Double-noise auto-encoder process fault classification method

By adopting a semi-supervised dual-noise autoencoder method that integrates pseudo-labels and consistency regularization in industrial processes, the problems of noise interference in fault data and insufficient labels in industrial processes are solved, and efficient fault classification and robustness improvement are achieved , ensuring the stability of industrial production and resource security.

CN118940122APending Publication Date: 2024-11-12SHENYANG INSTITUTE OF CHEMICAL TECHNOLOGY
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202411112765.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-14
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

In industrial processes, fault data contains a large amount of irrelevant information, insufficient labels and severe noise interference, resulting in poor fault classification results. The existing deep learning methods are not robust in noise environments.

Method used

Using a semi-supervised dual-noise autoencoder (PLCR-SSDNAE) method that combines pseudo-label and consistency regularization, fault-related features are extracted through hierarchical training and layered fine-tuning, and data differences are enhanced by using Gaussian noise, combining pseudo-label and consistency Regularly adjust the model to improve robustness and classification accuracy.

Benefits of technology

Improve the accuracy and robustness of the fault classification model in the case of insufficient labels, can effectively identify the types of faults in industrial processes, reduce the impact of irrelevant information, and ensure the normal operation of industrial production and resource safety.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940122A_ABST
    Figure CN118940122A_ABST
Patent Text Reader

Abstract

The invention discloses a double-noise auto-encoder process fault classification method, and relates to an industrial process fault classification method, which comprises the following steps of: 1, deeply knowing an industrial background and an industrial process, and determining a fault type; 2, industrial process data are collected and divided; 3, preprocessing the process data; step 4, introducing two kinds of Gaussian noise into the training data, inputting a semi-supervised dual-noise auto-encoder network (PLCR-SSDNAE) fusing a pseudo tag and consistency regularization to learn fault related characteristics, and optimizing network parameters by minimizing a loss function; and 5, inputting test data into the trained model to predict a data label, and evaluating the performance of the model according to an evaluation index. During online application, data categories are judged through the fault classification model, whether faults occur in the production process and what faults occur are monitored in real time, and normal operation of the industrial process is guaranteed. The effectiveness of the method on industrial process fault classification is verified through experiments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of industrial process fault diagnosis, and specifically relates to a fault classification method for industrial processes using a dual noise autoencoder that integrates pseudo labels and consistency regularization. Background Art

[0002] In recent years, industrial processes have become increasingly complex. If a fault occurs in the production process, it is likely to cause serious losses. Therefore, while detecting the fault in a timely manner, accurately and efficiently identifying the fault type can effectively help technicians make corresponding maintenance in a timely manner to avoid accidents. A large number of data-driven fault classification methods have been successfully applied to the field of fault classification, but as process data becomes larger and more complex, with high dimensionality and nonlinearity, this has brought difficulties to fault classification methods based on statistical analysis and shallow learning. At present, many fault classification methods based on deep learning have been successfully applied to the field of process fault classification.

[0003] When a fault occurs, process data contains a lot of information, some of which is irrelevant to the fault. Process data is usually easy to collect, but data labels need to be manually labeled. Manual labeling is time-consuming and laborious, and is difficult to achieve in actual processes. In addition, industrial processes are interfered by many factors, and the generated process data contains a lot of noise, which affects the fault classification results. Autoencoder is an unsupervised learning model that can effectively extract data features and reduce dimensions, but the extracted features contain a lot of information irrelevant to the fault. The output layer of the supervised autoencoder does not reconstruct its input layer data, but predicts the fault category, so that features more relevant to the fault type can be extracted. However, supervised autoencoders (SSAE) require a large number of labels to participate in training, and it is difficult to achieve good results when there are insufficient labels. Pseudo-labeling, as a simple and effective semi-supervised learning method of deep neural networks, can use a small amount of labeled data to train the model, and then use the model to predict unlabeled data and generate labels, but this method is too dependent on the quality of labeled data and is not robust. To this end, two types of noise are introduced into the data, pseudo labels are combined with consistency regularization, and the loss function of the supervised autoencoder is improved, so that it can extract fault-related features with a small amount of labeled data, thereby improving the accuracy and robustness of the classification model. Summary of the invention

[0004] The purpose of the present invention is to propose a dual noise autoencoder process fault classification method, which is a semi-supervised dual noise autoencoder (PLCR-SSDNAE) method that integrates pseudo labels and consistency regularization. The present invention considers the problem that when a fault occurs, the process data contains a lot of information about irrelevant faults, the labels are insufficient and there is noise interference. In order to better utilize a small amount of labeled data and learn fault-related information, a stacked supervised autoencoder is used for feature extraction. The method mainly adopts the layered training and layered fine-tuning methods to train each layer of supervised autoencoders and then stack them. Preliminary training is performed using labeled data to generate pseudo labels for unlabeled data. By adding Gaussian noise to data samples to achieve weak interference and strong interference of data, the model is adjusted by combining consistency regularization with pseudo labels to improve robustness. In order to enhance the difference between process data and normal data generated by different fault types, fault feature enhancement is performed based on normal data. Online classification of process faults is realized, classification accuracy is improved, state faults are discovered in time, production efficiency is improved and property losses are avoided.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A dual noise autoencoder process fault classification method, the method mainly comprises the following steps:

[0007] S1: In-depth understanding of industrial production processes, clarifying the types of industrial process failures and related variables;

[0008] S2: Collect relevant process data based on industrial field sensors or obtain normal samples and fault samples through industrial system simulation platform;

[0009] S3: Divide the data into training set and test set;

[0010] S4: preprocess the divided data set;

[0011] S5: Perform data enhancement based on improved dissimilarity analysis to amplify the difference between fault data and normal data; introduce two types of Gaussian noise into the training data to form strong and weak interference to the data; input the labeled training data into the model, preliminarily train the network, learn fault-related features, generate pseudo labels for unlabeled data, and adjust the network by combining the pseudo labels with consistency regularization for the unlabeled data;

[0012] S6: Train the model by minimizing the loss function and optimize the model parameters;

[0013] S7: Input the test data into the trained classification model to predict the data label, determine the data category, and adjust the production process in real time according to the data category to ensure the normal progress of industrial production.

[0014] The dual noise autoencoder process fault classification method requires familiarity with the production process, fault types and data processing of the industrial process.

[0015] The dual-noise autoencoder process fault classification method, the overall steps of the classification method are divided into a training part and a testing part; using improved dissimilarity analysis for data enhancement, using PLCR-SSDNAE to extract fault-related features, reducing information on irrelevant faults, introducing two types of noise to form strong and weak interference to the data, and combining pseudo labels with consistency regularization to adjust the model to improve the robustness and classification accuracy of the model.

[0016] The dual noise autoencoder process fault classification method integrating pseudo labels and consistency regularization, wherein the training part is:

[0017] (1) Understand the industrial process and preprocess the industrial process data while dividing it into training and test sets;

[0018] (2) Improved dissimilarity analysis of the data to enhance the difference between fault data and normal data;

[0019] (3) Construct a semi-supervised dual-noise autoencoder network model, set model parameters, introduce two types of noise into the enhanced data and input it into the network to extract fault-related features and train the network, and use pseudo labels combined with consistency regularization to adjust network parameters;

[0020] (4) Obtain the optimal model parameters by minimizing the loss function and obtain the fault classification model.

[0021] The dual noise autoencoder process fault classification method, the test part is:

[0022] (1) Preprocessing and fault enhancement of test data;

[0023] (2) The enhanced test data is input into the fault classification model for label prediction, which determines the data category and the type of fault that occurred, and effectively adjusts the anomalies that occur in the production process to ensure the normal operation of industrial production.

[0024] The advantages and effects of the present invention are:

[0025] 1. The present invention is based on the stacked supervised autoencoder. A dual-noise autoencoder process fault classification method that integrates pseudo-labels and consistency regularization is proposed. The fault features are enhanced based on normal data, and the difference between process data and normal data generated by different fault types is amplified. Two types of noise are introduced into the enhanced data to enable the classification model to have a certain anti-interference ability. A small amount of labeled data is used to preliminarily train the model, learn fault-related information, predict unlabeled data, select pseudo-labels that exceed a certain threshold and adjust the model in combination with unlabeled data, so that fault-related features can be extracted for fault classification when labels are insufficient, thereby improving the learning efficiency of fault-related information and the accuracy of the classification model.

[0026] 2. In the model adjustment stage, a new loss function based on pseudo-labels and consistency regularization is designed to prevent model overfitting and improve model robustness. This prevents the model from being overly dependent on data quality and having poor anti-interference capabilities. Compared with other models, the dual-noise autoencoder network that integrates pseudo-labels and consistency regularization has higher classification accuracy when labels are insufficient.

[0027] 3. The present invention meets the needs of real-time monitoring of actual industrial production operation conditions and judging the types of industrial process faults. When used online, the present invention can directly judge the data category through the classification model, effectively adjust the anomalies that occur in the production process, ensure the normal operation of industrial production, and avoid the loss of resources and property. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the technical route of the present invention;

[0029] Figure 2 This is a diagram showing the fault classification results of the Tennessee-Eastman process of the present invention;

[0030] Figure 3 This is a diagram showing the fault classification results of the three-phase flow process of the present invention;

[0031] Figure 4 It is a schematic diagram of the overall structure of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be described in detail below with reference to the embodiments shown in the accompanying drawings.

[0033] Related term explanation: SSAE: Stacked Supervised Autoencoder; Pseudo-labeling: Pseudo-labeling method; Consistency Regularization: Consistency Regularization; PLCR-SSDNAE: Semi-supervised double noise autoencoder that combines pseudo-labeling and consistency regularization.

[0034] The present invention designs a dual-noise autoencoder network that integrates pseudo labels and consistency regularization to solve the problem that process data in industrial processes contain a lot of irrelevant fault information, insufficient labels and noise interference, and a fault classification method for determining the category of process data. It includes the following steps:

[0035] S1: Master the theoretical knowledge and operation procedures of industrial processes, and identify the types of industrial process failures and related variables;

[0036] S2: Collect and organize relevant process data collected by industrial field sensors or obtain normal samples and fault samples through industrial system simulation platform;

[0037] S3: Divide the collected data into training set and test set;

[0038] S4: preprocess the training data and test data;

[0039] S5: Perform data enhancement based on improved dissimilarity analysis to amplify the difference between fault data and normal data; introduce two types of Gaussian noise into the training data to form strong and weak interference to the data; input the labeled training data into the model, preliminarily train the network, learn fault-related features, generate pseudo labels for unlabeled data, and adjust the network by combining the pseudo labels with consistency regularization for the unlabeled data;

[0040] S6: Train the model and optimize the model parameters by minimizing the loss function and calculating the two model evaluation indicators, precision (P) and recall (R);

[0041] S7: Input the preprocessed test data into the trained classification model to predict the data label, determine the data category, and adjust the production process in real time according to the data category to ensure the normal progress of industrial production.

[0042] As a further improvement, the present invention uses the improved dissimilarity analysis method described in step S5 to perform data enhancement, introduces two types of Gaussian noise to form strong and weak interference to the data, and preliminarily trains the network to learn fault-related information, predict unlabeled enhanced samples, and generate pseudo labels. In the adjustment stage, the network is adjusted layer by layer using a method combined with pseudo-label consistency regularization, and finally the coding layers are stacked to extract fault-related features to improve classification accuracy and robustness. The specific steps are as follows:

[0043] S51 improved data enhancement method only takes N normal samples X in the training set nor N×M , find its covariance matrix R, perform eigenvalue decomposition on R, and obtain the transformation matrix P. Using the transformation matrix P as the basis, map all samples in the training set and the test set. The specific formula is as follows:

[0044]

[0045]

[0046] S52: Supervise the autoencoder layer by layer, save the relevant parameters, extract features from the input samples, that is, the hidden layer variable h, and the output of the kth layer is:

[0047]

[0048] S53: The loss function of the kth layer is:

[0049]

[0050] S54: In the adjustment phase, the trained SSAE is used to predict the unlabeled sample xuw with weak Gaussian noise added after fault enhancement. Each layer will generate the prediction probability for each category Take the index of the maximum value of the predicted probability As pseudo labels. Pseudo labels with confidence exceeding a certain threshold τ are selected and combined with samples xus with strong Gaussian noise added, and hierarchical adjustments are made based on the previously trained weight parameters. At this time, the loss function combined with pseudo label consistency regularization can be expressed as:

[0051]

[0052] Among them, α is the indicator function, which means the predicted probability of the pseudo label When it exceeds a certain threshold τ, the value is 1, otherwise it is 0. τ is a scalar hyperparameter, which is the threshold for screening pseudo-labels. Only when the threshold is exceeded, the pseudo-label is included in the loss calculation. When the pseudo-label threshold does not exceed τ, the loss is 0 and does not participate in the model adjustment. H represents the cross entropy loss. When adjusting, first extract features to train the next layer, and then fine-tune the current layer to prevent the prediction error from gradually accumulating and seriously affecting the classification results. Only pseudo-labels that exceed the threshold and the corresponding sample information are used, which greatly reduces the deviation of the model caused by inaccurate prediction of some data. Through the adjustment of consistency regularization, the model is made more robust to prevent overfitting.

[0053] S61: The present invention uses an improved loss function training model, integrates pseudo labels and consistency regularization into the guided feature learning process, optimizes model parameters, and improves the accuracy and robustness of the classification model. The specific process is as follows:

[0054]

[0055] As the instruction manual Figure 4The method of dual noise autoencoder process fault classification based on pseudo label and consistency regularization is shown. The method realizes fault classification of industrial process based on pseudo label, consistency regularization and stacked supervised autoencoder network. In order to intuitively evaluate the performance of the model, the present invention uses two indicators, precision (P) and recall (R), to quantify the prediction effect of the model. Formulas (8) and (9) define P and R:

[0056]

[0057] Among them, TP represents the correctly classified samples belonging to this category, FP represents the samples that do not belong to this category are incorrectly classified as this category, and FN represents the samples that belong to this category but are incorrectly classified as other categories. Generally, the larger P and R are, the better the performance of the model.

[0058] The fault classification process of the present invention is as follows:

[0059] Training phase:

[0060] (1) Data standardization: z-score standardization was used.

[0061] (2) For all standardized sample data, fault enhancement is performed based on normal samples.

[0062] (3) Add the same weak Gaussian noise to the labeled training samples and the unlabeled training samples, and add an additional strong Gaussian noise to the unlabeled training samples.

[0063] (4) The supervised AE model is trained hierarchically with labeled training samples. The features extracted by each layer of the model are used as the input of the next layer. After the training is completed layer by layer, the model parameters are saved, and the model is used to predict the category of the unlabeled training samples with weak Gaussian noise added to generate prediction results.

[0064] (5) When the predicted probability of the predicted label exceeds a certain threshold τ, the pseudo-label is saved, and the model is trained again with the unlabeled training samples corresponding to the pseudo-label and the unsupervised loss to adjust the model.

[0065] (6) Use the Adam optimizer and loss function to fine-tune the network parameters. After training, save the parameters, stack the encoding layers of each layer, and use the output layer of the last layer as the final output of the network.

[0066] Testing phase:

[0067] The same method is used in the test phase and the training phase for data preprocessing and fault enhancement, and the trained PLCR-SSDNAE model is used to obtain the classification results. The precision and recall evaluation indicators are calculated to verify the effect of the model.

[0068] [Example 1]

[0069] The TE process is a simulation system developed by Eastman Chemical Company based on real chemical processes, which is mainly used to test the effectiveness of process control and monitoring methods. The present invention selects 52 process measurement variables, and the training data contains 500 normal samples, a total of 18 types of faults, and each fault contains 480 samples. 200 samples are randomly selected from each type of data as labeled data for training, and the rest are unlabeled data. The number of normal samples in the test data is 3840, with a total of 18 types of faults, and 800 samples of each type of fault.

[0070] As can be seen from Table 1, the PLCR-SSDNAE fault classification method has the highest classification accuracy for all faults except faults 6, 7 and 18, with an average accuracy of 74.19%. Figure 2 The value in the middle represents the recall rate. The SSAE method is obviously more chaotic than PLCR-SSDNAE, with more misclassified samples and higher false recognition rate. The robustness of the present invention is stronger. The proposed method improves faults 5 and 10 particularly significantly. Overall, the proposed PLCR-SSDNAE fault classification method has better effect.

[0071] Table 1: Comparison of TE process classification accuracy

[0072]

[0073]

[0074] [Example 2]

[0075] The three-phase flow facility at Cranfield University provides a controllable and measurable flow rate of water, oil and air for a pressurized system. It can provide single-phase air, water and oil, or a mixture of these fluids at a required rate. The present invention selects 24 variables, and selects 4 fault type data and 1 normal type data for fault classification when the air and water flow rates vary. The detailed introduction of each fault type can be found in the literature.

[28] Found in . 200 samples are randomly selected from each category of the labeled training set data, 400 samples are selected from each category of the unlabeled training set, and 1000 samples are selected from each category of the test set, for a total of 5000 samples.

[0076] Table 2: Comparison of classification accuracy of three-phase flow processes

[0077]

[0078] In Table 2 and the instruction manual Figure 3It can be seen that for the four faults, the present invention has achieved the best results, with an overall average accuracy of 98.83%. The confusion matrix also shows the effectiveness of the proposed method, with the most obvious improvement in the accuracy of fault 2 and fault 4. The other three methods have a certain probability of confusing fault 2, fault 4 and fault 1, and PLCR-SSDNAE significantly reduces this misidentification.

[0079] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the embodiments. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A dual-noise autoencoder process fault classification method, characterized in that: The method comprises the following steps: S1: Deepen the industrial production process and identify the types of industrial process failures and related variables; S2: Collect relevant process data based on industrial field sensors or obtain normal samples and fault samples through industrial system simulation platform; S3: Divide the data into training set and test set; S4: preprocess the divided data set; S5: Perform data enhancement based on improved dissimilarity analysis to amplify the differences between faulty and normal data. Two types of Gaussian noise are introduced into the training data to create strong and weak interference. Labeled training data is fed into the model to initially train the network and learn fault-related features. Pseudo-labels are generated for unlabeled data. The unlabeled data is then combined with the pseudo-labels and consistency regularization to adjust the network. S6: Train the model by minimizing the loss function and optimizing the model parameters; S7: Input the test data into the trained classification model to predict the data label, determine the data category, and adjust the production process in real time according to the data category to ensure the normal progress of industrial production.

2. A dual-noise autoencoder process fault classification method according to claim 1, characterized in that: This method is familiar with the production process, fault types and data processing of industrial processes; The overall steps of the classification method are training and testing. Improved dissimilarity analysis is used for data enhancement. PLCR-SSDNAE is used to extract fault-related features and reduce information about irrelevant faults. Two types of noise are introduced to form strong and weak interference on the data. Pseudo-labels are combined with consistency regularization to adjust the model and improve its robustness and classification accuracy.

3. The dual noise autoencoder process fault classification method according to claim 1, characterized in that: The training part is: (1) Understand the industrial process and preprocess the industrial process data while dividing it into training and test sets; (2) Improved dissimilarity analysis of the data to enhance the difference between fault data and normal data; (3) Construct a semi-supervised dual-noise autoencoder network model, set the model parameters, introduce two types of noise into the enhanced data and input it into the network to extract fault-related features and train the network, and use pseudo labels combined with consistency regularization to adjust the network parameters; (4) Obtain the optimal model parameters by minimizing the loss function and obtain the fault classification model.

4. The dual noise autoencoder process fault classification method according to claim 1, characterized in that: The test parts are: (1) Preprocessing and fault enhancement of test data; (2) The enhanced test data is input into the fault classification model for label prediction, which determines the data category and the type of fault that has occurred, and effectively adjusts the anomalies that occur in the production process to ensure the normal operation of industrial production.

Citation Information

Cited By

  • Self-supervision enhanced semi-supervised wafer failure prediction method and system

    CN119557571A

  • A self-supervised enhanced semi-supervised wafer failure prediction method and system

    CN119557571B

  • Multi-scene lithium ion battery fault diagnosis method and system

    CN120703590A