Industrial process fault classification method adaptive to multi-source semi-supervised condition constraint domain

Abstract

The invention discloses an industrial process fault classification method adaptive to a multi-source semi-supervised condition constraint domain, and relates to an industrial process soft measurement method. Collecting different working condition data and constructing a source domain and target domain data set; preprocessing the data set through maximum and minimum normalization; storing the classifier model; in the testing stage, target domain data is used for testing, classification is carried out through the stored feature extractor model and classifier model, the generalization ability and classification precision of the model are improved, the marking cost of faults in the industrial process is reduced, and the stability and reliability of the model are improved. The system is integrated into an industrial control system, and real-time monitoring and fault diagnosis of equipment and a process are realized. By continuously learning and adapting to new data, the model can maintain efficient fault classification capability, and safe and stable operation of the industrial process is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the field of industrial process fault classification, and specifically relates to a multi-source domain semi-supervised domain adaptation (MSSCCDA) method, which aligns target domain features through more comprehensive information from multi-source domains and an adversarial domain adaptation method to achieve feature migration, so as to facilitate classification of unlabeled target domain data. Background Art

[0002] In modern industrial processes, industrial process fault classification is an important research field in modern industrial systems, which aims to identify and classify various potential faults by analyzing and processing data from industrial processes. Industrial processes usually involve complex mechanical, electrical and chemical systems, which may fail during operation due to various reasons, such as equipment aging, operating errors, environmental changes, etc. Timely and accurate identification and classification of these faults is of great significance to ensure the safety, reliability and economy of industrial systems.

[0003] Fault classification is a key step in fault diagnosis. By classifying faults into different types, targeted guidance can be provided for further processing and repair of faults. Traditional fault classification methods rely on expert knowledge and experience, and usually adopt rule-based and model-based methods. These methods can achieve good results in specific application scenarios. However, with the increasing complexity of industrial processes, these methods are unable to cope with large-scale, high-dimensional, and multivariate data and are difficult to meet actual needs. In recent years, the rise of data-driven methods, especially data-driven methods based on deep learning, has received widespread attention and application in the field of fault classification of industrial processes due to its powerful feature extraction capabilities and adaptability. However, in general data-driven fault classification methods, it is assumed that the training data extracted from the source domain and the test data extracted from the target domain should have the same feature space distribution. This assumption is often challenged in actual industrial applications. Considering many factors, such as changes in set points, different operating conditions, mechanical wear, and personnel operation methods, data sets collected under different working conditions usually have different feature distributions, which is called the multi-condition problem. At the same time, in actual industrial applications, each fault cannot obtain enough labels, which also makes it difficult to establish a fault identification model.

[0004] In view of the above problems, in order to be applied to complex industrial processes, the present invention implements fault classification in a method of using multi-source domain feature extraction and semi-supervised conditional constraint domain feature alignment. The introduction of multiple source domains can expand the diversity of training data, reduce the risk of the model relying on the characteristics of a single source domain, and improve the robustness of the model on different data. Semi-supervised conditional constraint domain feature alignment reduces the feature difference between the source domain and the target domain through the adversarial training idea of ​​adversarial discriminant domain adaptation, uses a small amount of labeled data in the target domain to pre-train the feature extractor as the initialization of the target domain feature extractor, and combines the center clustering trained with labeled target domain data as a conditional constraint in the adversarial training stage, jointly optimizes the target domain feature extractor, generates features to try to deceive the domain discriminator, and achieves inter-domain alignment and intra-domain category alignment, thereby realizing classification for unlabeled data. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-source semi-supervised conditional constraint domain adaptive industrial process fault classification method to solve the problem of inconsistent feature distribution and insufficient fault type labels caused by multiple working conditions. The present invention uses a convolutional neural network combined with a triplet loss function to extract optimized features on multi-source domain feature extraction to reduce the difference between source domain classes. And through adversarial training combined with conditional constraint loss, the source domain and target domain features are aligned to complete feature migration. Reduce the annotation cost of faults in industrial processes and improve the stability and reliability of the model. Real-time monitoring and fault diagnosis of equipment and processes are achieved. Maintain efficient fault classification capabilities to ensure the safe and stable operation of industrial processes.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A multi-source semi-supervised conditional constraint domain adaptive industrial process fault classification method, the method steps are as follows:

[0008] S1: Master the industrial process theory and basic operation procedures, and understand the sources of multi-condition data;

[0009] S2: Collect and organize different working condition data and corresponding labels collected by industrial field instruments or obtained by industrial system simulation platform;

[0010] S3: Set different working condition data as multiple source domain data and a target domain data, the target domain dataset is divided into labeled data Unlabeled data

[0011] S4: Preprocess the data set by maximum and minimum normalization;

[0012] S5: Mix multiple source domain data into one dataset by using the convolutional neural network Gs (X s θ s ) Combined with triplet loss to extract optimized features and train the classifier C(F s θ C ), save the classifier model parameters;

[0013] S6: Pre-train the feature extractor G using labeled data from the target domain dataset t (X L θ t ), and obtain the mixed target domain data features through the pre-trained feature extractor;

[0014] S7: Train the center clustering with labeled target domain features and calculate the center loss Mixed data features and source domain features participate in adversarial training and calculate adversarial loss Use the center loss as the constraint condition of the target domain category, and combine it with the adversarial loss to jointly optimize the target domain feature extractor. After reaching the minimum loss, save the optimized target domain feature extractor model and the training model.

[0015] S8: The preprocessed target domain data is used as a test set to extract features through the feature extractor model saved in the training part, and the extracted features are used to construct pseudo labels and classify through the classifier model saved in the training part;

[0016] S9: Draw a confusion map by combining the data labels and the constructed pseudo labels to test the model;

[0017] The multi-source semi-supervised conditional constraint domain adaptive industrial process fault classification method includes being familiar with and mastering the operation flow of the industrial process and the processing and application of data.

[0018] The method for industrial process fault classification based on multi-source semi-supervised conditional constraint domain adaptation utilizes more comprehensive feature information in each domain of the multi-source combined with semi-supervised conditional constraint domain adaptation to align source domain and target domain features; the overall steps of the method are divided into a training part and a testing part:

[0019] Training part:

[0020] Mix multiple source domain data into one dataset and perform data preprocessing (normalization);

[0021] Convolutional neural network combined with triplet loss to extract optimized features;

[0022] Train the classifier and save the classifier model;

[0023] Use the labeled data (normalized) in the target domain dataset to pre-train the feature extractor, whose model structure is the same as the source domain feature extractor;

[0024] The mixed target domain data features (including labeled data and unlabeled data (normalized)) are obtained through the pre-trained feature extractor;

[0025] Train center clustering with labeled features and calculate center loss; Mix data features and source domain features in adversarial training and calculate adversarial loss; Use center loss as a constraint on the target domain category and combine adversarial loss to jointly optimize the target domain feature extractor;

[0026] After reaching the minimum loss, save the optimized target domain feature extractor model for testing;

[0027] Test part:

[0028] (1) Test data preprocessing (normalization);

[0029] (2) extracting features by training the partially saved target domain feature extractor model;

[0030] (3) The extracted features are used to construct pseudo labels and classify through the classifier model saved in the training part;

[0031] (4) Draw a confusion map by combining the data labels and the constructed pseudo-labels, and test the model.

[0032] The advantages and positive effects of the present invention are:

[0033] 1. In modern industrial processes, data distribution under different equipment, working conditions, and operating conditions may vary significantly. By integrating multi-source domain data, the generalization ability of the model can be improved, so that it can still maintain a high classification accuracy when facing failures under different conditions. In addition, the use of multi-source domains can also make full use of the information in each domain to improve the robustness and reliability of the model.

[0034] 2. Fault data in industrial processes are often difficult to fully label, and the labeling cost is high, while unlabeled data is relatively abundant. When the target domain data is limited, semi-supervised conditional constrained domain adaptation uses the method of transfer learning to train the model using labeled source domain data and a small amount of labeled target domain data, effectively improving the accuracy of fault classification. At the same time, semi-supervised conditional constrained domain adaptation performs well in dealing with data distribution differences. It narrows the distribution gap between the source domain and the target domain through the idea of ​​confrontation, making the model more stable and accurate in the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the technical route of the present invention;

[0036] Figure 2 is the confusion matrix of the four methods under the TE process;

[0037] Figure 3 is the confusion matrix of the four methods under the TFF process;

[0038] Figure 4 It is the overall structural flow chart of the model of the present invention. DETAILED DESCRIPTION

[0039] A process fault classification method based on multi-source domain semi-supervised conditional constraint domain adaptation, the method comprising the following steps:

[0040] S1: Master the industrial process theory and basic operation procedures, and understand the sources of multi-condition data;

[0041] S2: Collect and organize different working condition data and corresponding labels collected by industrial field instruments or obtained by industrial system simulation platform;

[0042] S3: Set different working condition data as multiple source domain data and a target domain data, the target domain dataset is divided into labeled data Unlabeled data

[0043] S4: Preprocess the data set by maximum and minimum normalization;

[0044] S5: Mix multiple source domain data into one dataset by using the convolutional neural network G s (X s θ s ) Combined with triplet loss to extract optimized features and train the classifier C(F s θ C ), save the classifier model parameters;

[0045] S6: Pre-train the feature extractor G using labeled data from the target domain dataset t (X L θ t ), whose model structure is the same as the source domain feature extractor, and obtains mixed target domain data features (labeled data and unlabeled data) through the pre-trained feature extractor;

[0046] S7: Train the center clustering with labeled target domain features and calculate the center loss Mixed data features and source domain features participate in adversarial training and calculate adversarial loss Use the center loss as the constraint condition of the target domain category, and combine it with the adversarial loss to jointly optimize the target domain feature extractor. After reaching the minimum loss, save the optimized target domain feature extractor model for testing;

[0047] S8: The preprocessed target domain data is used as a test set to extract features through the feature extractor model saved in the training part, and the extracted features are used to construct pseudo labels and classify through the classifier model saved in the training part;

[0048] S9: Draw a confusion map by combining the data labels and the constructed pseudo labels and test the model;

[0049] The present invention uses step S5 specifically as follows:

[0050] S51: Convolutional neural network extracts features, the formula is as follows (1):

[0051]

[0052] X S is the source domain dataset, θ s is the source domain extractor G s Parameters, is the i-th data in the source domain dataset. The convolutional neural network uses ReLU as the activation function;

[0053] S52: Triplet loss optimizes multi-source domain features. The trained source domain features are optimized through triplet loss, which includes three parts: anchor, positive example, and negative example. The anchor and the positive example have the same class label, while the negative example has samples with different labels. The formula is as follows (2):

[0054]

[0055] is the characteristic of the anchor point, is a characteristic of a positive example, is the feature of negative examples, and margin is a hyperparameter used to control the distance difference between similar data. The features optimized by triplet loss ensure that the features of the same category of data between the source domain and the source domain are more closely distributed.

[0056] S53: Extract optimized features and train a classifier using the cross entropy loss method. The specific formula is as follows (3):

[0057]

[0058] θ C are the parameters of the classifier. Save the trained classifier model.

[0059] The present invention uses step 6 as follows:

[0060] S61: Pre-train the target domain feature extractor with labeled data, and use the pre-trained feature extractor as the initialization feature extractor for the mixed target domain data. The formula is as follows (4):

[0061]

[0062] The present invention uses step 7 as follows:

[0063] S71: The labeled target domain features are clustered using the center loss method, as shown in formula (5):

[0064]

[0065] The center loss of training is used as a conditional constraint and introduced into the adversarial training to guide the alignment of category features in the source domain and the target domain.

[0066] S72: Mixed data features and source domain features participate in adversarial training, and the adversarial loss is calculated as shown in formula (6)

[0067]

[0068] S73: Semi-supervised conditional constraint adversarial loss optimization can be expressed as formula (7):

[0069]

[0070] Updating the feature extractor by minimizing the conditional constraint adversarial loss is expressed as formula (8):

[0071]

[0072] λ is a hyperparameter used to control the influence weight of the center loss, where θ t are the parameters of the target domain feature extractor, η is the learning rate, It is the gradient of the total loss relative to the parameters of the target domain feature extractor. Because through the semi-supervised conditional constraint adversarial loss optimization, the target domain feature extractor can gradually generate features that can deceive the domain discriminator to ensure the alignment between the source domain and the target domain while also ensuring the alignment of categories within the domain, thus improving the generalization performance of the model.

[0073] Hereinafter, the present invention will be described in detail through exemplary embodiments.

[0074] See also Figure 4 , the present invention provides a multi-source domain semi-supervised conditional constraint domain adaptation process fault classification method, the steps of which are:

[0075] Training part:

[0076] (1) Mix multiple source domain data into one dataset and perform data preprocessing (normalization);

[0077] (2) Convolutional neural network combined with triplet loss to extract optimized features;

[0078] (3) Train the classifier and save the classifier model;

[0079] (4) Use the labeled data (normalized) in the target domain dataset to pre-train the feature extractor, whose model structure is the same as the source domain feature extractor;

[0080] (5) Obtaining mixed target domain data features (including labeled data and unlabeled data (normalized)) through a pre-trained feature extractor;

[0081] (6) Train center clustering with labeled features and calculate center loss. Mix data features and source domain features in adversarial training and calculate adversarial loss. Use center loss as a constraint on the target domain category and combine adversarial loss to jointly optimize the target domain feature extractor;

[0082] (7) After reaching the minimum loss, save the optimized target domain feature extractor model for testing; Testing part:

[0083] (1) Test data preprocessing (normalization);

[0084] (2) extracting features by training the partially saved target domain feature extractor model;

[0085] (3) The extracted features are used to construct pseudo labels and classify through the classifier model saved in the training part;

[0086] (4) Draw a confusion map by combining the data labels and the constructed pseudo labels to test the model;

[0087] Example 1

[0088] The Tennessee Eastman process is a widely used benchmark system in the field of industrial process control and fault diagnosis. It evaluates the performance of control and diagnostic algorithms in complex industrial processes. The dataset of the present invention includes data of normal operation and various fault types. Each dataset includes 12 input variables, 41 measurement variables and 28 interference variables. A total of ten faults are selected for the experiment. Each category in the source domain contains 600 samples, and each category in the target domain contains 200 samples.

[0089] Table 1 Classification accuracy of the four methods in the TE process

[0090]

[0091]

[0092] From Table 1 and the instruction manual Figure 2It can be seen that the average classification accuracy of the model of the present invention is better than that of other models. In categories 1 to 3, all methods can be well classified. From categories 4 to 9, the classification of DANN and ADDA is obviously confusing. The reason is that both DANN and ADDA are unsupervised domain adaptation methods, which only align inter-domain features but not intra-domain features. If the features of the category data are not obvious, it will lead to serious misclassification. MSSDA is an ablation experiment, and its effect is also significantly lower than that of the method of the present invention.

[0093] Example 2

[0094] The multiphase flow device developed by Cranfield University is designed to provide controllable and measurable flow for pressurized systems of water, oil and air. The data set includes data of normal operation and various fault types, and each data set includes 24 variables. The present invention uses mode 2 and mode 3 as source domains 1 and 2, and mode 1 as the target domain. Categories 1 to 5 represent normal samples and fault 1 to fault 4 samples respectively. Each category in the source domain contains 600 samples, and each category in the target domain contains 200 samples.

[0095] Table 2 Classification accuracy of four methods under TFF conditions

[0096]

[0097] From Table 2 and the instruction manual Figure 3 DANN and ADDA have confusion in the classification of categories 1 and 2, SSDA has confusion in the classification of categories 4 and 5, and although the present invention has misclassification between categories 1 and 5, overall, the average accuracy of the present invention is better than the first three methods.

[0098] The specific embodiments of the present invention have been disclosed above. The present invention is applicable to the fields mentioned in the specification and other suitable fields. Without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the description illustrations and specific related details given in the specification.

Claims

1. A multi-source semi-supervised conditional constraint domain adaptation industrial process fault classification method, characterized in that: The method steps are as follows: S1: Master the industrial process theory and basic operation procedures, and understand the sources of multi-condition data; S2: Collect and organize different working condition data and corresponding labels collected by industrial field instruments or obtained by industrial system simulation platform; S3: Set different working condition data as multiple source domain data and a target domain data, the target domain dataset is divided into labeled data Unlabeled data S4: Preprocess the data set by maximum and minimum normalization; S5: Mix multiple source domain data into one dataset by using the convolutional neural network G s (X s ; θ s ) Combined with triplet loss to extract optimized features and train the classifier C(F s θ C ), save the classifier model parameters; S6: Pre-train the feature extractor G using labeled data from the target domain dataset t (X L θ t ), and obtain the mixed target domain data features through the pre-trained feature extractor; S7: Train the center clustering with labeled target domain features and calculate the center loss Mixed data features and source domain features participate in adversarial training and calculate adversarial loss Use the center loss as the constraint condition of the target domain category, and combine it with the adversarial loss to jointly optimize the target domain feature extractor. After reaching the minimum loss, save the optimized target domain feature extractor model and the training model. S8: The preprocessed target domain data is used as a test set to extract features through the feature extractor model saved in the training part, and the extracted features are used to construct pseudo labels and classify through the classifier model saved in the training part; S9: Combine the data labels and constructed pseudo labels to draw a confusion map and test the model.

2. According to claim 1, a multi-source semi-supervised conditional constraint domain adaptive industrial process fault classification method is characterized in that: The method includes familiarizing with the operating procedures of industrial processes and the processing and application of data.

3. The method for industrial process fault classification based on multi-source semi-supervised conditional constraint domain adaptation according to claim 1 is characterized in that: This method uses more comprehensive feature information in each domain of multiple sources combined with semi-supervised conditional constrained domain adaptation to align source domain and target domain features; The overall steps of its method are divided into training part and testing part: Training part: (1) Mix multiple source domain data into one dataset and perform data preprocessing (normalization); (2) Convolutional neural network combined with triplet loss to extract optimized features; (3) Train the classifier and save the classifier model; (4) Use the labeled data (normalized) in the target domain dataset to pre-train the feature extractor, whose model structure is the same as the source domain feature extractor; (5) Obtaining mixed target domain data features (including labeled data and unlabeled data (normalized)) through a pre-trained feature extractor; (6) Train center clustering with labeled features and calculate center loss; Mix data features and source domain features in adversarial training and calculate adversarial loss; Use center loss as a constraint condition for the target domain category and combine adversarial loss to jointly optimize the target domain feature extractor; (7) After reaching the minimum loss, save the optimized target domain feature extractor model for testing; Test part: (1) Test data preprocessing (normalization); (2) extracting features by training the partially saved target domain feature extractor model; (3) The extracted features are used to construct pseudo labels and classify through the classifier model saved in the training part; (4) Draw a confusion map by combining the data labels and the constructed pseudo-labels, and test the model.

Images