includes:
[0032] Step 401: Transfer the fault-free test data X te The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr ;
[0033] Step 402: Transfer the deviation fault test data X te_bias The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr_bias ;
[0034] Step 403: Transfer the stuck fault test data X te_stuck The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr_stuck ;
[0035] Step 404: Transfer the Gaussian noise fault test data X te_gauss_noise The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr_gauss_noise ;
[0036] Step 405: Transfer the peak fault test data X te_spike The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr_spike .
[0037] The sensor fault robust multi-output soft measurement system based on adversarial learning described in this invention includes an adversarial example generation module, a spatial feature extraction module, a linear attention mechanism module, a temporal feature extraction module, and a prediction module;
[0038] The adversarial example generation module uses the fault-free dataset D tr ={(X tr ,Y tr The training set is fed into an adversarial example generation module consisting of gated recurrent units and fully connected layers, and adversarial examples X are generated using the fast gradient sign method. adv ;
[0039] The spatial feature extraction module extracts the adversarial sample X generated by the adversarial sample generation module. adv The input is fed into a spatial feature extraction module composed of a one-dimensional convolutional neural network to obtain feature H containing spatial information from process variables. s ;
[0040] The linear attention mechanism module will incorporate features H containing spatial information from process variables. s The input is fed into the linear attention mechanism module to obtain the result related to X. adv A selection vector R of the same size. The size of the elements of this vector represents the contribution of the corresponding process variable to the output. The larger the value of the element, the more important the corresponding process variable is to the prediction result. Conversely, the smaller the value, the less important the variable is to the prediction result or the sensor collecting the variable is malfunctioning.
[0041] The temporal feature extraction module, along with the prediction module, is reused with the adversarial example generation module. This module extracts the adversarial example X generated by the adversarial example generation module. adv The Hadam product of the selection vector R generated by the linear attention mechanism module and the input to the temporal feature extraction module yields the feature H containing spatiotemporal information. s-t And reduce the impact of sensor failure data on prediction results;
[0042] The prediction module consists of two sub-modules, which receive features H containing spatiotemporal information extracted by the temporal feature extraction module. s-t The predicted values Y of the two key variables were obtained respectively. pr1 Y pr2 .
[0043] The beneficial effects of this invention are as follows: This invention consists of three parts: an adversarial example generation stage, a training stage, and a multivariate prediction stage. Compared with traditional single-output soft measurement methods, the method of this invention incorporates an adversarial example generation module, which generates adversarial examples using easily collected fault-free data, thereby replacing sensor fault data for model training. This solves the problem of difficult sensor fault data collection and effectively improves the model's robustness 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, this invention effectively mines the spatiotemporal information in process variables and reduces the adverse impact of sensor fault data on prediction results, thereby improving the model's prediction performance for key variables and its robustness to sensor faults. By adding a prediction module containing two sub-modules to the proposed method, this invention has the ability to predict multiple key variables simultaneously. Finally, by adjusting the model's hyperparameters and loss function, the model achieves good results. Attached Figure Description
[0044] Figure 1 This is a flowchart illustrating a sensor fault robust multi-output soft measurement method based on adversarial learning as described in this invention. Figure 1 ;
[0045] Figure 2 This is a flowchart illustrating a sensor fault robust multi-output soft measurement method based on adversarial learning as described in this invention. Figure 2 ;
[0046] Figure 3 This is a flowchart illustrating a sensor fault robust multi-output soft measurement method based on adversarial learning as described in this invention. Figure 3 ;
[0047] Figure 4 This is a schematic diagram of the structure of a sensor fault robust multi-output soft measurement system based on adversarial learning as described in this invention. Figure 1 ;
[0048] Figure 5 This is a schematic diagram of the structure of a sensor fault robust multi-output soft measurement system based on adversarial learning as described in this invention. Figure 2 . Detailed Implementation
[0049] Specific implementation method one: as follows Figures 1 to 3 As shown, a sensor fault robust multi-output soft measurement method based on adversarial learning includes the following steps:
[0050] Step 1: Obtain the historical dataset D = {(X,Y)} and split it into the model's training set D. tr ={(X tr ,Y tr )} and fault-free test set D te ={(X te ,Y te Data processing is performed on the training set and the fault-free test set; specifically including:
[0051] Step 101: Process the acquired dataset to obtain the training set D for the model. tr ={(X tr ,Y tr )} and fault-free test set D te ={(X te ,Y te )},in X represents process variables, Y represents process variables, t represents time steps, p represents the number of process variables, and k represents the number of key variables;
[0052] Step 102: Obtain the training set D of the model. tr ={(X tr ,Y tr)} and fault-free test set D te ={(X te ,Y te After that, perform "maximum-min" normalization on each of them. The maximum-min normalization process is as follows: in represents the normalized variable value, x represents the unprocessed variable value, min represents the minimum value of the variable, and max represents the maximum value of the variable;
[0053] Step 2: Process the X tr Inputting the data into an untrained soft sensor model yields provisional predicted values Y1 for key variables. Adversarial samples X are then generated using the fast gradient sign method. adv , using (X adv Y tr Training the model; specifically including:
[0054] Step 201: Process the fault-free training data X tr Input into an untrained soft sensor model to obtain the predicted value Y1 of the key variable;
[0055] Step 202: Obtain the predicted value Y1 of the key variable, and use the fast gradient sign method to generate adversarial samples X. adv The specific steps are as follows:
[0056] Step 20201: Calculate the predicted value Y1 and the true value Y. tr The loss function L used is the mean squared error loss function.
[0057] Step 20202: Calculate the loss function L against X. tr The partial derivatives;
[0058] Step 20203: Generate adversarial examples X using the fast gradient sign method. adv ,
[0059] Step 203, use (X) adv Y tr The soft measurement model is trained using the root mean square error loss function.
[0060] Step 3: Transfer the process variable data X from the processed fault-free test set. te Based on the mathematical model of sensor faults, a fault is injected to obtain 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_spikeSpecifically, it includes:
[0061] Step 301: Utilize fault-free test data X te Mathematical model of sensor faults and 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: Utilize fault-free test data X te The construction of a mathematical model for sensor faults is stuck, hindering the testing of sensor fault 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: Utilize fault-free test data X te Constructing Gaussian noise sensor fault test data X using a mathematical model of sensor faults 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: Utilize fault-free test data X te Constructing a mathematical model of sensor faults and peak sensor fault test data Y pr_spike Randomly select n process variables to add faults, and randomly select some time points from each of these n process variables, adding 10 to the original data;
[0065] Step 4: Transfer 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 The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr Y pr_bias Y pr_stuck Y pr_gauss_noise Y pr_spike Specifically, it includes:
[0066] Step 401: Transfer the fault-free test data X te The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr ;
[0067] Step 402: Transfer the deviation fault test data X te_bias The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr_bias ;
[0068] Step 403: Transfer the stuck fault test data X te_stuck The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr_stuck ;
[0069] Step 404: Transfer the Gaussian noise fault test data X te_gauss_noise The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr_gauss_noise ;
[0070] Step 405: Transfer the peak fault test data X te_spike The input is fed into the trained soft sensor model to obtain the predicted values Y of the key variables. pr_spike .
[0071] Specific implementation method two: such as Figure 4 and Figure 5 As shown, a sensor fault robust multi-output soft measurement system based on adversarial learning includes an adversarial example generation module, a spatial feature extraction module, a linear attention mechanism module, a temporal feature extraction module, and a prediction module.
[0072] The adversarial example generation module uses the fault-free dataset D tr ={(X tr ,Y tr The training set is fed into an adversarial example generation module consisting of gated recurrent units and fully connected layers, and adversarial examples X are generated using the fast gradient sign method. adv ;
[0073] The spatial feature extraction module extracts the adversarial sample X generated by the adversarial sample generation module. adv The input is fed into a spatial feature extraction module composed of a one-dimensional convolutional neural network to obtain feature H containing spatial information from process variables. s ;
[0074] The linear attention mechanism module will incorporate features H containing spatial information from process variables. s The input is fed into the linear attention mechanism module to obtain the result related to X. adv A selection vector R of the same size. The size of the elements of this vector represents the contribution of the corresponding process variable to the output. The larger the value of the element, the more important the corresponding process variable is to the prediction result. Conversely, the smaller the value, the less important the variable is to the prediction result or the sensor collecting the variable is malfunctioning.
[0075] The temporal feature extraction module, along with the prediction module, is reused with the adversarial example generation module. This module extracts the adversarial example X generated by the adversarial example generation module. adv The Hadam product of the selection vector R generated by the linear attention mechanism module and the input to the temporal feature extraction module yields the feature H containing spatiotemporal information.s-t And reduce the impact of sensor failure data on prediction results;
[0076] The prediction module consists of two sub-modules, which receive features H containing spatiotemporal information extracted by the temporal feature extraction module. s-t The predicted values Y of the two key variables were obtained respectively. pr1 Y pr2 .
[0077] Working principle
[0078] The aforementioned sensor fault robust multi-output soft measurement system based on adversarial learning comprises an adversarial example generation module 1, a spatial feature extraction module 2, a linear attention mechanism module 3, a temporal feature extraction module 11, and a prediction module 12. First, in the adversarial example generation module 1, the training set D... tr ={(X tr ,Y tr The input is given to 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 adv The input is fed into a spatial feature extraction module composed of a one-dimensional convolutional neural network to obtain feature H containing spatial information from process variables. s ; Again, the feature H containing spatial information in the process variables will be included. s The input is fed into the linear attention mechanism module to obtain the result related to X. adv A selection vector R of the same size is used, where the size of each element represents the contribution of the corresponding process variable to the output. A larger element value indicates a more important process variable to the prediction result, while a smaller value indicates that the variable is unimportant or that the sensor collecting the variable is malfunctioning. Then, adversarial examples X generated by the adversarial example generation module are used. adv The Hadam product of the selection vector R generated by the linear attention mechanism module and the input to the temporal feature extraction module yields the feature H containing spatiotemporal information. s-t And reduce the impact of sensor fault data on prediction results; finally, the features H containing spatiotemporal information extracted by the time feature extraction module are used. s-t The predicted values of two key variables are obtained by inputting them into the prediction module.
[0079] like Figure 5 As shown, the adversarial example generation module 1 includes:
[0080] Temporal feature extraction module 11, during the adversarial example generation stage, is used to receive the training set D. tr ={(X tr ,Y tr This allows us to extract the process variable X. tr It contains useful information.
[0081] In the adversarial example generation stage, the prediction module 12 receives features extracted by the temporal feature extraction module 11 to obtain temporary predicted values of key variables.
[0082] Specifically, the time feature extraction module 11 extracts useful information from process variables, then the prediction module 12 obtains temporary predicted values of key variables, and finally, adversarial examples X are generated using the fast gradient sign method. adv The formula is as follows: X adv =X tr +ε·M·sgn(▽ Xadv L(Y1,Y tr )).
[0083] The aforementioned sensor fault robust multi-output soft measurement system based on adversarial learning improves prediction accuracy under sensor fault conditions through the collaborative work of multiple modules. First, an adversarial example generation module generates adversarial examples using a fault-free dataset and a fast gradient sign method. Next, a spatial feature extraction module extracts features containing spatial information, followed by a linear attention mechanism module that generates a selection vector to evaluate the contribution of each process variable to the prediction result. Then, the adversarial examples and selection vector are combined and input into a temporal feature extraction module to capture spatiotemporal information and reduce the impact of sensor faults on the prediction. Finally, a prediction module generates predicted values for key variables based on the spatiotemporal features. The overall system effectively improves robustness to sensor faults through adversarial learning technology, ensuring the accuracy of multi-output soft measurements.
[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.
Claims
1. A sensor fault-robust multi-output soft measurement method based on adversarial learning, characterized in that, The system for implementing the measurement method includes an adversarial example generation module, a spatial feature extraction module, a linear attention mechanism module, a temporal feature extraction module, and a prediction module; The adversarial example generation module uses a fault-free dataset. As a training set, the training set is input into an adversarial example generation module consisting of gated recurrent units and fully connected layers, and adversarial examples are generated using the fast gradient sign method. ; The spatial feature extraction module extracts adversarial examples generated by the adversarial example generation module. The input is fed into a spatial feature extraction module composed of a one-dimensional convolutional neural network to obtain features containing spatial information from process variables. ; The linear attention mechanism module incorporates features containing spatial information from process variables. The input is fed into the linear attention mechanism module to obtain the result. Selection vectors of the same size The size of the elements in 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. Conversely, the smaller the value, the less important the variable is to the prediction result or the sensor collecting the variable is malfunctioning. The temporal feature extraction module, along with the prediction module, is reused with the adversarial example generation module. This module extracts the adversarial examples generated by the adversarial example generation module. The selection vector generated by the linear attention mechanism module The Hadam product is then input into the temporal feature extraction module to obtain features containing spatiotemporal information. And reduce the impact of sensor failure data on prediction results; The prediction module consists of two sub-modules, which receive features containing spatiotemporal information extracted by the temporal feature extraction module. The predicted values of the two key variables were obtained respectively. , ; The steps of the sensor fault robust multi-output soft measurement method based on adversarial learning include: Step 1: Obtain historical dataset Split into the training set of the model and fault-free test set Data processing is performed on the training set and the fault-free test set; specifically including: Step 101: Process the acquired dataset to obtain the training set for the model. and fault-free test set ,in , , Represents process variables, Represents process variables, Represents time step, Represents the number of process variables, Represents the number of key variables; Step 102: Obtain the training set of the model. and fault-free test set Then, perform "maximum-min" normalization on each of them. The maximum-min normalization process is as follows: ,in This represents the normalized variable value. Represents the unprocessed variable value. Represents the minimum value among variables. Represents the maximum value among variables; Step 2: Process the... Input into an untrained soft sensor model to obtain provisional predicted values for key variables. Generate adversarial samples using the fast gradient sign method. ,use( , Training the model; specifically including: Step 201: Process the fault-free training data Input into an untrained soft sensor model to obtain predicted values for key variables. ; Step 202: Obtain the predicted values of key variables. Generate adversarial samples using the fast gradient sign method The specific steps are as follows: Step 20201: Calculate the predicted value with truth value loss function between The mean squared error loss function is used. Step 20202: Calculate the loss function right The partial derivatives; Step 20203: Generate adversarial examples using the fast gradient sign method. , Step 203, use ( , The soft measurement model is trained using the root mean square error loss function. Step 3: Process variable data from the processed fault-free test set. Based on the mathematical model of sensor faults, faults are injected to obtain sensor deviation fault test data. Sensor stuck fault test data Sensor Gaussian noise fault test data Sensor spike fault test data Specifically, it includes: Step 301: Utilize fault-free test data Construction of Mathematical Model for Sensor Faults and Deviation Sensor Fault Test Data Randomly select n process variables to add faults, and add 0.3 to the normalized data of these n process variables; Step 302: Utilize fault-free test data The construction of mathematical models for sensor faults is stuck, hindering the testing of sensor fault data. Randomly select n process variables to add faults, and add 0.3 to the normalized data of these n process variables; Step 303: Utilize fault-free test data Constructing Gaussian noise sensor fault test data using a mathematical model of sensor faults 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: Utilize fault-free test data Constructing peak sensor fault test data using mathematical models of sensor faults Randomly select n process variables to add faults, and randomly select some time points from each of these n process variables, adding 10 to the original data; Step 4: Transfer the fault-free test data and deviation fault test data , stuck fault test data Gaussian noise fault test data Peak Fault Test Data The values are input into the trained soft sensor model to obtain the predicted values of the key variables. , , , , Specifically, it includes: Step 401: Transfer the fault-free test data The values are input into the trained soft sensor model to obtain the predicted values of the key variables. ; Step 402: Transfer the deviation fault test data The values are input into the trained soft sensor model to obtain the predicted values of the key variables. ; Step 403: Obtain the stuck fault test data The values are input into the trained soft sensor model to obtain the predicted values of the key variables. ; Step 404: Transfer Gaussian noise fault test data The values are input into the trained soft sensor model to obtain the predicted values of the key variables. ; Step 405: Transfer the peak fault test data The values are input into the trained soft sensor model to obtain the predicted values of the key variables. .