A multi-source semi-supervised condition-constrained domain adaptation method for industrial process fault classification

Abstract

The application relates to an industrial process soft measurement method, and specifically relates to a multi-source semi-supervised condition constraint domain adaptive industrial process fault classification method, which comprises the following steps: familiarizing with an actual industrial process; collecting different working condition data and constructing source domain target domain data sets; pre-processing the data sets through maximum minimum normalization; saving a classifier model; in a test stage, target domain data is used for testing, and classification is performed through the saved feature extractor model and the classifier model, so that the generalization ability and classification precision of the model are improved, the labeling cost of faults in an industrial process is reduced, and the stability and reliability of the model are improved. The application is integrated into an industrial control system to realize real-time monitoring and fault diagnosis of equipment and processes. Through continuous learning and adaptation to new data, the model can maintain efficient fault classification ability, and the safe and stable operation of an industrial process is ensured.

Description

Technical Field

[0001] This invention belongs to the field of industrial process fault classification, specifically involving a multi-source domain semi-supervised domain adaptation (MSSCCDA) method. By using more comprehensive information from multiple sources and adversarial domain adaptation methods, the target domain features are aligned to achieve feature transfer, thereby facilitating the classification of unlabeled target domain data. Background Technology

[0002] In modern industrial processes, industrial process fault classification is an important research area in modern industrial systems. It aims to identify and classify various potential faults by analyzing and processing data from industrial processes. Industrial processes typically involve complex mechanical, electrical, and chemical systems that may fail during operation due to various reasons, such as equipment aging, operational errors, and environmental changes. Timely and accurate identification and classification of these faults are crucial for ensuring the safety, reliability, and economy of industrial systems.

[0003] Fault classification is a crucial step in fault diagnosis. By categorizing faults into different types, targeted guidance can be provided for further handling and repair. Traditional fault classification methods rely on expert knowledge and experience, typically employing rule-based or model-based approaches, which can achieve good results in specific application scenarios. However, with the increasing complexity of industrial processes, these methods become inadequate when handling large-scale, high-dimensional, and multivariate data, failing to meet practical needs. In recent years, the rise of data-driven methods, especially deep learning-based methods, has garnered widespread attention and application in industrial process fault classification due to their powerful feature extraction capabilities and adaptability. However, general data-driven fault classification methods assume that training data extracted from the source domain and test data extracted from the target domain should have the same feature space distribution. This assumption is often challenged in practical industrial applications. Considering numerous factors, such as changes in setpoints, different operating conditions, mechanical wear, and personnel operating methods, datasets collected under different working conditions often exhibit differences in feature distribution, a phenomenon known as the multi-condition problem. Furthermore, in practical industrial applications, insufficient labels can be obtained for each fault type, further complicating the establishment of fault identification models.

[0004] To address the above problems and for application in complex industrial processes, this invention proposes a method for fault classification using multi-source domain feature extraction and semi-supervised conditional constraint domain feature alignment. The introduction of multiple source domains expands the diversity of training data, reduces the risk of the model relying on the characteristics of a single source domain, and improves the model's robustness across different data. Semi-supervised conditional constraint domain feature alignment reduces the feature differences between the source and target domains through adversarial training, using a small amount of labeled data from the target domain to pre-train the feature extractor as initialization. During the adversarial training phase, it combines the central clustering trained on labeled target domain data as a conditional constraint to jointly optimize the target domain feature extractor, generating features that attempt to deceive the domain discriminator, achieving inter-domain alignment and intra-domain class alignment, thereby enabling the classification of unlabeled data. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-source, semi-supervised, conditionally constrained domain-adaptive industrial process fault classification method, addressing the problems of inconsistent feature distributions and insufficient fault type labels caused by multiple operating conditions. This invention utilizes convolutional neural networks combined with a triplet loss function to extract optimized features in multi-source domain feature extraction, reducing inter-class differences between source domains. Furthermore, it aligns source and target domain features through adversarial training combined with conditional constraint loss, achieving feature transfer. This reduces the labeling cost of faults in industrial processes, improves model stability and reliability, enables real-time monitoring and fault diagnosis of equipment and processes, maintains efficient fault classification capabilities, and ensures the safe and stable operation of industrial processes.

[0006] The objective of this invention is achieved through the following technical solution:

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

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

[0009] S2: Collect and organize different operating condition data and corresponding tags from industrial field instruments or industrial system simulation platforms;

[0010] S3: Set different operating condition data as multiple source domain data A target domain dataset, which is divided into labeled data. Unlabeled data

[0011] S4: Preprocess the dataset using max-min normalization;

[0012] S5: Combine data from multiple source domains into a single dataset using a convolutional neural network G.s (X s ;θ s Combine triplet loss to extract optimized features and train a 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 mixed target domain data features through a pre-trained feature extractor;

[0014] S7: Train center clustering using labeled target domain features, and calculate center loss. Hybrid data features and source domain features are used in adversarial training to calculate adversarial loss. The center loss is used as a constraint on the target domain category. Combined with the adversarial loss, the target domain feature extractor is optimized. After the minimum loss is reached, the optimized target domain feature extractor model is saved, and the training model is saved.

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

[0016] S9: Combine the labels in the data with the constructed pseudo-labels to draw a confusion graph and test the model;

[0017] The aforementioned multi-source semi-supervised conditional constraint domain adaptation industrial process fault classification method includes familiarity with and application of industrial process operation procedures and data processing.

[0018] The aforementioned multi-source semi-supervised conditional constraint domain adaptation method for industrial process fault classification utilizes more comprehensive feature information from various multi-source domains, combined with semi-supervised conditional constraint domain adaptation to align source and target domain features. The overall method consists of a training phase and a testing phase.

[0019] Training section:

[0020] Combine data from multiple source domains into a single dataset and perform data preprocessing (normalization);

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

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

[0023] The feature extractor is pre-trained using labeled (normalized) data from the target domain dataset, and its model structure is the same as that of the source domain feature extractor.

[0024] The features of the mixed target domain data (including labeled data and unlabeled data (after normalization)) are obtained by a pre-trained feature extractor.

[0025] Clustering is performed using labeled features, and the center loss is calculated. Mixed data features and source domain features are used for adversarial training, and the adversarial loss is calculated. The center loss is used as a constraint on the target domain category, and the adversarial loss is combined with the adversarial loss to optimize the target domain feature extractor.

[0026] After minimizing the loss, save the optimized target domain feature extractor model for testing.

[0027] Test section:

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

[0029] (2) Extract features using the target domain feature extractor model saved in the training section;

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

[0031] (4) Combine the labels of the data and the constructed pseudo-labels to draw a confusion graph and test the model.

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

[0033] 1. In modern industrial processes, data distribution can vary significantly depending on the equipment, operating conditions, and other factors. Integrating multi-source domain data can improve the generalization ability of a model, enabling it to maintain high classification accuracy when facing faults under different conditions. Furthermore, using multi-source domain data can fully utilize information from each domain, enhancing the model's robustness and reliability.

[0034] 2. Fault data in industrial processes is often difficult to fully label, and labeling is costly, while unlabeled data is relatively abundant. Semi-supervised conditional domain adaptation, when target domain data labeling is limited, utilizes transfer learning to train the model using labeled source domain data and a small amount of labeled target domain data, effectively improving fault classification accuracy. Simultaneously, semi-supervised conditional domain adaptation excels in handling data distribution differences, narrowing the distribution gap between the source and target domains through adversarial thinking, making the model's performance in the target domain more stable and accurate. Attached Figure Description

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

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

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

[0038] Figure 4 This is a flowchart illustrating the overall structure of the model of this invention. Detailed Implementation

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

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

[0041] S2: Collect and organize different operating condition data and corresponding tags from industrial field instruments or industrial system simulation platforms;

[0042] S3: Set different operating condition data as multiple source domain data A target domain dataset, which is divided into labeled data. Unlabeled data

[0043] S4: Preprocess the dataset using max-min normalization;

[0044] S5: Combine data from multiple source domains into a single dataset using a convolutional neural network G. s (X s ;θ s Combine triplet loss to extract optimized features and train a 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 Its model structure is the same as the source domain feature extractor, and it obtains mixed target domain data features (labeled data and unlabeled data) through a pre-trained feature extractor.

[0046] S7: Train center clustering using labeled target domain features, and calculate center loss. Hybrid data features and source domain features are used in adversarial training to calculate adversarial loss. The center loss is used as a constraint on the target domain category. Combined with the adversarial loss, the target domain feature extractor is optimized. After the minimum loss is reached, the optimized target domain feature extractor model is saved for testing.

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

[0048] S9: Combine the labels in the data with the constructed pseudo-labels to draw a confusion graph and test the model;

[0049] The specific application of step S5 in this invention is as follows:

[0050] S51: Convolutional neural network extracts features, as shown in formula (1):

[0051]

[0052] X S For the source domain dataset, θ s For source domain extractor G s The parameters, Let be the i-th data point in the source domain dataset. Convolutional neural networks use ReLU as the activation function.

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

[0054]

[0055] These are characteristics of anchor points. These are characteristics of positive examples. Features are the characteristics of negative examples, and margin is a hyperparameter used to control the distance difference between similar data. Features optimized using triplet loss ensure a closer distribution of features among data of the same category between different source domains.

[0056] S53: Extract the optimized features and train a classifier using the cross-entropy loss method, as shown in formula (3).

[0057]

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

[0059] The specific application of step 6 in this invention is as follows:

[0060] S61: A target domain feature extractor is pre-trained using labeled data, and then used as the initial feature extractor for the mixed target domain data. Its formula is as follows (4):

[0061]

[0062] The specific application of step 7 in this invention is as follows:

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

[0064]

[0065] By using the center loss of the training as a conditional constraint, it is introduced into the adversarial training to guide the alignment of class features in the source and target domains.

[0066] S72: Mixed data features and source domain features are used 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 Equation (8):

[0071]

[0072] λ is a hyperparameter used to control the influence weight of the central loss, where θ t These are the parameters of the target domain feature extractor, where η is the learning rate. This is the gradient of the total loss relative to the parameters of the target domain feature extractor. Because the target domain feature extractor optimizes against the loss through semi-supervised conditional constraints, it can gradually generate features that can deceive the domain discriminator, ensuring both inter-domain alignment between the source and target domains and intra-domain class alignment, thereby improving the model's generalization performance.

[0073] The present invention will now be described in detail through exemplary embodiments.

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

[0075] Training section:

[0076] (1) Combine data from multiple source domains into a single dataset and perform data preprocessing (normalization);

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

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

[0079] (4) Use labeled data (after normalization) in the target domain dataset to pre-train the feature extractor, and its model structure is the same as that of the source domain feature extractor.

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

[0081] (6) Clustering with labeled features is trained, and the center loss is calculated. Mixed data features and source domain features are used for adversarial training, and the adversarial loss is calculated. The center loss is used as a constraint on the target domain category, and the adversarial loss is combined with the adversarial loss to optimize the target domain feature extractor.

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

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

[0084] (2) Extract features using the target domain feature extractor model saved in the training section;

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

[0086] (4) Combine the labels of the data and the constructed pseudo-labels to draw a confusion graph and 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 this invention includes data on normal operation and various fault types. Each dataset includes 12 input variables, 41 measurement variables, and 28 disturbance variables. Ten types of faults were 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 four methods in TE processes.

[0090]

[0091]

[0092] From Table 1 and the instruction manual appendix Figure 2As can be seen, the average classification accuracy of the model of this invention is better than other models. For categories 1 to 3, all methods can classify well. For categories 4 to 9, DANN and ADDA exhibit significant misclassification. This is because DANN and ADDA are both unsupervised domain adaptation methods, achieving only inter-domain feature alignment but not intra-domain feature alignment. If the features of the category data are not clearly distinguishable, it will lead to serious misclassification. MSSDA, based on ablation experiments, also performs significantly worse than the method of this invention.

[0093] Example 2

[0094] The multiphase flow apparatus developed by Cranfield University aims to provide controllable and measurable flow rates for pressurized systems of water, oil, and air. The datasets include data from normal operation and various fault types, with each dataset containing 24 variables. This invention uses modes 2 and 3 as source domains 1 and 2, and mode 1 as the target domain. Categories 1 to 5 represent normal samples and fault samples 1 to 4, 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 in the TFF case.

[0096]

[0097] From Table 2 and the instruction manual appendix Figure 3 DANN and ADDA exhibit misclassification in categories 1 and 2, while SSDA shows misclassification in categories 4 and 5. Although the present invention also exhibits misclassification in categories 1 and 5, its overall average accuracy is superior to the previous three methods.

[0098] The specific embodiments of the present invention have been disclosed above. The present invention is applicable to the fields described 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 descriptive illustrations and specific related details given in the specification.

Claims

1. A multi-source semi-supervised conditional constraint domain adaptation method for industrial process fault classification, characterized in that, The method steps are as follows: S1: Master the theory and basic operating procedures of industrial processes, and understand the sources of multi-condition data; S2: Collect and organize different operating condition data and corresponding tags from industrial field instruments or industrial system simulation platforms; S3: Set different operating condition data as multiple source domain data A target domain dataset, which is divided into labeled data. Unlabeled data ; S4: Preprocess the dataset using max-min normalization; S5: Combine data from multiple source domains into a single dataset using a convolutional neural network. Combine triplet loss to extract optimized features and train a classifier. Save the classifier model parameters; S6: Pre-train a feature extractor using labeled data from the target domain dataset. And obtain hybrid target domain data features through a pre-trained feature extractor; S7: Train center clustering using labeled target domain features, and calculate center loss. Mixed target domain data features and source domain features are used in adversarial training to calculate adversarial loss. The center loss is used as a constraint on the target domain category. Combined with the adversarial loss, the target domain feature extractor is optimized. After the minimum loss is reached, the optimized target domain feature extractor model is saved, and the training model is saved. S8: Use the preprocessed target domain data as a test set to extract features through the feature extractor model saved in the training part. The extracted features are then used to construct pseudo-labels and classify the data through the classifier model saved in the training part. S9: Combine the labels in the data with the constructed pseudo-labels to draw a confusion graph and test the model.

Images