Multi-channel domain adaptation network-based chemical process cross-apparatus fault diagnosis method and apparatus

Abstract

The present disclosure relates to a multi-channel domain adaptation network-based chemical process cross-apparatus fault diagnosis method and apparatus. The method comprises: acquiring a fault diagnosis task for a device to be diagnosed under a target process flow; on the basis of the fault diagnosis task and said device, determining target domain data required for executing the fault diagnosis task, the target domain data comprising a plurality of types of operation data of said device operating in a first target domain, and the first target domain being determined on the basis of the target process flow; and inputting the target domain data into a trained multi-channel domain adaptation network for calculation, to obtain a diagnosis result for the fault diagnosis task. The chemical process cross-apparatus fault diagnosis method and apparatus according to embodiments of the present disclosure can achieve cross-domain migration of fault-category information, achieve fault diagnosis with high accuracy and low false alarm rate on different apparatuses in similar process flows, and have an effective cross-apparatus generalization capability.

Description

A Method and Apparatus for Cross-Unit Fault Diagnosis in Chemical Processes Based on Multi-channel Domain Adaptive Networks

[0001] This disclosure claims priority to Chinese Patent Application No. 202410937242.7, filed on July 12, 2024, entitled "Method and Apparatus for Fault Diagnosis of Chemical Processes Across Units Based on Multi-channel Domain Adaptive Network", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This disclosure relates to the field of chemical process control technology, and in particular to a method and apparatus for cross-unit fault diagnosis of chemical processes based on a multi-channel domain adaptive network. Background Technology

[0003] With the rapid development of artificial intelligence and industrial big data technologies, many data-driven process control technologies have been proposed to address fault diagnosis in chemical processes. Fault diagnosis methods in these technologies are mainly based on statistical probability modeling and deep learning, with deep learning-based methods still dominating.

[0004] However, deep learning-based fault diagnosis methods are mainly used for fault diagnosis of single equipment, such as bearings and belts. Because single equipment has fewer monitored variables and more obvious fault characteristics, the diagnostic effect is relatively good. In contrast, chemical processes involve multidimensional coupling of variables and dynamic fluctuations in operating conditions, making the application scenarios for single equipment far more complex. Currently, there are no effective technical solutions for cross-unit fault diagnosis in the field of chemical process control technology. General diagnostic models are scarce, and their classification accuracy urgently needs improvement.

[0005] Public content

[0006] In view of this, this disclosure proposes a cross-unit fault diagnosis method and device for chemical processes based on a multi-channel domain adaptive network, which can realize cross-domain transfer of fault category information, achieve high accuracy and low false alarm rate fault diagnosis on different units in similar process flows, and has effective cross-unit general capability.

[0007] According to one aspect of this disclosure, a method for cross-unit fault diagnosis in chemical processes based on a multi-channel domain adaptation network is provided. The method includes: acquiring a fault diagnosis task for a device to be diagnosed under a target process flow; determining target domain data required to perform the fault diagnosis task based on the fault diagnosis task and the device to be diagnosed, wherein the target domain data includes various operating data of the device to be diagnosed operating under a first target domain, the first target domain being determined according to the target process flow; inputting the target domain data into a trained multi-channel domain adaptation network for calculation to obtain a diagnosis result for the fault diagnosis task; wherein the method further includes a training process for the multi-channel domain adaptation network, the training process including the following steps: acquiring a target domain sample dataset and a source domain sample dataset, wherein the target domain sample dataset includes multiple different sample devices operating under multiple different process flows. The target domain sample data under the second target domain includes source domain sample data of each sample device operating in multiple different source domains. Each target domain sample data and each source domain sample data includes multiple normal operation data and multiple abnormal operation data. Each sample device, each second target domain, and each source domain are determined according to different process flows in the chemical process. Each normal operation data and each abnormal operation data in the source domain sample dataset are labeled to obtain a source domain labeled dataset. A training sample set is constructed based on the source domain labeled dataset and the target domain sample dataset. The initialized multi-channel domain adaptation network is trained based on a preset loss function and the training sample set to obtain a trained multi-channel domain adaptation network. The loss function includes a feature distribution difference function constructed based on the maximum mean difference induced by a multinomial kernel.

[0008] In one possible implementation, the multi-channel domain adaptation network includes a feature extraction module and a domain adaptation module. The initialization of the multi-channel domain adaptation network is trained based on a preset loss function and the training sample set to obtain a trained multi-channel domain adaptation network. This includes: processing the training sample set using the feature extraction module to obtain a processing result; determining the source domain loss function value and target domain loss function value of the feature extraction module based on the processing result, wherein the processing result includes source domain features, target domain features, source domain classification result, and target domain classification result; processing the source domain features and target domain features using the domain adaptation module to obtain fault category prediction results for each abnormal operation data in the target domain sample data; determining the relational loss function value of the domain adaptation module based on each fault category prediction result; determining the feature distribution difference function value using a polynomial kernel function and the training sample set; determining the total loss function value of the current multi-channel domain adaptation network based on the source domain loss function value, the target domain loss function value, the relational loss function value, and the feature distribution difference function value; and updating the parameters of the current multi-channel domain adaptation network based on the total loss function value.

[0009] In one possible implementation, the feature extraction module includes a source domain branch and a target domain branch; wherein, processing the training sample set using the feature extraction module to obtain processing results, and determining the source domain loss function value and target domain loss function value of the feature extraction module based on the processing results, includes: using the source domain branch to extract features from the data of each source domain in the source domain sample data to obtain source domain features of each source domain, inferring the source domain classification result of the corresponding source domain based on the source domain features of each source domain, and determining the source domain loss function value of the feature extraction module based on the source domain classification result and the true domain label of each source domain; using the target domain branch to extract features from the target domain sample data to obtain target domain features of each second target domain, inferring the target domain classification result of the corresponding second target domain based on the target domain features, and determining the target domain loss function value of the feature extraction module based on the target domain classification result and the true domain label of each second target domain.

[0010] In one possible implementation, the domain adaptation module processes the source domain features and the target domain features to obtain fault category prediction results for each abnormal operation data in the target domain sample data. Based on each fault category prediction result, the relational loss function value of the domain adaptation module is determined. This includes: using the domain adaptation module to calculate the target domain features and the source domain features to obtain domain-adapted sample features; determining the fault category prediction results for the target domain sample data based on the domain-adapted sample feature degree and a predetermined fault category; and determining the relational loss function value of the domain adaptation module based on the fault category prediction results and the true fault labels of the target domain sample data.

[0011] In one possible implementation, the source domain features include source domain-specific features representing information unique to the source domain, the target domain features include target domain-specific features representing information unique to the second target domain, and the domain adaptation module includes a first fully connected layer network pair and a second fully connected layer network pair; wherein, determining the feature distribution difference function value using a polynomial kernel function and the training sample set includes: determining a first feature distribution difference function value based on the source domain sample data according to the definition rules of the polynomial kernel function, determining a second feature distribution difference function value based on the output of the first fully connected layer network pair, and determining a third feature distribution difference function value based on the output of the second fully connected layer network pair; and determining the feature distribution difference function value based on the first feature distribution difference function value, the second feature distribution difference function value, and the third feature distribution difference function value.

[0012] In one possible implementation, the target domain data is input into a trained multi-channel domain adaptation network for computation to obtain a diagnostic result for the fault diagnosis task. This includes: extracting sub-domain features from the target domain data to obtain sub-domain specific features representing information unique to the first target domain; and extracting global features from the target domain data to obtain global universal features representing common information shared by the first target domain and each of the source domains. Domain migration features are then determined based on the sub-domain specific features and the global universal features. Fault diagnosis is performed based on the domain migration features to obtain a diagnostic result for the fault diagnosis task.

[0013] In one possible implementation, subdomain feature extraction is performed on the target domain data to obtain subdomain-specific features representing information unique to the target domain, and global feature extraction is performed on the target domain data to obtain global universal features representing common information shared by the target domain and each of the source domains. This includes: performing subdomain feature extraction on the target domain data at a time scale to obtain subdomain temporal features; performing subdomain feature extraction on the subdomain temporal features at a spatial scale to obtain the subdomain-specific features; and performing global feature extraction on the target domain data at a time scale to obtain global temporal features; and performing global feature extraction on the global temporal features at a spatial scale to obtain the global-specific features.

[0014] According to another aspect of this disclosure, a fault diagnosis device for a chemical process across units based on a multi-channel domain adaptation network is provided, comprising: an acquisition module for acquiring a fault diagnosis task for a device to be diagnosed under a target process flow; a determination module for determining target domain data required to perform the fault diagnosis task based on the fault diagnosis task and the device to be diagnosed, the target domain data including various operating data of the device to be diagnosed operating under a first target domain, the first target domain being determined according to the target process flow; and a calculation module for inputting the target domain data into a trained multi-channel domain adaptation network for calculation to obtain a diagnosis result for the fault diagnosis task; wherein, the device further comprises a training module, the training module being used to: acquire a target domain sample dataset and a source domain sample dataset, the target domain sample dataset including multiple different sample devices operating under multiple different process flow conditions; The target domain sample data under the same second target domain, the source domain sample dataset includes source domain sample data of each sample device operating under multiple different source domains, wherein each target domain sample data and each source domain sample data includes multiple normal operation data and multiple abnormal operation data, each sample device, each second target domain, and each source domain are determined according to different process flows in the chemical process; each normal operation data and each abnormal operation data in the source domain sample dataset are labeled to obtain a source domain labeled dataset, and a training sample set is constructed based on the source domain labeled dataset and the target domain sample dataset; the initialized multi-channel domain adaptation network is trained based on a preset loss function and the training sample set to obtain a trained multi-channel domain adaptation network, wherein the loss function includes a feature distribution difference function constructed based on the maximum mean difference induced by a multinomial kernel.

[0015] In one possible implementation, the multi-channel domain adaptation network includes a feature extraction module and a domain adaptation module. The initialization of the multi-channel domain adaptation network is trained based on a preset loss function and the training sample set to obtain a trained multi-channel domain adaptation network. This includes: processing the training sample set using the feature extraction module to obtain a processing result; determining the source domain loss function value and target domain loss function value of the feature extraction module based on the processing result, wherein the processing result includes source domain features, target domain features, source domain classification result, and target domain classification result; processing the source domain features and target domain features using the domain adaptation module to obtain fault category prediction results for each abnormal operation data in the target domain sample data; determining the relational loss function value of the domain adaptation module based on each fault category prediction result; determining the feature distribution difference function value using a polynomial kernel function and the training sample set; determining the total loss function value of the current multi-channel domain adaptation network based on the source domain loss function value, the target domain loss function value, the relational loss function value, and the feature distribution difference function value; and updating the parameters of the current multi-channel domain adaptation network based on the total loss function value.

[0016] In one possible implementation, the feature extraction module includes a source domain branch and a target domain branch; wherein, processing the training sample set using the feature extraction module to obtain processing results, and determining the source domain loss function value and target domain loss function value of the feature extraction module based on the processing results, includes: using the source domain branch to extract features from the data of each source domain in the source domain sample data to obtain source domain features of each source domain, inferring the source domain classification result of the corresponding source domain based on the source domain features of each source domain, and determining the source domain loss function value of the feature extraction module based on the source domain classification result and the true domain label of each source domain; using the target domain branch to extract features from the target domain sample data to obtain target domain features of each second target domain, inferring the target domain classification result of the corresponding second target domain based on the target domain features, and determining the target domain loss function value of the feature extraction module based on the target domain classification result and the true domain label of each second target domain.

[0017] In one possible implementation, the domain adaptation module processes the source domain features and the target domain features to obtain fault category prediction results for each abnormal operation data in the target domain sample data. Based on each fault category prediction result, the relational loss function value of the domain adaptation module is determined. This includes: using the domain adaptation module to calculate the target domain features and the source domain features to obtain domain-adapted sample features; determining the fault category prediction results for the target domain sample data based on the domain-adapted sample feature degree and a predetermined fault category; and determining the relational loss function value of the domain adaptation module based on the fault category prediction results and the true fault labels of the target domain sample data.

[0018] In one possible implementation, the source domain features include source domain-specific features representing information unique to the source domain, the target domain features include target domain-specific features representing information unique to the second target domain, and the domain adaptation module includes a first fully connected layer network pair and a second fully connected layer network pair; wherein, determining the feature distribution difference function value using a polynomial kernel function and the training sample set includes: determining a first feature distribution difference function value based on the source domain sample data according to the definition rules of the polynomial kernel function, determining a second feature distribution difference function value based on the output of the first fully connected layer network pair, and determining a third feature distribution difference function value based on the output of the second fully connected layer network pair; and determining the feature distribution difference function value based on the first feature distribution difference function value, the second feature distribution difference function value, and the third feature distribution difference function value.

[0019] In one possible implementation, the target domain data is input into a trained multi-channel domain adaptation network for computation to obtain a diagnostic result for the fault diagnosis task. This includes: extracting sub-domain features from the target domain data to obtain sub-domain specific features representing information unique to the first target domain; and extracting global features from the target domain data to obtain global universal features representing common information shared by the first target domain and each of the source domains. Domain migration features are then determined based on the sub-domain specific features and the global universal features. Fault diagnosis is performed based on the domain migration features to obtain a diagnostic result for the fault diagnosis task.

[0020] In one possible implementation, subdomain feature extraction is performed on the target domain data to obtain subdomain-specific features representing information unique to the target domain, and global feature extraction is performed on the target domain data to obtain global universal features representing common information shared by the target domain and each of the source domains. This includes: performing subdomain feature extraction on the target domain data at a time scale to obtain subdomain temporal features; performing subdomain feature extraction on the subdomain temporal features at a spatial scale to obtain the subdomain-specific features; and performing global feature extraction on the target domain data at a time scale to obtain global temporal features; and performing global feature extraction on the global temporal features at a spatial scale to obtain the global-specific features.

[0021] According to another aspect of this disclosure, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the above-described method when executing instructions stored in the memory.

[0022] According to another aspect of this disclosure, a non-volatile computer-readable storage medium is provided that stores computer program instructions thereon, wherein the computer program instructions, when executed by a processor, implement the above-described method.

[0023] According to another aspect of this disclosure, a computer program product is provided, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is run in a processor of an electronic device, the processor in the electronic device performs the above-described method.

[0024] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0025] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0026] Figures 1 and 2 show schematic diagrams of a cross-unit fault diagnosis method for chemical processes based on a multi-channel domain adaptive network according to embodiments of the present disclosure.

[0027] Figure 3 shows a schematic diagram of a multi-channel domain adaptation network provided according to an embodiment of the present disclosure.

[0028] Figure 4 shows a schematic diagram of a feature extraction module provided according to an embodiment of the present disclosure.

[0029] Figure 5 shows a schematic diagram of a multi-head attention mechanism provided according to an embodiment of the present disclosure.

[0030] Figure 6 shows a schematic diagram of a converter model provided according to an embodiment of the present disclosure.

[0031] Figure 7 shows a schematic diagram illustrating the variation trend of the maximum mean difference induced by the Gaussian kernel under different bandwidth parameters.

[0032] Figure 8 shows a schematic diagram of the process flow of the Tennessee-Eastman process.

[0033] Figure 9 shows a comparative schematic diagram of the operation of the Tennessee-Eastman process provided according to an embodiment of the present disclosure under normal and fault conditions.

[0034] Figure 10 shows a schematic diagram of a cross-condition fault diagnosis confusion matrix provided according to an embodiment of the present disclosure.

[0035] Figure 11 shows a radar diagram illustrating the comparison of fault diagnosis time in the Tennessee-Eastman process using different deep transfer learning methods provided according to embodiments of the present disclosure.

[0036] Figures 12 and 13 show block diagrams of a chemical process cross-unit fault diagnosis device based on a multi-channel domain adaptive network according to embodiments of the present disclosure. Detailed Implementation

[0037] Various exemplary embodiments, features, and aspects of this disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise. The term "exemplary" used herein means "serving as an example, embodiment, or illustration." Any embodiment illustrated herein as "exemplary" is not necessarily to be construed as superior to or better than other embodiments. Furthermore, numerous specific details are set forth in the following detailed description to better illustrate this disclosure. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure. To facilitate understanding of the technical solutions provided by the embodiments of this disclosure by those skilled in the art, the technical environment for implementing the technical solutions will be described first.

[0038] The main reason why current chemical process fault diagnosis methods are not well applied to chemical production processes under multiple operating conditions is that these methods are primarily based on deep learning. However, the distribution of monitored process data changes under different process parameters, quality requirements, and equipment operating conditions. Traditional deep learning fault diagnosis methods assume that the distributions of training and test data are consistent. Furthermore, chemical process data has high dimensionality and nonlinear characteristics. Directly applying these models to multi-condition production processes will lead to method failure. Traditional methods involve fine-tuning supervised algorithm models, but this approach is prone to overfitting when there are significant differences in the distributions of the source and target domains. This results in overlearning specific distribution characteristics of the source domain data while neglecting more globally applicable features that align with the target domain features.

[0039] To address the aforementioned technical problems, this disclosure provides a cross-unit fault diagnosis method for chemical processes based on a Multi-Channel Domain Adaptive Network (MSDAN). The method includes: acquiring a fault diagnosis task for the equipment to be diagnosed under a target process flow; determining target domain data required to perform the fault diagnosis task based on the fault diagnosis task and the equipment to be diagnosed, wherein the target domain data includes various operating data of the equipment to be diagnosed operating under a first target domain, which is determined according to the target process flow; inputting the target domain data into a trained multi-channel domain adaptive network for calculation to obtain a diagnosis result for the fault diagnosis task; wherein the method further includes a training process for the multi-channel domain adaptive network, the training process including the following steps: acquiring a target domain sample dataset and a source domain sample dataset, wherein the target domain sample dataset includes multiple different sample devices operating under a target process flow. The system comprises target domain sample data under multiple different second target domains, and source domain sample datasets including source domain sample data of various sample devices operating under multiple different source domains. Each target domain sample dataset and each source domain sample dataset includes multiple normal operation data and multiple abnormal operation data. Each sample device, each second target domain, and each source domain is determined based on different process flows in the chemical process. Each normal operation data and each abnormal operation data in the source domain sample dataset is labeled to obtain a source domain labeled dataset. A training sample set is constructed based on the source domain labeled dataset and the target domain sample dataset. The initialized multi-channel domain adaptation network is trained based on a preset loss function and the training sample set to obtain the trained multi-channel domain adaptation network. The loss function includes a feature distribution difference function constructed based on the maximum mean difference induced by a multinomial kernel. This enables cross-domain transfer of fault category information, achieving high-accuracy, low-false-report fault diagnosis on different devices in similar process flows, and possesses effective cross-device generalization capabilities.

[0040] Figures 1 and 2 are schematic diagrams of a cross-unit fault diagnosis method for chemical processes based on a multi-channel domain adaptation network according to an embodiment of the present disclosure. Figure 3 is a schematic diagram of a multi-channel domain adaptation network according to an embodiment of the present disclosure. Figure 4 is a schematic diagram of a feature extraction module according to an embodiment of the present disclosure. Figure 5 is a schematic diagram of a multi-head attention mechanism according to an embodiment of the present disclosure. Figure 6 is a schematic diagram of a converter model according to an embodiment of the present disclosure. Figure 7 is a schematic diagram of the variation trend of the maximum mean difference induced by Gaussian kernels under different bandwidth parameters. Figure 8 is a schematic diagram of the process flow of the Tennessee-Eastman process. Figure 9 is a schematic diagram comparing the operation of the Tennessee-Eastman process under normal and fault conditions according to an embodiment of the present disclosure. Figure 10 is a schematic diagram of a cross-condition fault diagnosis confusion matrix according to an embodiment of the present disclosure. Figure 11 is a radar schematic diagram showing the comparison results of fault diagnosis time of different deep transfer learning methods according to an embodiment of the present disclosure in the Tennessee-Eastman process.

[0041] The cross-unit fault diagnosis method for chemical processes (hereinafter referred to as the diagnosis method) provided in this disclosure is illustrated below with reference to Figures 1 to 11. This diagnosis method includes not only the actual fault diagnosis method based on a multi-channel domain adaptation network, but also the training method for the multi-channel domain adaptation network. The multi-channel domain adaptation network, i.e., the fault diagnosis model, includes a feature extraction module and a domain adaptation module. The number of feature extraction modules can be set to one or more according to actual needs.

[0042] As shown in Figure 1, the cross-unit fault diagnosis method for chemical processes provided in this embodiment may include the following steps S101 to S103 in the online diagnosis stage for a fault diagnosis model. Steps S101 to S103 enable cross-domain transfer of fault category information, achieving high-accuracy and low-false-alarm-rate fault diagnosis for the equipment to be diagnosed.

[0043] Step S101: Obtain the fault diagnosis task for the equipment to be diagnosed under the target process flow.

[0044] The target process flow can be determined based on the process flow of interest in the actual chemical process. Correspondingly, the equipment to be diagnosed can also be determined based on the process flow of interest in the actual chemical process, and the number can be one or multiple.

[0045] Step S102: Determine the target domain data required to perform the fault diagnosis task based on the fault diagnosis task and the device to be diagnosed.

[0046] The target domain data includes various operating data of the device to be diagnosed operating under the first target domain. The specific operating data can be determined according to the actual process flow and the actual device to be diagnosed in the chemical process. This disclosure does not limit this.

[0047] The first target domain is also determined based on the target process flow. The first target domain used in practical applications and the second target domain involved in training are essentially both target domains, meaning new datasets from which knowledge learned from the source domain needs to be applied. Typically, the data distribution of the target domain differs from that of the source domain, and it may have little or no labeled data. The source domain, in transfer learning, refers to the dataset with labeled data, usually used to train the model to learn useful features and knowledge. Correspondingly, target domain data refers to the dataset from which prediction or classification needs to be performed. Specifically, multi-channel domain adaptation networks, i.e., fault diagnosis models, need to apply knowledge learned from the source domain to infer target domain data and target sample data. Multi-source domain data, also known as the source domain sample dataset (hereinafter referred to as the source domain sample dataset), refers to datasets from multiple different source domains. Multi-source domain data can provide more diversity and richer information, helping to improve the learning effect of the target domain.

[0048] Step S103: Input the target domain data into the trained multi-channel domain adaptation network for calculation to obtain the diagnostic results for the fault diagnosis task.

[0049] In one possible implementation, step S103 may include: extracting subdomain features from the target domain data to obtain subdomain-specific features representing information unique to the first target domain; and extracting global features from the target domain data to obtain global universal features representing common information shared by the first target domain and various source domains. For example, subdomain-specific features and global universal features can be determined by a feature extraction module. Domain migration features are then determined based on the subdomain-specific features and global universal features. Fault diagnosis is performed based on the domain migration features to obtain a diagnostic result for the fault diagnosis task. For example, the total features determined by the subdomain-specific features and global universal features can be processed based on a domain adaptation module to obtain the final diagnostic result. In this way, by extracting subdomain-specific features and global universal features from the target domain data respectively, the characteristics of the target domain data can be captured more comprehensively. Subdomain-specific features cover unique information specific to the first target domain, while global universal features cover common information between the target domain and multiple source domains. This fusion of dual features can more accurately identify fault modes in the first target domain, thereby improving the accuracy of fault diagnosis.

[0050] In one possible implementation, step S103, which involves extracting sub-domain features from the target domain data to obtain sub-domain-specific features representing information unique to the target domain, and extracting global features from the target domain data to obtain global universal features representing common information shared by the target domain and all source domains, may include: extracting sub-domain features at a time scale from the target domain data to obtain sub-domain temporal features; extracting sub-domain features at a spatial scale from the sub-domain temporal features to obtain sub-domain-specific features; and extracting global features at a time scale from the target domain data to obtain global temporal features, and extracting global features at a spatial scale from the global temporal features to obtain global-specific features. The process of extracting sub-domain features at a time scale can be implemented based on a Transformer model, and the process of extracting global features at a spatial scale can be implemented based on a Multi-Scale Convolutional Neural Network (MSCNN), as detailed below. In this way, by extracting features from the target domain data at both the time and spatial scales, more detailed and comprehensive feature information can be captured. Feature extraction at the time scale helps to understand the changing trends and periodicity of data over time, while feature extraction at the spatial scale can reveal the patterns and differences in the spatial distribution of data. This multi-dimensional feature extraction method significantly enhances the diagnostic model's ability to capture both target domain-specific and common information, thereby improving the accuracy of fault diagnosis.

[0051] In this way, by acquiring real-time operational data such as target domain data in chemical processes, extracting features from the target domain data to obtain subdomain-specific features and global general features, determining domain migration features, and determining diagnostic results based on domain migration features, a series of steps can achieve high-accuracy and low-false-report fault diagnosis on multiple different units in similar processes or under different operating conditions. The extraction of global general features not only helps the multi-channel domain adaptation network perform better in the target domain, but also promotes the migration and sharing of knowledge from the source domain to the target domain. These general features reflect the common rules and characteristics between different domains, providing strong support for the rapid adaptation of the multi-channel domain adaptation network in new fields. At the same time, by combining subdomain-specific features and global general features for fault diagnosis, the multi-channel domain adaptation network can more accurately identify specific fault modes in the target domain, improve the pertinence and effectiveness of diagnosis, and thus enhance the model's cross-unit versatility.

[0052] As shown in Figure 2, the cross-unit fault diagnosis method for chemical processes provided in this embodiment of the present disclosure further includes a training process for a multi-channel domain adaptation network. The training process includes the following steps S201 to S203 in the offline modeling stage for the fault diagnosis model. Through steps S201 to S203, key features can be effectively extracted from the data in the source domain and the second target domain, and the domain adaptation mechanism can be used to reduce the differences between domains, thereby improving the fault diagnosis capability of the multi-channel domain adaptation network in the target domain. This cross-domain adaptation capability enables the multi-channel domain adaptation network to maintain high diagnostic accuracy in different target domains. At the same time, the use of a multinomial kernel function to calculate the feature distribution difference function value helps to quantify the difference between the feature distributions of the source domain and the target domain, thereby guiding the multi-channel domain adaptation network to gradually reduce this difference during the training process. This allows the multi-channel domain adaptation network to not only perform well on the training set, but also maintain high generalization performance on unseen test sets.

[0053] Step S201: Obtain the target domain sample dataset and the source domain sample dataset.

[0054] The target domain sample dataset includes target domain sample data from multiple different sample devices operating in multiple different second target domains. These multiple different sample devices and multiple different second target domains can be determined based on similar or different process flows in chemical engineering. Taking sample device A as an example, target domain sample data for sample device A under each second target domain can be obtained. This data may include one or more types of operational data.

[0055] The source domain sample dataset includes source domain sample data for each sample device operating in multiple different source domains. These multiple different sample devices and multiple different source domains can be determined based on similar or different process flows in chemical engineering. Taking sample device B as an example, source domain sample data for sample device B in various source domains can be obtained. This data may include one or more types of operational data.

[0056] Each target domain sample data and each source domain sample data includes multiple normal operation data and multiple abnormal operation data, so that the multi-channel domain adaptation network can perform transfer learning based on these normal operation data and abnormal operation data.

[0057] The number and type of sample devices, the number of second target domains, the number of source domains, and the type and quantity of running data can all be flexibly set according to actual needs.

[0058] Step S202: Label each normal operation data and each abnormal operation data in the source domain sample dataset to obtain the source domain labeled dataset. Construct a training sample set based on the source domain labeled dataset and the target domain sample dataset.

[0059] By labeling only the source domain sample dataset, it helps to enhance the adaptability of multi-condition networks when using multi-channel domain adaptation networks for actual diagnosis, and can also cope with scenarios where there is little fault data and a lack of labels in the target domain that needs attention.

[0060] Step S203: Train the initialized multi-channel domain adaptation network based on the preset loss function and training sample set to obtain the trained multi-channel domain adaptation network.

[0061] The loss of a multi-channel domain adaptation network can include three parts: domain class loss (source domain loss and target domain loss), relation loss, and feature distribution difference loss. Correspondingly, the loss function can include the source domain loss function and target domain loss function of the feature extraction module, the relation loss function of the domain adaptation module, and the feature distribution difference function constructed based on the polynomial kernel-induced maximum mean discrepancy (PK-MMD).

[0062] The parameters of the multi-channel domain adaptation network can be initialized according to actual needs, and this embodiment of the disclosure does not limit this. During the training of the multi-channel domain adaptation network, the parameters of the current multi-channel domain adaptation network can be updated based on the specific value corresponding to the loss function. In one possible implementation, step S203 may include: processing the training sample set using the feature extraction module to obtain the processing result; determining the source domain loss function value and target domain loss function value of the feature extraction module based on the processing result; the processing result may include source domain features, target domain features, source domain classification result, and target domain classification result; processing the source domain features and target domain features using the domain adaptation module to obtain the fault category prediction result of each abnormal operation data in the target domain sample data; determining the relational loss function value of the domain adaptation module based on the fault category prediction result; determining the feature distribution difference function value using the multinomial kernel function and the training sample set; determining the total loss function value of the current multi-channel domain adaptation network based on the source domain loss function value, target domain loss function value, relational loss function value, and feature distribution difference function value; and updating the parameters of the current multi-channel domain adaptation network based on the total loss function value. In this way, by comprehensively considering the source domain loss function value, target domain loss function value, relational loss function value, and feature distribution difference function value to determine the total loss function value, the multi-channel domain adaptation network can comprehensively evaluate the performance of the fault diagnosis model during training and update the parameters based on these evaluation results. This multi-level evaluation mechanism helps to improve the reliability and accuracy of the diagnostic results.

[0063] The calculation process of the loss function is illustrated below using an example structure of a multi-channel domain adaptation network. In some embodiments, as shown in Figure 3, the multi-channel domain adaptation network structure includes two main modules: a feature extraction module and a domain adaptation module.

[0064] This diagnostic method incorporates a novel feature extraction module, Transformer-MSCNN, within a multi-channel domain adaptation network. This module is formed by concatenating a Transformer model with a multi-scale convolutional neural network. For example, as shown in Figure 4, normal operation data (or normal samples) and abnormal operation data (or fault samples with labels) from the source domain sample data are input into the feature extraction module via different channels. The temporal features determined by the Transformer are processed by a global general feature extraction channel and a sub-domain specific feature extraction channel to form a feature map containing key information about the source domain. This provides strong source domain knowledge support for subsequent sample classification tasks. The global general feature extraction channel aims to summarize common features between domains through domain adaptation, while the sub-domain specific feature extraction channel extracts domain-specific discriminative features to facilitate fault classification and domain label classification. Transformer and MSCNN focus on the sample data at both temporal and spatial scales, respectively, to capture richer feature information across different dimensions. The Transformer primarily focuses on the global relationships between different variables, which helps extract cross-domain shared information in transfer tasks. MSCNN focuses on local features of data, complementing the ability of multi-channel domain adaptive networks to learn data features at different scales.

[0065] In the feature extraction module of this embodiment, an attention mechanism is also introduced to enhance the multi-channel adaptive network's understanding of data features. The multi-head attention mechanism used is another variation of the self-attention mechanism. The multi-head attention mechanism executes multiple self-attention unit operations in parallel on the input data. Each self-attention operation can be considered a new subspace. Its workflow is shown in Figure 5, consisting of three steps: input data segmentation, attention weight calculation, and feature fusion and output. The initial calculation of the multi-head attention mechanism is similar to the original self-attention mechanism. The query, key, and value vectors (referred to as Q, K, and V vectors) obtained after linear transformation of the original data sequence are linearly transformed by their respective transformation matrices and then input into the scaled dot-product attention module. The difference is that its dot-product calculation needs to be performed h times, where h is the number of attention heads. Subsequently, the mechanism calculates the attention scores between vectors and performs a weighted summation of the values. Then, it concatenates all the scaled dot-product results and outputs them after a linear transformation to obtain the feature results enhanced by the multi-head attention mechanism. This mechanism endows multi-channel adaptive networks with the ability to learn their correlation information within different representation subspaces, and is also key to the Transformer model's ability to capture complex variable relationships. For example, the calculation of scaled dot product attention is as follows:

[0066] In equation (1), h i represents the output of the i-th attention head; f represents the scaled dot product attention function, which calculates the attention weights by performing a dot product on the input Q, K, and V vectors, and then performs a weighted summation on the V vector to generate the output; W i q This represents the Query transformation matrix, used to linearly transform the input Q vector to a subspace; This represents the Key transformation matrix, used to linearly transform the input K vector to a subspace; denoted as Value transformation matrix, used to linearly transform the input V vector to the subspace; q represents the Query vector, which is used to match the key in the attention mechanism to calculate the attention weights; k represents the Key vector, which is matched with q to calculate the attention weights; v represents the Value vector, which is the vector of the final weighted sum to obtain the output result.

[0067] In some embodiments, as shown in Figure 6, the main structure of the Transformer model consists of an encoder and a decoder. The encoder processes the input sequence, and the decoder generates the output sequence (emphasizing temporal features). Both are formed by stacking feedforward neural networks (FFNs) using a multi-head attention mechanism. The self-attention layer enhances the model's representational ability by learning information from different positions in the sequence in parallel. Residual connections are added at the output of each sub-layer of the Transformer model to avoid the gradient vanishing problem during training. The decoder outputs the probability distribution of each output element, thereby analyzing the positional information and importance of each element in the input sequence. Regarding hyperparameter settings, the number of attention heads in both the encoder and decoder of the Transformer can be set to 3. The convolutional kernels of MSCNN can be set according to actual needs. Fully connected layers are connected before and after to integrate the features extracted from each unit. The learning rate is 0.001 by default, and the network parameters are updated through backpropagation.

[0068] The loss function of the feature extractor module can be summarized by the general representation shown in equation (2):

[0069] In equation (2), L′ represents the loss function of the feature extractor module, where n is the total number of samples, x represents the dimension of the variable, and y is the true label of the sample. Assign classification labels to the samples.

[0070] In some embodiments, the feature extraction module can arrange an equal number of branch channels according to the total number of source domains. Each channel is further divided into a subdomain-specific feature extraction channel and a global general feature extraction channel, used to simultaneously extract cross-domain common features and the unique features of each source domain from multi-source domain data. The types of branch channels can include source domain branches and target domain branches as shown in Figure 3. Both source domain branches and target domain branches include corresponding subdomain-specific feature extraction channels and global general feature extraction channels. The subdomain-specific feature extraction channels in the source domain branches are used to extract subdomain features (or unique features) of the data input to that source domain branch to obtain the corresponding source domain-specific features. The global general feature extraction channels in the source domain branches are used to extract global features (or cross-domain common features) of the data input to that source domain branch to obtain the corresponding source domain general features. Source domain features can include source domain-specific features and source domain general features. The subdomain-specific feature extraction channel in the target domain branch is used to extract subdomain features from the data input to that target domain branch to obtain the corresponding target domain-specific features. The global general feature extraction channel in the target domain branch is used to extract global features from the data input to that target domain branch to obtain the corresponding target domain general features. Target domain features can include both target domain-specific and target domain general features. Whether in the online diagnostic phase or the offline modeling phase, essentially, source domain features represent the source domain aspect features of the corresponding source domain, and target domain features represent the target domain aspect features of the corresponding target domain. Source domain-specific features in source domain features represent information unique to the corresponding source domain, and target domain-specific features in target domain features represent information unique to the corresponding target domain. The specific source and target domains are determined based on the actual situation. The number of branch channels in the feature extraction module can be set according to actual needs.

[0071] The domain class loss of the feature extraction module, one of the loss functions, is determined by the source domain loss function and the target domain loss function. Step S203, which involves processing the training sample set using the feature extraction module to obtain processing results, and determining the source domain loss function value and target domain loss function value of the feature extraction module based on these results, may include: using the source domain branch in the feature extraction module to extract features from the data of each source domain in the source domain sample data, obtaining source domain features for each source domain; inferring the source domain classification result for the corresponding source domain based on the source domain features; and determining the source domain loss function value of the feature extraction module based on the source domain classification results and the true domain labels of each source domain; and using the target domain branch in the feature extraction module to extract features from the target domain sample data, obtaining target domain features for each second target domain; inferring the target domain classification result for the corresponding second target domain based on the target domain features; and determining the target domain loss function value of the feature extraction module based on the target domain classification results and the true domain labels of each second target domain.

[0072] For example, the source domain loss function value of the feature extraction module can be determined by the following equation (3):

[0073] In equation (3), L ds Let n represent the source domain loss function. bs This represents the total number of samples used in a specific source domain during a single training session. This represents the source domain classification result (or genus classification result) of the i-th sample in the source domain. This represents the true domain label of the i-th sample in the source domain. A sample belonging to this specific source domain, i.e., the source domain being correctly classified, indicates that the sample belongs to that specific source domain. When the source domain is misclassified

[0074] The target domain branch of the feature extraction module directly processes the target domain sample data for testing fault classification accuracy, without processing the source domain data. The source domain data is specifically used for parameter adjustment during training, such as adjusting the parameters of the diagnostic model through backpropagation. For example, the target domain loss function value of the feature extraction module can be determined by the following equation (4):

[0075] In equation (4), L dt Let n represent the target domain loss function. bt This represents the total number of samples used for a specific target domain during a single training session. This represents the target domain classification result (or genus classification result) of the i-th sample in the target domain. This represents the true domain label of the i-th sample in the target domain, when the sample in the target domain is correctly classified. When the target domain is misclassified

[0076] This diagnostic method also incorporates a domain adaptation module within the multi-channel domain adaptation network. After feature extraction and filtering, the contributions of all samples unsuitable for migration are eliminated, while key information is assigned greater weight. For example, the multi-channel data features output by the feature extraction module can be weighted and summed using the following formula (5) to form a feature map for classification by the domain adaptation module:

[0077] In equation (5), f s,t The feature map calculated by the domain adaptation module represents the comprehensive features of the entire feature space; This represents the feature map output by the i-th feature extraction module, which is determined jointly by Transformer and MSCNN. For example, the input data is first processed by Transformer to obtain temporal features, and then the temporal features are processed by MSCNN to obtain the output results of the feature extraction modules corresponding to Transformer and MSCNN. This output result may include subdomain-specific features and global general features. In the online diagnosis stage of the fault diagnosis model, the subdomain-specific features and global general features under the first target domain can be determined based on the trained Transformer and MSCNN. Thus, the output results of the feature extraction module are obtained based on the subdomain-specific features and global general features. Then, based on the actual number of feature extraction modules set, the feature map, i.e., the domain transfer features, calculated by the domain adaptation module is determined.

[0078] The last fully connected layer in the domain adaptation module, together with the normalization exponent (softmax) function, constitutes the sample classifier, which can be expressed as equation (6):

[0079] In equation (6), represents the label of the i-th fault category obtained from the diagnosis; c represents the index of the fault category, with the number of categories ranging from 1 to C; f represents the preset label for the i-th fault category. s,t Representation domain transfer features, This represents the similarity score between the feature map domain transfer features and the preset fault category label on fault category c.

[0080] The relational loss of the domain adaptation module, one of the loss functions, is determined by the relational loss function. Step S203 involves processing the source and target domain features using the domain adaptation module to obtain the fault category prediction results for each abnormal operation data in the target domain sample data. Based on these fault category prediction results, the relational loss function value of the domain adaptation module is determined. This may include: calculating the target and source domain features using the domain adaptation module to obtain the domain adaptation sample features; determining the fault category prediction results for the target domain sample data based on the domain adaptation sample feature degree and the predetermined fault category; and determining the relational loss function value of the domain adaptation module based on the fault category prediction results and the actual fault labels of the target domain sample data. For example, the relational loss function value of the domain adaptation module can be determined by the following formula (7):

[0081] In equation (7), L r Represents the relational loss function, n bs This represents the total number of samples used in a specific source domain during a single training session. This represents the label of the i-th fault category obtained from the diagnosis. This represents the i-th real fault label in the target domain sample data.

[0082] Since subdomain-specific features are more difficult to align with global general features, this diagnostic method simultaneously introduces a polynomial kernel-induced maximum mean difference algorithm between multilayer networks in different channels to achieve feature alignment. For example, as shown in Figure 3, PK-MMD calculation can be introduced at three points in the multichannel domain adaptation network, namely PK-MMD1, PK-MMD2, and PK-MMD3. In some embodiments, PK-MMD1 can be placed on the last convolutional network of the subdomain-specific feature extraction channel within the feature extraction module; PK-MMD2 can be placed between the first fully connected layer network pair and PK-MMD3 can be placed between the second fully connected layer network pair within the domain adaptation module, which includes the first fully connected layer network pair and the second fully connected layer network pair. The first fully connected layer network pair can be processed based on the source domain sample data to obtain the corresponding output results, and the second fully connected layer network pair is processed similarly. The specific processing procedure of the fully connected layer network pair is determined according to the actual structure of the diagnostic model.

[0083] One of the loss functions, the feature distribution difference function, may include three corresponding loss functions: PK-MMD1, PK-MMD2, and PK-MMD3. PK-MMD1 corresponds to the first feature distribution difference function, PK-MMD2 to the second feature distribution difference function, and PK-MMD3 to the third feature distribution difference function. The feature distribution difference function measures the distribution difference between source domain sample data in different source domains. The feature distribution difference function is used as one of the loss functions to adjust the weights of the diagnostic model. The first, second, and third feature distribution difference functions can be calculated synchronously after the weights of each layer in the diagnostic model are determined to update the parameters of the diagnostic model. This embodiment does not limit the order in which they are calculated.

[0084] For example, step S203, determining the feature distribution difference function value using a multinomial kernel function and the training sample set, may include: determining the first feature distribution difference function value based on the source domain sample data according to the definition rules of the multinomial kernel function; determining the second feature distribution difference function value based on the output of the first fully connected layer network pair; and determining the third feature distribution difference function value based on the output of the second fully connected layer network pair; and determining the feature distribution difference function value based on the first, second, and third feature distribution difference function values. The output of the first fully connected layer network pair is determined based on the source domain sample data, and the output of the second fully connected layer network pair is determined similarly. In this example, the first, second, and third feature distribution difference function values ​​can be calculated simultaneously.

[0085] The entire training process needs to minimize the loss function to output the trained multichannel adaptation network. In some embodiments, during the training process, corresponding weight functions λ can be preset for the three parts of loss: domain class loss, relation loss, and feature distribution difference loss. Since PK-MMD is used in three places in the multichannel domain adaptation network, the weighted total loss function of the multichannel adaptation network can be expanded as follows (8): L=λ0L r +λ1L PK-MMD1 +λ2L PK-MMD2 +λ3L PK-MMD3 +λ4(L ds +L dt Equation (8)

[0086] In equation (8), L represents the total loss function, and λ0 represents the relationship loss function L. r The corresponding weights, λ1 represents the weights corresponding to the first feature distribution difference function, L PK-MMD1 Let L represent the difference function of the first feature distribution, λ2 represent the weights corresponding to the difference function of the second feature distribution, and L... PK-MMD2 Let L represent the difference function of the second characteristic distribution, λ3 represent the weights corresponding to the difference function of the third characteristic distribution, and L represent the difference function of the second characteristic distribution. PK-MMD3 Let λ represent the third characteristic distribution difference function, and λ₄ represent the source domain loss function L. ds and target domain loss function L dt The corresponding weights are determined. The total loss value is obtained through equation (8) above. After backpropagation, the model weights are updated, and this process is repeated until the total loss defined by equation (8) is minimized, at which point the diagnostic model training is complete. During model testing, the model weights remain unchanged. The classification accuracy is calculated by testing with target domain sample data.

[0087] The PK-MMD used in the embodiments of this disclosure will now be described. Maximum Mean Discrepancy (MMD) is a commonly used statistical distance metric in transfer learning, used to measure the difference between the probability distributions of data in the source and target domains. For a given source domain... and target domain Their probability distributions are X and Y, respectively, and the maximum mean difference is defined by the following equation (9):

[0088] In equation (9), φ(·) represents the kernel function, which maps the input x to a Hilbert space. In practical applications, since the probability distributions X and Y of the source and target domains are difficult to express explicitly or are unknown in the task, MMD is often approximated by empirical estimation of samples, such as the following equation (10):

[0089] In equation (10), n and m refer to the source domain, respectively. and target domain The number of samples in the feature space, k(·) represents a positive definite kernel function used to calculate the similarity of sample pairs in the feature space.

[0090] MMD has evolved into various variants by selecting different kernel functions, among which Gaussian kernel-induced maximum mean discrepancy (GK-MMD) is the most commonly used. However, it has three significant drawbacks in application. First, GK-MMD relies on the average distance for calculation, neglecting the influence of higher-order moments, which limits the accuracy of calculating the differences in the distribution of transferable features, thus affecting the overall quality of transfer learning. Second, GK-MMD has high computational complexity, as its computational complexity is quadratic with the number of samples, i.e., O(n^2). 2 +2mn+m 2 Handling large-scale datasets leads to significant memory consumption and high computational costs. Thirdly, the algorithm is highly sensitive to the Gaussian kernel parameter settings. The performance of the Gaussian kernel depends on the choice of the bandwidth parameter, lacking the flexibility to adjust other parameters. Inappropriate bandwidth parameter selection can lead to overfitting or underfitting. As shown in Figure 7, assuming the two domain datasets satisfy Gaussian distributions N(0,1) and N(μt,1) respectively, the GK-MMD metric results show significant fluctuations with small changes in kernel width. Therefore, models based on GK-MMD struggle to converge at the optimal parameter point during distribution adaptation.

[0091] To overcome the aforementioned shortcomings of this diagnostic method, this disclosure proposes a polynomial kernel-induced maximum mean difference as a measure of differences in data distribution across different domains. By replacing the Gaussian kernel function with a polynomial kernel function, the input features can be mapped to a higher-dimensional space, allowing for the calculation of dot products to capture nonlinear relationships in the data. Furthermore, the algorithm performance can be intuitively adjusted by changing the kernel parameters. For example, the polynomial kernel function form shown in equation (11) can be used: k(x,y)=(ax T y+b) c c = 1, 2, ... Equation (11)

[0092] In equation (11), k(x,y) represents the polynomial kernel function, used to calculate the kernel value between input vectors x and y. It maps the input features to a high-dimensional space to capture the nonlinear relationship of the data; the slope a represents the contribution of the higher-order moments to the difference in sample distribution. The larger the slope value, the greater the influence of the higher-order moments on the distribution difference; the intercept b is a constant term and is usually set to 1; c represents the order of the polynomial, which determines the dimension of the feature space mapped to; k(·) represents the polynomial of the element-wise product between input vectors x and y. Thus, we can obtain That is, the calculation formula for PK-MMD, in which x i x j y i y j This represents the feature vector of the source or target domain, which is determined based on the actual processing procedure. According to the binomial theorem, the empirical estimate of PK-MMD can be expressed as: When c = 1, the average distance is When c>1, the calculated value is the weighted sum of the distances between the central moments of each order of the sample pair. The difference in the expected value of the polynomial kernel function under the two sample distributions reflects the similarity between the sample pairs. The computational complexity of PK-MMD is O(c(c+3) / 2+ε(n+m)), where O(ε) represents the time cost of matrix multiplication. The polynomial kernel, by mapping to a higher-dimensional space to extract the nonlinear features of the data, makes PK-MMD perform better when handling complex datasets. Its flexible and variable parameters also provide opportunities to adjust the model complexity; choosing appropriate kernel parameters can balance the model's metric accuracy and computational efficiency.

[0093] This disclosure proposes a novel cross-unit fault diagnosis method for chemical processes. It utilizes Transformer-MSCNN for cross-domain feature extraction, followed by feature filtering and linear transformation to convert the features into a feature map. This feature map then enters a domain adaptation module, calculating the similarity correlation between samples using shared weights. PK-MMD aligns the data distribution across domains, constructs a loss function, and adjusts the weights. This method effectively achieves cross-domain transfer of fault category information, demonstrating excellent recognition capabilities for various faults and exhibiting superior overall performance. It enables the diagnostic model to learn common features from different source domains, which is of significant practical importance for improving the model's general cross-unit diagnostic capabilities. This fault diagnosis method fully extracts fault data features from multiple operating conditions in the training sample set, transferring source domain features to the target domain. This allows the target operating condition to maintain high-precision fault identification even when training samples are scarce and lack label information.

[0094] The fault diagnosis workflow provided in this embodiment is divided into two stages: offline modeling and online diagnosis. Source domain data with fault labels is fed into a multi-channel domain adaptation network model for offline training. In the online stage, the pre-trained model is activated to minimize the difference in the distribution of target domain features and source domain features, thereby realizing the cross-domain transfer of fault category information.

[0095] In some embodiments, the training of a multi-channel domain adaptation network (i.e., a fault diagnosis model) can be achieved through the following offline modeling process. First, the structural parameters of the neural network corresponding to the feature extraction module in the multi-channel domain adaptation network are set according to actual needs, and its connection parameters are randomly initialized. The structural parameters of the feature extraction module may include the number of input layer units, the number of hidden layer units, the size of each batch of data, the training period using all samples in the training set, and random deactivation values ​​to avoid overfitting. The connection parameters may include weights, biases, and hyperparameters for Transformer and CNN architectures. Next, source domain sample data and target domain sample data (normal samples from the source domain, fault samples from the source domain, normal samples from the second target domain, and fault samples from the second target domain) are used as training data and input into the feature extraction module of the multi-channel domain adaptation network through different channels. After acquiring the training data, the feature extraction module focuses on both spatial and temporal aspects to capture feature information. After feature extraction and filtering, all samples whose contributions are unsuitable for transfer are removed, and key information is given higher weights. The multi-channel data features output by the feature extraction module are linearly transformed and converted into feature maps for classification by the domain adaptation module. Then, in the domain adaptation module, the similarity correlation between feature maps is calculated by sharing weights, and constraints are applied through a loss function to achieve cross-domain transfer of features under different operating conditions. The loss function can be constructed using a derivative of Maximum Mean Difference (MMD), namely PK-MMD. PK-MMD can be set at three points: the last convolutional layer in the feature extraction module, and between the two pairs of fully connected layers in the domain adaptation module. This allows for a better measurement of the distribution difference between source and target domain sample data. The last fully connected layer of the domain adaptation module, together with the softmax function, forms a sample classifier to classify the target domain sample data. The model is trained using training data collected under different operating conditions. By learning the data features in the training data, the optimal connection parameters of the multi-channel domain adaptation network are obtained. Next, the monitoring data of the validation set is input into the trained multi-channel domain adaptation network to calculate the confusion matrix, fault diagnosis time (FDT), and diagnostic accuracy. The multi-channel domain adaptation network is then evaluated to obtain a multi-channel domain adaptation network that meets the preset evaluation conditions.

[0096] In some embodiments, cross-unit fault detection in chemical processes can be achieved through the following online monitoring steps. First, real-time operating data of the unit to be diagnosed in the chemical process, i.e., target domain data, is acquired and standardized to obtain input data. Then, the input data is input to a feature extraction module for calculation to obtain a feature map. In the feature extraction module, subdomain-specific features representing the current target domain can be extracted using subdomain-specific feature extraction channels, and general features representing the source domain used in training the multi-channel domain adaptation network can be extracted using global general feature extraction channels. Transformer and MSCNN can be used to focus on different parts of the input data from both temporal and spatial perspectives, respectively. Finally, the feature map is input to a domain adaptation module for calculation to obtain the final diagnostic result, which indicates the fault status of the unit to be diagnosed.

[0097] Thus, the novel cross-unit fault diagnosis method for chemical processes proposed in this embodiment utilizes Transformer and MSCNN to extract cross-domain features from sample data in each domain. After feature filtering, the multi-channel features undergo linear transformation to convert them into feature maps. The feature maps enter the domain adaptation module, which calculates the similarity correlation between sample data by sharing weights. By aligning the data distribution of each domain through PK-MMD, constructing a loss function, and adjusting the weights, it can effectively achieve cross-domain transfer of fault category information. In particular, it has good identification capabilities for various faults, enabling the multi-channel domain adaptation network to learn common features from different source domains. The overall performance is excellent, which has significant practical implications for improving the model's general diagnostic capabilities across units.

[0098] To verify the feasibility and effectiveness of the method proposed in this disclosure, experiments were conducted on the Tennessee-Eastman Process (TEP) to validate its effectiveness. Performance comparisons were also performed with classic fault diagnosis methods: Deep Transfer Network Composed of Stacked Auto-Encoder (DTN), Deep Adaptation Network with CNN (DAN-CNN), Joint Adaptation Network with CNN (JAN-CNN), and Dynamic Adversarial Adaptation Network (DAAN).

[0099] The Tennessee-Eastman (TE) process is a classic industrial process simulation dataset in the field of fault diagnosis. By simulating real chemical production processes and introducing various fault conditions into the control environment, it has become a commonly used benchmark simulation model for multivariate process monitoring methods. The main chemical reactions involved in this process are as follows: A(g) + C(g) + D(g) → G(l), A(g) + C(g) + E(g) → H(l), A(g) + E(g) → F(l), and 3D(g) → 2F(l). This irreversible exothermic chemical reaction process includes four gaseous feedstocks, two liquid products, and one byproduct. The reaction environment is protected by an inert gas. The four gaseous feedstocks are A(g), C(g), D(g), and E(g), the two liquid products are G(l) and H(l), and the byproduct is F(l). The TE process encompasses five key process units, including a reactor, condenser, circulating compressor, gas-liquid separator, and stripping tower. The original TE dataset contained 12 operational variables, 19 component variables, and 22 process measurement variables. It could simulate 20 different types of fault conditions to test the classification performance of fault diagnosis methods. Subsequently, six new steady-state conditions were set up in addition to the basic operating conditions based on reaction yield and product composition ratio, and 19 control strategies were added to the TE process, which are commonly used in multi-condition fault diagnosis and process control optimization studies of chemical processes. Later, 8 process measurement variables and 24 component variables were added. The process flow diagram of the TE simulation model used in this example study of the TE process is shown in Figure 8 and Figure 9, which can be used to obtain the fluctuation trends of variables in the TE process under normal and fault conditions.

[0100] This example uses the basic operating conditions of the TE simulation for the experiment, with a data acquisition cycle of 3 minutes. Analysis of all monitored variables in the reaction process revealed that the component variables of the new model exhibited very small trends and low data acquisition frequency. Therefore, only 12 manipulated variables and 30 continuously measured variables were selected for modeling. All variables are labeled with their serial numbers in Figure 8 to more intuitively illustrate the measurement locations and potential coupling relationships. Due to the low sampling frequency of the component variables, this example only selects manipulated variables and continuously measured variables to build the dataset, totaling 42 variables. Since the control system of the latest TEP model can effectively control faults 15, 16, 23, and 28, these four types of disturbances are not considered (i.e., Table 1 does not include the disturbance information corresponding to disturbance serial numbers 15, 16, 23, and 28). The disturbances studied in this example are shown in Table 1 below. Table 1 illustrates the various disturbances in the TEP dataset.

[0101] Table 1. Introduction to various perturbations in the TEP dataset.

[0102] The TE process was used as a case study to validate the method's effectiveness in cross-condition transfer fault diagnosis. TE process simulation data for five different operating conditions were collected and used in the experiment. For the transfer learning experiment, labeled data from one operating condition was selected as the training set to diagnose the operational status of other unlabeled operating conditions.

[0103] Table 2 shows the accuracy of MSDAN in unsupervised fault diagnosis of the TE process across the target domain under different operating conditions. Based on the results in the table, the average diagnostic rate of the MSDAN method in all experiments was calculated to be 95.7%. to Representing five source domains, to The five target domains are represented. Observing the results in Table 2, it can be found that MSDAN performs well in pairwise migration and mutual diagnosis tasks in most operating conditions, and the fault diagnosis accuracy of each migration experiment exceeds 90%, indicating that the model maintains strong robustness in complex scenarios across operating conditions.

[0104] Table 2. Accuracy of MSDAN in unsupervised fault diagnosis across target domains of TE processes.

[0105] Figure 10 shows the fault diagnosis confusion matrix of MSDAN with condition 1 as the training set and condition 4 as the test set. Only faults 9 and 15, which are difficult to diagnose, did not reach 90% accuracy. From the raw data, fault 9 only caused random changes in the feed temperature (represented by variable 24), and fault 15, representing condenser cooling water valve stickiness, did not cause drastic fluctuations in the measured variables. Therefore, the characteristics of most variables for these two faults are similar to those under normal conditions, making accurate diagnosis difficult. However, the experimental results also demonstrate that the remaining 18 faults were accurately identified by the deep transfer network, and the overall diagnostic accuracy is satisfactory.

[0106] Furthermore, this example defines a new performance metric for fault diagnosis methods, namely FDT. Fault diagnosis time is defined as the total time elapsed from the introduction of variable disturbances into the system and the generation of fault data characteristics until the type of fault is correctly identified three times consecutively. In actual chemical production, the process flow is large and complex. Faced with a large amount of data and control signals in the process monitoring system, operation engineers often only pay attention to the alarm information and consider what control measures to take to deal with the abnormal event after the system has issued alarms for the same type of fault multiple times. At the same time, three consecutive correct classifications also indicate that the model's diagnostic effect on this type of fault tends to be stable, giving operators more confidence to take action based on the model's output results. Table 3 below shows the comparison of fault diagnosis time (min) results of different deep transfer learning methods in the TE process. It shows the comparison of fault diagnosis time of different deep transfer learning methods when trained and transferred to the TE process from condition 1 to condition 4. The results show that MSDAN has the shortest average diagnosis time and can make accurate judgments from the beginning of the fault for most fault types.

[0107] Table 3 Comparison of Fault Diagnosis Time (min) for Different Deep Transfer Learning Methods in the TE Process

[0108] To make the fault diagnosis time indicator more intuitive, the longest fault diagnosis time for each type of fault was used as the benchmark, and the diagnosis times for other faults of the same type were standardized and plotted as a radar chart as shown in Figure 11. In Figure 11, the closer the closed curve is to the center, the faster the diagnosis speed. It can be clearly seen that MSDAN's fault diagnosis is the most efficient, and there is no case where the diagnosis time for a certain type of fault is significantly longer than other methods. This shows that the MSDAN model has the ability to identify faults earlier and more stably while ensuring diagnostic accuracy.

[0109] Given the superior classification performance of MSDAN by extracting source and target domain features separately using multiple channels, this example further extends the MSDAN model structure by enriching the training set samples of the source domain with other operating condition data not used in pairwise transfer experiments. More source domain feature extraction and domain adaptation channels are added to the network structure to evaluate the model's feature learning ability for multiple source domain data. Table 4 shows the accuracy results of different depth transfer learning models in multi-source domain cross-operating condition fault diagnosis tasks. Taking 94.5% in Table 4(a) as an example, 94.5% represents the diagnostic accuracy when the model is transferred to target domain 2 after training with data from source domain (0+1). The results demonstrate that the overall classification performance tends to improve when the model is trained with more source domain data. In most cases, MSDAN can fully learn the fault features of different operating conditions represented by multiple source domains, improving the fault diagnosis accuracy for the target operating condition and performing the best among all methods. Overall, MSDAN also has strong flexibility in model design. When the experiment includes more data from different domains with similar feature distributions, MSDAN can be extended into an enhanced model that learns more effective features from multi-source domain data and transfers them to the target domain task, thus providing design ideas for building a more general fault diagnosis model.

[0110] Table 4. Accuracy results of transfer learning models with different depths in multi-source domain cross-condition fault diagnosis tasks.

[0111] This disclosure proposes a cross-unit fault diagnosis method for chemical processes based on a Multi-channel Domain Adaptive Network (MSDAN). It not only introduces a novel feature extraction network built on Transformer-MSCNN, employing a modified polynomial kernel-induced maximum mean difference (PK-MMD) as a measure of the difference in probability distributions between source and target domain data, but also proposes that during training, source and target domain data are fed into different channels in parallel as input to the MSDAN. Simultaneously, normal and fault samples from the source domain are also labeled and fed into different channels of the feature extractor to extract common inter-domain features and specific domain discriminative features for fault classification and domain label classification. Furthermore, in the online diagnosis stage, the difference in feature distributions between the target and source domains is minimized, achieving cross-domain transfer of fault category information. The cross-unit fault diagnosis method for chemical processes proposed in this disclosure significantly reduces the diagnostic model's dependence on labeled information, lowers the demand for manually labeled data, and can fully learn the fault characteristics of different operating conditions represented by multiple source domains, improving the fault diagnosis accuracy for the target operating condition represented by the target domain. This provides a foundation for subsequent fault diagnosis research on cross-unit industrial cases.

[0112] Figure 12 shows a block diagram of a cross-unit fault diagnosis device for chemical processes based on a multi-channel domain adaptation network according to an embodiment of the present disclosure. As shown in Figure 12, the cross-unit chemical process 100 includes an acquisition module 101, a determination module 102, a calculation module 103, and a training module 104. The acquisition module 101 is used to acquire a fault diagnosis task for the equipment to be diagnosed under a target process flow. The determination module 102 is used to determine the target domain data required to perform the fault diagnosis task based on the fault diagnosis task and the equipment to be diagnosed. The target domain data includes various operating data of the equipment to be diagnosed operating under a first target domain, which is determined according to the target process flow. The calculation module 103 is used to input the target domain data into the trained multi-channel domain adaptation network for calculation to obtain a diagnosis result for the fault diagnosis task. Training module 104 is used for: acquiring a target domain sample dataset and a source domain sample dataset, wherein the target domain sample dataset includes target domain sample data of multiple different sample devices operating in multiple different second target domains, and the source domain sample dataset includes source domain sample data of each of the sample devices operating in multiple different source domains, wherein each of the target domain sample data and each of the source domain sample data includes multiple normal operation data and multiple abnormal operation data, and each of the sample devices, each of the second target domains, and each of the source domains are determined according to different process flows in the chemical process; labeling each normal operation data and each of the abnormal operation data in the source domain sample dataset to obtain a source domain labeled dataset; constructing a training sample set based on the source domain labeled dataset and the target domain sample dataset; training the initialized multi-channel domain adaptation network based on a preset loss function and the training sample set to obtain a trained multi-channel domain adaptation network, wherein the loss function includes a feature distribution difference function constructed based on the maximum mean difference induced by a multinomial kernel.

[0113] In one possible implementation, the multi-channel domain adaptation network includes a feature extraction module and a domain adaptation module. The initialization of the multi-channel domain adaptation network is trained based on a preset loss function and the training sample set to obtain a trained multi-channel domain adaptation network. This includes: processing the training sample set using the feature extraction module to obtain a processing result; determining the source domain loss function value and target domain loss function value of the feature extraction module based on the processing result, wherein the processing result includes source domain features, target domain features, source domain classification result, and target domain classification result; processing the source domain features and target domain features using the domain adaptation module to obtain fault category prediction results for each abnormal operation data in the target domain sample data; determining the relational loss function value of the domain adaptation module based on each fault category prediction result; determining the feature distribution difference function value using a polynomial kernel function and the training sample set; determining the total loss function value of the current multi-channel domain adaptation network based on the source domain loss function value, the target domain loss function value, the relational loss function value, and the feature distribution difference function value; and updating the parameters of the current multi-channel domain adaptation network based on the total loss function value.

[0114] In one possible implementation, the feature extraction module includes a source domain branch and a target domain branch; wherein, processing the training sample set using the feature extraction module to obtain processing results, and determining the source domain loss function value and target domain loss function value of the feature extraction module based on the processing results, includes: using the source domain branch to extract features from the data of each source domain in the source domain sample data to obtain source domain features of each source domain, inferring the source domain classification result of the corresponding source domain based on the source domain features of each source domain, and determining the source domain loss function value of the feature extraction module based on the source domain classification result and the true domain label of each source domain; using the target domain branch to extract features from the target domain sample data to obtain target domain features of each second target domain, inferring the target domain classification result of the corresponding second target domain based on the target domain features, and determining the target domain loss function value of the feature extraction module based on the target domain classification result and the true domain label of each second target domain.

[0115] In one possible implementation, the domain adaptation module processes the source domain features and the target domain features to obtain fault category prediction results for each abnormal operation data in the target domain sample data. Based on each fault category prediction result, the relational loss function value of the domain adaptation module is determined. This includes: using the domain adaptation module to calculate the target domain features and the source domain features to obtain domain-adapted sample features; determining the fault category prediction results for the target domain sample data based on the domain-adapted sample feature degree and a predetermined fault category; and determining the relational loss function value of the domain adaptation module based on the fault category prediction results and the true fault labels of the target domain sample data.

[0116] In one possible implementation, the source domain features include source domain-specific features representing information unique to the source domain, the target domain features include target domain-specific features representing information unique to the second target domain, and the domain adaptation module includes a first fully connected layer network pair and a second fully connected layer network pair; wherein, determining the feature distribution difference function value using a polynomial kernel function and the training sample set includes: determining a first feature distribution difference function value based on the source domain sample data according to the definition rules of the polynomial kernel function, determining a second feature distribution difference function value based on the output of the first fully connected layer network pair, and determining a third feature distribution difference function value based on the output of the second fully connected layer network pair; and determining the feature distribution difference function value based on the first feature distribution difference function value, the second feature distribution difference function value, and the third feature distribution difference function value.

[0117] In one possible implementation, the target domain data is input into a trained multi-channel domain adaptation network for computation to obtain a diagnostic result for the fault diagnosis task. This includes: extracting sub-domain features from the target domain data to obtain sub-domain specific features representing information unique to the first target domain; and extracting global features from the target domain data to obtain global universal features representing common information shared by the first target domain and each of the source domains. Domain migration features are then determined based on the sub-domain specific features and the global universal features. Fault diagnosis is performed based on the domain migration features to obtain a diagnostic result for the fault diagnosis task.

[0118] In one possible implementation, subdomain feature extraction is performed on the target domain data to obtain subdomain-specific features representing information unique to the target domain, and global feature extraction is performed on the target domain data to obtain global universal features representing common information shared by the target domain and each of the source domains. This includes: performing subdomain feature extraction on the target domain data at a time scale to obtain subdomain temporal features; performing subdomain feature extraction on the subdomain temporal features at a spatial scale to obtain the subdomain-specific features; and performing global feature extraction on the target domain data at a time scale to obtain global temporal features; and performing global feature extraction on the global temporal features at a spatial scale to obtain the global-specific features.

[0119] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0120] This disclosure also proposes a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the above-described method. The computer-readable storage medium can be volatile or non-volatile.

[0121] This disclosure also proposes an electronic device, including: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to implement the above method when executing the instructions stored in the memory.

[0122] This disclosure also provides a computer program product, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is run in a processor of an electronic device, the processor in the electronic device performs the above-described method.

[0123] Figure 13 shows a block diagram of a cross-plant fault diagnosis apparatus for chemical processes based on a multi-channel domain adaptive network according to an embodiment of the present disclosure. For example, apparatus 1900 may be provided as a server or terminal device. Referring to Figure 13, apparatus 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by memory 1932 for storing instructions executable by the processing component 1922, such as application programs. The application programs stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, processing component 1922 is configured to execute instructions to perform the methods described above.

[0124] Device 1900 may also include a power supply component 1926 configured to perform power management of device 1900, a wired or wireless network interface 1950 configured to connect device 1900 to a network, and an input / output interface 1958 (I / O interface). Device 1900 can operate on an operating system, such as Windows Server, stored in memory 1932. TM Mac OS X TM Unix TM Linux TM FreeBSD TM Or similar.

[0125] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of the apparatus 1900 to perform the described method. This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure. A computer-readable storage medium may be a tangible device capable of holding and storing instructions used by an instruction execution device. A computer-readable storage medium may be, for example—but not limited to—an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combinations thereof. Computer-readable storage media as used herein is not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires. The computer-readable program instructions described herein can be downloaded from the computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. Networks may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. Each computing / processing device's network adapter card or network interface receives computer-readable program instructions from the network and forwards these instructions to computer-readable storage media within the respective computing / processing device. The computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server.In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from computer-readable program instructions. This electronic circuitry can execute computer-readable program instructions to implement various aspects of this disclosure.

[0126] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0127] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0128] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0129] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0130] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for cross-unit fault diagnosis in chemical processes based on multi-channel domain adaptive networks, characterized in that, The method includes: Obtain fault diagnosis tasks for the equipment to be diagnosed under the target process flow; Based on the fault diagnosis task and the device to be diagnosed, target domain data required to perform the fault diagnosis task is determined. The target domain data includes various operating data of the device to be diagnosed running under a first target domain, which is determined according to the target process flow. The target domain data is input into the trained multi-channel domain adaptation network for calculation to obtain the diagnostic results for the fault diagnosis task. The method further includes a training process for the multi-channel domain adaptation network, the training process comprising the following steps: Obtain target domain sample datasets and source domain sample datasets. The target domain sample dataset includes target domain sample data of multiple different sample devices operating under multiple different second target domains. The source domain sample dataset includes source domain sample data of each of the sample devices operating under multiple different source domains. Each target domain sample data and each source domain sample data includes multiple normal operation data and multiple abnormal operation data. Each sample device, each second target domain, and each source domain is determined according to different process flows in the chemical process. Each normal operation data and each abnormal operation data in the source domain sample dataset are labeled to obtain the source domain labeled dataset. A training sample set is constructed based on the source domain labeled dataset and the target domain sample dataset. The initialized multi-channel domain adaptation network is trained based on the preset loss function and the training sample set to obtain the trained multi-channel domain adaptation network. The loss function includes a feature distribution difference function constructed based on the maximum mean difference induced by a multinomial kernel.

2. The method according to claim 1, characterized in that, The multi-channel domain adaptation network includes a feature extraction module and a domain adaptation module; The multi-channel domain adaptation network is trained based on a preset loss function and the training sample set to obtain the trained multi-channel domain adaptation network, including: The training sample set is processed using the feature extraction module to obtain the processing result. Based on the processing result, the source domain loss function value and target domain loss function value of the feature extraction module are determined. The processing result includes source domain features, target domain features, source domain classification result, and target domain classification result. The source domain features and the target domain features are processed by the domain adaptation module to obtain the fault category prediction results of each abnormal operation data in the target domain sample data. The relational loss function value of the domain adaptation module is determined based on the fault category prediction results. The feature distribution difference function value is determined using a multinomial kernel function and the training sample set; The total loss function value of the current multi-channel domain adaptation network is determined based on the source domain loss function value, the target domain loss function value, the relation loss function value, and the feature distribution difference function value, and the parameters of the current multi-channel domain adaptation network are updated based on the total loss function value.

3. The method according to claim 2, characterized in that, The feature extraction module includes a source domain branch and a target domain branch; The feature extraction module processes the training sample set to obtain a processing result, and based on the processing result, determines the source domain loss function value and target domain loss function value of the feature extraction module, including: The source domain branch is used to extract features from the data of each source domain in the source domain sample data to obtain the source domain features of each source domain. Based on the source domain features of each source domain, the source domain classification result of the corresponding source domain is inferred. Based on the classification result of each source domain and the true domain label of each source domain, the source domain loss function value of the feature extraction module is determined. The target domain sample data is used to extract features by using the target domain branch to obtain the target domain features of each second target domain. The target domain classification result of the corresponding second target domain is inferred based on the target domain features. The target domain loss function value of the feature extraction module is determined based on the target domain classification result and the real domain label of each second target domain.

4. The method according to claim 2, characterized in that, The domain adaptation module processes the source domain features and the target domain features to obtain fault category prediction results for each abnormal operation data in the target domain sample data. Based on each fault category prediction result, the relational loss function value of the domain adaptation module is determined, including: The domain adaptation module is used to calculate the target domain features and the source domain features to obtain the domain-adapted sample features; Based on the domain adaptation sample feature degree and the predetermined fault category, the fault category prediction result of the target domain sample data is determined; The relational loss function value of the domain adaptation module is determined based on the fault category prediction results and the actual fault labels of the target domain sample data.

5. The method according to claim 2, characterized in that, The source domain features include source domain-specific features representing information unique to the source domain, the target domain features include target domain-specific features representing information unique to the second target domain, and the domain adaptation module includes a first fully connected layer network pair and a second fully connected layer network pair. The feature distribution difference function value is determined using a multinomial kernel function and the training sample set, including: According to the definition rules of the polynomial kernel function, the first feature distribution difference function value is determined based on the source domain sample data, the second feature distribution difference function value is determined based on the output of the first fully connected layer network pair, and the third feature distribution difference function value is determined based on the output of the second fully connected layer network pair. The feature distribution difference function value is determined based on the first feature distribution difference function value, the second feature distribution difference function value, and the third feature distribution difference function value.

6. The method according to any one of claims 1 to 5, characterized in that, The target domain data is input into a trained multi-channel domain adaptation network for computation to obtain diagnostic results for the fault diagnosis task, including: Subdomain feature extraction is performed on the target domain data to obtain subdomain-specific features representing information unique to the first target domain, and global feature extraction is performed on the target domain data to obtain global general features representing common information shared by the first target domain and each of the source domains. Domain migration features are determined based on the subdomain-specific features and the global general features; Based on the domain migration characteristics, fault diagnosis is performed to obtain diagnostic results for the fault diagnosis task.

7. The method according to claim 6, characterized in that, Subdomain feature extraction is performed on the target domain data to obtain subdomain-specific features representing information unique to the target domain. Global feature extraction is then performed on the target domain data to obtain global universal features representing common information shared by the target domain and each of the source domains, including: Subdomain features are extracted at the time scale from the target domain data to obtain subdomain temporal features. Then, subdomain features are extracted at the spatial scale from the subdomain temporal features to obtain specific features of the subdomain. Global features are extracted from the target domain data on a time scale to obtain global temporal features. Global features are then extracted from the global temporal features on a spatial scale to obtain global specific features.

8. A cross-unit fault diagnosis device for chemical processes based on a multi-channel domain adaptive network, characterized in that, include: The acquisition module is used to acquire fault diagnosis tasks for the equipment to be diagnosed under the target process flow. The determination module is used to determine the target domain data required to perform the fault diagnosis task based on the fault diagnosis task and the device to be diagnosed. The target domain data includes various operating data of the device to be diagnosed running under a first target domain, which is determined according to the target process flow. The calculation module is used to input the target domain data into the trained multi-channel domain adaptation network for calculation to obtain the diagnosis result for the fault diagnosis task; The device further includes a training module, which is used for: Obtain target domain sample datasets and source domain sample datasets. The target domain sample dataset includes target domain sample data of multiple different sample devices operating under multiple different second target domains. The source domain sample dataset includes source domain sample data of each of the sample devices operating under multiple different source domains. Each target domain sample data and each source domain sample data includes multiple normal operation data and multiple abnormal operation data. Each sample device, each second target domain, and each source domain is determined according to different process flows in the chemical process. Each normal operation data and each abnormal operation data in the source domain sample dataset are labeled to obtain the source domain labeled dataset. A training sample set is constructed based on the source domain labeled dataset and the target domain sample dataset. The initialized multi-channel domain adaptation network is trained based on the preset loss function and the training sample set to obtain the trained multi-channel domain adaptation network. The loss function includes a feature distribution difference function constructed based on the maximum mean difference induced by a multinomial kernel.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the method of any one of claims 1 to 7 when executing instructions stored in the memory.

10. A non-volatile computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images