Big data processing method and system applied to industrial Internet

Through the model generation method based on graph self-attention network, combined with benchmark samples and migration sample data, the industrial equipment diagnostic model is optimized, which solves the problems of low efficiency and insufficient generalization capabilities of traditional methods, and realizes efficient and accurate industrial equipment Troubleshooting and health maintenance.

CN117991748BActive Publication Date: 2024-11-26BEIJING UNITED MEDIA TECH CO LTD
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202410031305.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-09
Publication Date
2024-11-26
Estimated Expiration
2044-01-09

AI Technical Summary

Technical Problem

Traditional industrial equipment fault diagnosis methods are inefficient, difficult to adapt to complex and changeable industrial environments, and have limited generalization capabilities, making it difficult to expand application fields.

Method used

Using a model generation method based on graph self-attention network, the industrial equipment diagnostic model is optimized through benchmark samples and migration sample data, and a more efficient second industrial equipment diagnostic model is generated, and the guided model is used to learn parameters, so as to improve the Generalization performance and adaptability of the model.

Benefits of technology

It realizes efficient accuracy and rapid migration of industrial equipment fault diagnosis, reduces the cost and time of new equipment launch, expands the application field of industrial equipment diagnostics, and improves the robustness and adaptability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117991748B_ABST
    Figure CN117991748B_ABST
Patent Text Reader

Abstract

The embodiment of the present application provides a big data processing method and system for industrial Internet. By utilizing a model generation method based on a graph self-attention network, the initial first industrial equipment diagnosis model is optimized and a more efficient second industrial equipment diagnosis model is generated. By using benchmark sample and migration sample data for model training, the generalization performance of the model in various industrial environments can be improved by using transfer learning methods while protecting the model from the risk of overfitting. In addition, in the scenario of processing large-scale industrial Internet big data, the application field of industrial equipment diagnosis is greatly expanded, which helps to achieve efficient and accurate fault diagnosis and health maintenance of industrial equipment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of industrial Internet technology, and specifically, to a big data processing method and system applied to industrial Internet. Background Art

[0002] In modern industrial production, industrial equipment fault diagnosis is a crucial link, which is of great significance for ensuring production safety and improving production efficiency. With the development of the industrial Internet, the scale and complexity of industrial Internet big data are increasing, which brings a series of challenges to industrial equipment fault diagnosis. Traditional industrial equipment diagnosis methods mainly rely on manual inspection or rule-based expert systems. These methods are not only inefficient, but also often cannot make accurate judgments when facing complex and changing industrial environments.

[0003] In addition, due to the lack of an effective learning mechanism, traditional diagnostic methods are difficult to adapt to new equipment and working conditions. For another example, other existing methods mainly include model-based methods and data-driven methods. Model-based methods rely on accurate modeling of the internal mechanisms of industrial equipment, which is often difficult to achieve in practical applications; while data-driven methods use a large amount of historical data to train models for fault diagnosis, but such methods usually need to be retrained and optimized when facing different industrial scenarios, and their generalization ability is limited, making it difficult to expand the application field of industrial equipment diagnosis. Summary of the invention

[0004] In order to at least overcome the above-mentioned deficiencies in the prior art, the purpose of the embodiments of the present application is to provide a big data processing method and system for industrial Internet.

[0005] Acquire a first industrial equipment diagnosis model, wherein the first industrial equipment diagnosis model is a neural network model generated by learning network parameter information in a first industrial equipment diagnosis process using benchmark sample industrial control data, and the first industrial equipment diagnosis model includes a first graph self-attention network;

[0006] Performing feature conversion on the migration sample industrial control data to generate migration sample conversion data, and performing graph self-attention feature analysis on the migration sample conversion data according to the first graph self-attention network to generate graph self-attention features;

[0007] Performing parameter learning on the first image self-attention network according to the image self-attention features and the migration sample industrial control data to generate a second image self-attention network;

[0008] Optimizing the first industrial equipment diagnosis model according to the second graph self-attention network to generate a guiding model and a guided model;

[0009] Based on the fuzzy diagnosis result data generated by the guidance model using the benchmark sample industrial control data and the migration sample industrial control data, parameter learning is performed on the guided model to generate a second industrial equipment diagnosis model. The second industrial equipment diagnosis model is used to perform industrial equipment diagnosis on the benchmark industrial control data or the migration industrial equipment data in the target industrial Internet big data during the first industrial equipment diagnosis process.

[0010] In a possible implementation of the first aspect, performing feature conversion on the migration sample industrial control data to generate migration sample conversion data, and performing graph self-attention feature parsing on the migration sample conversion data according to the first graph self-attention network includes:

[0011] Performing data obfuscation processing on the migration sample industrial control data to generate the migration sample industrial control data after the data obfuscation processing as the migration sample conversion data;

[0012] Loading the migrated sample industrial control data after the data obfuscation processing into the first graph self-attention network to perform graph self-attention feature analysis to generate obfuscated graph self-attention features;

[0013] The step of performing parameter learning on the first graph self-attention network according to the graph self-attention feature and the migration sample industrial control data to generate a second graph self-attention network includes:

[0014] Loading the self-attention features of the confusion graph into a confusion restoration network for feature restoration processing to generate confusion restoration data, wherein the confusion restoration data reflects the data result generated by reconstructing the migration sample industrial control data after the data obfuscation processing;

[0015] Parameters of the first graph self-attention network are learned according to feature differences between the obfuscated restored data and the migrated sample industrial control data to generate the second graph self-attention network.

[0016] In a possible implementation of the first aspect, the migration sample industrial control data includes migration sample device session data, and the migration sample device session data includes k state session events, where k is an integer greater than 1;

[0017] The performing data obfuscation processing on the migration sample industrial control data to generate the migration sample industrial control data after the data obfuscation processing as the migration sample conversion data includes:

[0018] Performing session vector analysis on the migration sample device session data to generate session vector path data corresponding to the migration sample device session data, wherein the session vector path data includes session link vectors corresponding to the k state session events respectively;

[0019] Performing data obfuscation processing on the session vector path data to generate an obfuscated processing vector as the migration sample industrial control data after the data obfuscation processing, wherein the obfuscated processing vector includes x hidden session link vectors, where x<k and x is a positive integer;

[0020] The step of loading the self-attention features of the confusion map into a confusion restoration network for feature restoration processing to generate confusion restoration data includes:

[0021] Loading the confusion graph self-attention features into the confusion restoration network, performing industrial equipment diagnosis on the state session events corresponding to the x hidden session link vectors, and generating x derived state session events;

[0022] The reconstructed device session data corresponding to the migration sample device session data is obtained according to the x derived state session events, and the reconstructed device session data is used as the obfuscated restoration data.

[0023] In a possible implementation of the first aspect, performing feature conversion on the migration sample industrial control data to generate migration sample conversion data, and performing graph self-attention feature parsing on the migration sample conversion data according to the first graph self-attention network includes:

[0024] Performing feature expansion on the migration sample industrial control data to generate expanded migration sample industrial control data as the migration sample conversion data, wherein the expanded migration sample industrial control data includes valid training data and invalid training data corresponding to the migration sample industrial control data;

[0025] Loading the migration sample industrial control data, the valid training data and the invalid training data into the first graph self-attention network for graph self-attention feature analysis, generating migration graph self-attention features corresponding to the migration sample industrial control data, valid graph self-attention features corresponding to the valid training data and invalid graph self-attention features corresponding to the invalid training data;

[0026] The performing parameter learning on the first image self-attention network according to the image self-attention feature and the migration sample industrial control data to generate a second image self-attention network includes:

[0027] The first graph self-attention network is trained in a direction of minimizing the feature difference between the transition graph self-attention feature and the valid graph self-attention feature and maximizing the feature difference between the transition graph self-attention feature and the invalid graph self-attention feature, and parameter learning is performed on the first graph self-attention network to generate the second graph self-attention network.

[0028] In a possible implementation of the first aspect, the fuzzy diagnosis result data generated by the guiding model based on the benchmark sample industrial control data and the migration sample industrial control data, performing parameter learning on the guided model to generate the second industrial equipment diagnosis model includes:

[0029] In the polling optimization of the qth training stage, the guided model optimized and generated in the q-1th training stage is optimized according to the fuzzy diagnosis result data generated by the guidance model optimized and generated in the q-1th training stage through the reference sample industrial control data and the migration sample industrial control data, so as to generate the guided model optimized and generated in the qth training stage and the guidance model optimized and generated in the qth training stage;

[0030] Initializing the network parameter information of the graph self-attention network in the guided model generated by optimization in the qth training stage as the network parameter information of the second graph self-attention network, to generate the guided model initialized in the qth training stage;

[0031] According to the fuzzy diagnosis result data generated by the guidance model optimized and generated according to the benchmark sample industrial control data and the migration sample industrial control data in the qth training stage, parameter learning is performed on the guided model initialized in the qth training stage to generate the second industrial equipment diagnosis model.

[0032] In a possible implementation manner of the first aspect, generating the guided model generated by optimization in the qth training stage and the guiding model generated by optimization in the qth training stage includes:

[0033] After obtaining the guided model generated by optimization in the qth training stage, obtaining first network parameter information of the guided model generated by optimization in the qth training stage and second network parameter information of the guiding model generated by optimization in the q-1th training stage;

[0034] Performing parameter integration on the first network parameter information and the second network parameter information based on a set optimization rule to generate integrated network parameter information;

[0035] The guidance model generated by optimization in the q-1th training stage is optimized according to the integrated network parameter information to generate the guidance model generated by optimization in the qth training stage.

[0036] In a possible implementation of the first aspect, the fuzzy diagnosis result data generated by the guiding model based on the benchmark sample industrial control data and the migration sample industrial control data, performing parameter learning on the guided model to generate the second industrial equipment diagnosis model includes:

[0037] Strengthening and extending the benchmark sample industrial control data and the migration sample industrial control data respectively to generate strengthened and extended benchmark sample data and strengthened and extended migration sample data;

[0038] Performing industrial equipment diagnosis on the enhanced extended benchmark sample data and the enhanced extended migration sample data respectively according to the guided model to generate benchmark industrial equipment diagnostic data and migration industrial equipment diagnostic data, wherein the migration industrial equipment diagnostic data reflects the industrial equipment diagnostic data of the migration sample industrial control data in the first industrial equipment diagnostic process, and the benchmark industrial equipment diagnostic data reflects the industrial equipment diagnostic data of the benchmark sample industrial control data in the first industrial equipment diagnostic process;

[0039] Acquire benchmark industrial equipment diagnostic annotation data, wherein the benchmark industrial equipment diagnostic annotation data reflects reference diagnostic data of the benchmark sample industrial control data in the first industrial equipment diagnostic process;

[0040] Performing lightweight extension on the migration sample industrial control data to generate lightweight extended migration sample data;

[0041] Performing fuzzy diagnosis prediction on the lightweight extension migration sample data according to the guidance model to generate fuzzy diagnosis result data; wherein the feature extension depth of the enhanced extension to the data is greater than the feature extension depth of the lightweight extension to the data;

[0042] Determining a benchmark training error parameter based on feature differences between the benchmark industrial equipment diagnostic data and the benchmark industrial equipment diagnostic annotated data;

[0043] Determining a migration training error parameter according to a feature difference between the migration industrial equipment diagnosis data and the fuzzy diagnosis result data;

[0044] Parameter learning is performed on the guided model according to the benchmark training error parameter and the transfer training error parameter to generate the second industrial equipment diagnostic model.

[0045] In a possible implementation of the first aspect, the method further includes:

[0046] When industrial equipment diagnosis is performed on the enhanced extended benchmark sample data and the enhanced extended migration sample data according to the guided model, domain feature consistency analysis is performed on the enhanced extended benchmark sample data and the enhanced extended migration sample data to generate a domain feature consistency error;

[0047] The step of performing parameter learning on the guided model according to the reference training error parameter and the migration training error parameter to generate the second industrial equipment diagnosis model includes:

[0048] The reference training error parameter, the migration training error parameter and the domain feature consistency error are integrated to generate a global error parameter;

[0049] performing parameter learning on the guided model according to the global error parameter to generate the second industrial equipment diagnostic model;

[0050] Wherein, the first industrial equipment diagnosis process includes a fault diagnosis task;

[0051] The step of performing fuzzy diagnosis prediction on the lightweight extended migration sample data according to the guidance model to generate fuzzy diagnosis result data includes:

[0052] Performing fuzzy diagnosis prediction on the lightweight extended migration sample data according to the guidance model, generating a plurality of reference fuzzy diagnosis result data and diagnosis probability values ​​respectively corresponding to the plurality of reference fuzzy diagnosis result data, wherein the plurality of reference fuzzy diagnosis result data reflect the fault diagnosis category generated after the fault diagnosis is performed on the lightweight extended migration sample data, and the diagnosis probability value reflects the probability value of the fault diagnosis category;

[0053] The reference fuzzy diagnosis result data having a diagnosis probability value greater than a set probability value among the plurality of reference fuzzy diagnosis result data are output as the fuzzy diagnosis result data.

[0054] In a possible implementation of the first aspect, the method further includes:

[0055] Performing fuzzy diagnosis prediction on the benchmark sample industrial control data according to the guidance model, generating a plurality of benchmark industrial equipment diagnosis annotation data and diagnosis probability values ​​corresponding to the plurality of benchmark industrial equipment diagnosis annotation data, wherein the plurality of diagnosis probability values ​​correspond to w industrial equipment fault labels, and w is a positive integer;

[0056] Calculate the average probability value of multiple diagnosis probability values ​​corresponding to the r-th industrial equipment fault label as the threshold probability value corresponding to the r-th industrial equipment fault label, r≤w and r is a positive integer;

[0057] The step of outputting the reference fuzzy diagnosis result data having a diagnosis probability value greater than a set probability value among the plurality of reference fuzzy diagnosis result data as the fuzzy diagnosis result data comprises:

[0058] If the reference fuzzy diagnosis result data corresponding to the r-th industrial equipment fault label among the multiple reference fuzzy diagnosis result data is greater than the threshold probability value corresponding to the r-th industrial equipment fault label, the reference fuzzy diagnosis result data corresponding to the r-th industrial equipment fault label is output as the fuzzy diagnosis result data.

[0059] According to one aspect of an embodiment of the present application, an industrial Internet system is provided, which includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement a big data processing method applied to the industrial Internet in any one of the aforementioned possible implementations.

[0060] According to one aspect of the embodiments of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in various optional implementations of the above three aspects.

[0061] In the technical solutions provided in some embodiments of the present application, the initial first industrial equipment diagnosis model is optimized and a more efficient second industrial equipment diagnosis model is generated by utilizing a model generation method based on a graph self-attention network. By using benchmark sample and migration sample data for model training, the generalization performance of the model in various industrial environments can be improved by using the transfer learning method while protecting the model from the risk of overfitting. In addition, in the scenario of processing large-scale industrial Internet big data, the application field of industrial equipment diagnosis is greatly expanded, which helps to achieve efficient and accurate fault diagnosis and health maintenance of industrial equipment.

[0062] By adopting a neural network model based on the self-attention mechanism, the characteristics of industrial control data can be parsed and learned more effectively. The self-attention network can automatically discover important features and hidden relationships when processing complex industrial control data, thereby improving the accuracy and robustness of the model. By performing feature conversion and graph self-attention feature parsing on the migrated sample industrial control data, a second graph self-attention network is generated, and the first industrial equipment diagnosis model is optimized accordingly, thereby realizing the rapid migration and optimization of the industrial equipment diagnosis model. This not only reduces the cost and time of new equipment going online, but also enables the model to have better generalization ability and adapt to industrial equipment of different types or states. By introducing the guidance model and the guided model, as well as the fuzzy diagnosis result data generated by the benchmark sample industrial control data and the migrated sample industrial control data, the learning process of the industrial equipment diagnosis model is further optimized, and the existing knowledge and newly collected data are fully utilized, making the model training more efficient and the diagnosis results more accurate. Finally, the generated second industrial equipment diagnosis model can be widely used in the first industrial equipment diagnosis process, and the benchmark industrial control data or migrated industrial equipment data in the target industrial Internet big data are used for industrial equipment diagnosis, which greatly expands the application field of industrial equipment diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be extracted in combination with these drawings without creative work.

[0064] Figure 1 A flowchart of a big data processing method applied to industrial Internet provided in an embodiment of the present application;

[0065] Figure 2 A schematic block diagram of the structure of an industrial Internet system for implementing the above-mentioned big data processing method applied to industrial Internet provided in an embodiment of the present application. DETAILED DESCRIPTION

[0066] The following description is intended to enable one of ordinary skill in the art to implement and combine the present application, and the description is provided in the context of a specific application scenario and its requirements. It will be apparent to one of ordinary skill in the art that various changes may be made to the disclosed embodiments, and that the general principles defined in the present application may be applied to other embodiments and application scenarios without departing from the principles and scope of the present application. Therefore, the present application is not limited to the described embodiments, but should be given the broadest scope consistent with the claims.

[0067] Figure 1 This is a flow chart of a big data processing method applied to industrial Internet provided by an embodiment of the present application. The big data processing method applied to industrial Internet is introduced in detail below.

[0068] Step S110, obtaining a first industrial equipment diagnosis model, wherein the first industrial equipment diagnosis model is a neural network model generated by learning network parameter information in a first industrial equipment diagnosis process using benchmark sample industrial control data, and the first industrial equipment diagnosis model includes a first graph self-attention network.

[0069] For example, in the field of new energy, such as wind power generation systems, the industrial Internet system can collect and analyze the historical operation data of wind power generation equipment as benchmark sample industrial control data. After obtaining these benchmark sample industrial control data, deep learning techniques, such as neural network models, are used to analyze and learn through a large amount of benchmark sample industrial control data to generate a first industrial equipment diagnosis model that can diagnose wind power generation equipment. The first industrial equipment diagnosis model contains a first graph self-attention network to capture the correlation between features.

[0070] The reference sample industrial control data may refer to industrial control data collected during normal or fault mode operation of a specific industrial equipment (such as a wind turbine). These reference sample industrial control data may include operating parameters, sensor readings, status monitoring data, etc. of the industrial equipment, which are used to establish and train the diagnostic model of the industrial equipment. For example, in a wind turbine, the reference sample industrial control data may include parameters such as wind speed, turbine speed, and generator output power.

[0071] The first industrial equipment diagnosis process is a process of training a neural network model using benchmark sample industrial control data. In this process, the neural network model will gradually master the normal operation and performance of industrial equipment when failures occur by learning various industrial equipment characteristics and patterns. For example, in a wind turbine, the first industrial equipment diagnosis process may include analyzing parameters such as wind speed and turbine speed to determine whether the wind turbine is operating normally.

[0072] Network parameter information learning can refer to deep learning analysis of benchmark sample industrial control data, where the neural network model learns the characteristics and associations of various industrial equipment and uses this information for pattern recognition and fault prediction. For example, in a wind turbine, network parameter information learning may include learning the relationship between wind speed and turbine speed to predict the performance of the turbine at different wind speeds.

[0073] The neural network model is an analytical model that simulates the computing mode of the human brain, and is composed of a large number of neurons (nodes) and the connections (edges) between them. The neural network model can be used to deal with complex problems, including pattern recognition, prediction, and natural language processing. In the first industrial equipment diagnosis model, the neural network model is used to identify and predict the failure of industrial equipment. For example, in the diagnosis of wind turbine faults, deep learning algorithms such as convolutional neural networks (CNN) or recurrent neural networks (RNN) may be used to establish a neural network model.

[0074] The first graph self-attention network is an important component of the neural network model, which uses the self-attention mechanism in graph theory to process and parse the feature relationships in the input data. Through the self-attention mechanism, the graph self-attention network can better capture the key information in the data and give it higher weight, thereby improving the diagnostic accuracy of the neural network model. For example, in the diagnosis of wind turbine faults, the first graph self-attention network can use the self-attention mechanism to analyze the relationship between sensor reading data and find out the key parameters related to the fault.

[0075] Step S120, performing feature conversion on the migration sample industrial control data to generate migration sample conversion data, and performing graph self-attention feature analysis on the migration sample conversion data according to the first graph self-attention network to generate graph self-attention features.

[0076] For example, in order to expand the application scope of the first graph self-attention network, the industrial Internet system can perform feature conversion on the operating data of other types of wind power equipment (migrated sample industrial control data) to generate migrated sample conversion data. Then, the graph self-attention features are parsed based on the first graph self-attention network to extract the graph self-attention features. The purpose of this is to improve the universality of the model by incorporating the operating data of different wind power equipment into model training.

[0077] Exemplarily, the migrated sample industrial control data may refer to data collected from other similar industrial equipment (such as other types of generators) in addition to the training data of a specific industrial equipment (such as a wind turbine). The migrated sample industrial control data can help the neural network model to better generalize to other similar industrial equipment and improve the adaptability and accuracy of the neural network model. For example, in wind turbine fault diagnosis, the migrated sample industrial control data may include data collected from other types of generators, such as hydroelectric generators or solar generators.

[0078] In the graph self-attention network, the process of converting the transfer sample conversion data into graph self-attention features. In this process, the attention weight of each node (data point) can be calculated based on the feature relationship of the input data, and these weights are used to generate the session vector. The self-attention feature is a generalized representation that captures the key information in the input data. For example, in wind turbine fault diagnosis, graph self-attention feature parsing can generate self-attention features that summarize the operating status of the wind turbine based on the relationship between sensor reading data.

[0079] Step S130, performing parameter learning on the first graph self-attention network according to the graph self-attention features and the migration sample industrial control data to generate a second graph self-attention network.

[0080] For example, the industrial Internet system then learns the parameters of the first graph self-attention network based on the graph self-attention features and the transferred sample industrial control data to generate the second graph self-attention network. In this process, the similarities and differences between different wind power equipment can be learned, making it more powerful and accurate in fault diagnosis.

[0081] Step S140: optimizing the first industrial equipment diagnosis model according to the second graph self-attention network to generate a guiding model and a guided model.

[0082] For example, after the second graph self-attention network is generated, the industrial Internet system will optimize the first industrial equipment diagnosis model based on this new second graph self-attention network to generate a guidance model and a guided model. Among them, the guidance model is an auxiliary model (teacher network) used to guide another student network (guided model); the guided model accepts the guidance of the guidance model and optimizes itself.

[0083] Step S150, based on the fuzzy diagnosis result data generated by the guidance model using the benchmark sample industrial control data and the migration sample industrial control data, the guided model is subjected to parameter learning to generate a second industrial equipment diagnosis model. The second industrial equipment diagnosis model is used to perform industrial equipment diagnosis on the benchmark industrial control data or the migration industrial equipment data in the target industrial Internet big data during the first industrial equipment diagnosis process.

[0084] For example, the industrial Internet system will use the benchmark sample industrial control data and the migrated sample industrial control data to generate fuzzy diagnosis result data through the guidance model, and then learn the parameters of the guided model based on these fuzzy diagnosis result data to generate a more accurate second industrial equipment diagnosis model. Finally, this second industrial equipment diagnosis model will be applied in the real-time monitoring and maintenance of the wind power generation system to perform fault diagnosis and equipment optimization of wind power generation equipment on the benchmark industrial control data or migrated industrial equipment data in the target industrial Internet big data.

[0085] Based on the above steps, the embodiment of the present application optimizes the initial first industrial equipment diagnosis model and generates a more efficient second industrial equipment diagnosis model by utilizing a model generation method based on a graph self-attention network. Using benchmark sample and migration sample data for model training can protect the model from the risk of overfitting while using the transfer learning method to improve the generalization performance of the model in various industrial environments. In addition, in the scenario of processing large-scale industrial Internet big data, the application field of industrial equipment diagnosis is greatly expanded, which helps to achieve efficient and accurate fault diagnosis and health maintenance of industrial equipment.

[0086] By adopting a neural network model based on the self-attention mechanism, the characteristics of industrial control data can be parsed and learned more effectively. The self-attention network can automatically discover important features and hidden relationships when processing complex industrial control data, thereby improving the accuracy and robustness of the model. By performing feature conversion and graph self-attention feature parsing on the migrated sample industrial control data, a second graph self-attention network is generated, and the first industrial equipment diagnosis model is optimized accordingly, thereby realizing the rapid migration and optimization of the industrial equipment diagnosis model. This not only reduces the cost and time of new equipment going online, but also enables the model to have better generalization ability and adapt to industrial equipment of different types or states. By introducing the guidance model and the guided model, as well as the fuzzy diagnosis result data generated by the benchmark sample industrial control data and the migrated sample industrial control data, the learning process of the industrial equipment diagnosis model is further optimized, and the existing knowledge and newly collected data are fully utilized, making the model training more efficient and the diagnosis results more accurate. Finally, the generated second industrial equipment diagnosis model can be widely used in the first industrial equipment diagnosis process, and the benchmark industrial control data or migrated industrial equipment data in the target industrial Internet big data are used for industrial equipment diagnosis, which greatly expands the application field of industrial equipment diagnosis.

[0087] In a possible implementation, step S120 may include:

[0088] Step S121 , performing data obfuscation processing on the migration sample industrial control data, and generating the migration sample industrial control data after the data obfuscation processing as the migration sample conversion data.

[0089] Step S122: Load the migrated sample industrial control data after the data obfuscation processing into the first graph self-attention network to perform graph self-attention feature analysis to generate obfuscated graph self-attention features.

[0090] For example, data obfuscation processing may refer to certain modifications or transformations of the migration sample industrial control data to increase the generalization ability of the neural network model. Data obfuscation processing may include operations such as random noise injection, data clipping, and data masking. For example, in wind turbine fault diagnosis, data obfuscation processing may include random noise injection into sensor reading data of a hydroelectric generator or a solar generator to simulate measurement errors. Thus, the migration sample industrial control data after data obfuscation processing is used as a new data source for neural network model training, which can improve the generalization ability of the neural network model. For example, in wind turbine fault diagnosis, the migration sample industrial control data after data obfuscation processing may include sensor reading data of a hydroelectric generator or a solar generator injected with random noise. Thus, the migration sample conversion data refers to the process of converting the migration sample industrial control data into a format suitable for neural network model input. That is, the migration sample conversion data is generated by performing data obfuscation processing on the migration sample industrial control data.

[0091] The obfuscated graph self-attention features refer to the results of loading the migrated sample industrial control data after data obfuscation into the first graph self-attention network for graph self-attention feature parsing. These features can capture key information and patterns in the obfuscated data. For example, in wind turbine fault diagnosis, the obfuscated graph self-attention features may include feature vectors that characterize the relationship between sensor readings of hydroelectric generators or solar generators injected with random noise.

[0092] In this embodiment, in order to improve the generalization ability of the neural network model when facing different types of wind turbines, it is first necessary to perform data obfuscation processing on the migration sample industrial control data. Typical data obfuscation processing techniques include data enhancement, data clipping, and noise injection. Taking data enhancement as an example, a certain proportion of parameters such as wind speed or turbine speed can be randomly increased to simulate changes in observed values ​​in actual scenarios. Similarly, certain parameter values ​​can be cut to simulate changes in the sensor reading range. For noise injection, a certain proportion of random noise can be added to the sensor readings or state parameters. Such data obfuscation processing can help the model cope with different data distributions and challenges during the learning process.

[0093] After the data obfuscation process is completed, the obfuscated data needs to be used as migration sample conversion data in the subsequent neural network model training and analysis process. For example, the wind turbine parameters that have been enhanced or cropped can be combined with the sensor readings with added noise to construct a new dataset that can represent the common characteristics and differences between different types of wind turbines.

[0094] Step S130 may include:

[0095] Step S131, loading the self-attention features of the confusion graph into the confusion restoration network for feature restoration processing to generate confusion restoration data, wherein the confusion restoration data reflects the data result generated by reconstructing the migration sample industrial control data after the data obfuscation processing.

[0096] Step S132: performing parameter learning on the first graph self-attention network according to feature differences between the obfuscated restored data and the migrated sample industrial control data to generate the second graph self-attention network.

[0097] The confusion reduction network is a network used to reduce the self-attention features of the confusion map to the original features. The goal of this confusion reduction network is to recover the useful information in the self-attention features of the confusion map by learning the inverse process of the confusion processing process. For example, in wind turbine fault diagnosis, the confusion reduction network may include a denoising neural network to remove noise from the sensor reading data of the hydroelectric generator or solar generator caused by random noise injection.

[0098] The obfuscated restored data refers to the data result generated by loading the obfuscated graph self-attention features into the obfuscated restoration network for feature restoration processing, which reflects the approximate situation of reconstructing the migration sample industrial control data after data obfuscation processing. For example, in wind turbine fault diagnosis, the obfuscated restored data may include denoised sensor reading data of hydroelectric generators or solar generators, which can be used to train models to identify and predict faults.

[0099] On this basis, we can learn how to improve the fault diagnosis performance of the migrated sample industrial control data based on the obfuscated restored data. For example, in wind turbine fault diagnosis, the feature differences between the obfuscated restored data and the migrated sample industrial control data (original unobfuscated data) can be used to train the first graph self-attention network, thereby generating a second graph self-attention network that can better identify and predict wind turbine faults.

[0100] After completing the data obfuscation process, the new migration sample industrial control data needs to be input into the first graph self-attention network for training. The graph self-attention network can capture complex dependencies and patterns in the data. In the wind power generation system, the first graph self-attention network can learn the mutual influence between sensor readings, equipment status, and operating parameters. After training, the first graph self-attention network will generate corresponding obfuscated graph self-attention features for each data point. These obfuscated graph self-attention features can capture the key information in the migration sample industrial control data and help improve the generalization performance of the neural network model between different types of wind turbines.

[0101] In a possible implementation manner, the migration sample industrial control data includes migration sample device session data, and the migration sample device session data includes k state session events, where k is an integer greater than 1.

[0102] Step S121 may include:

[0103] Step S1211 , performing session vector analysis on the migration sample device session data to generate session vector path data corresponding to the migration sample device session data, wherein the session vector path data includes session link vectors corresponding to the k state session events respectively.

[0104] For example, in the scenario of a wind power generation system, it is necessary to process session data from different wind power generation devices. In this embodiment, each migration sample device session data represents the operating state of a wind power generation device within a specific period of time. This operating state includes k state session events, such as device startup, normal operation, shutdown, failure, etc. It should be noted that the number of state session events k is an integer greater than 1.

[0105] In order to use the neural network model to analyze these migration sample device session data, it is necessary to first convert the migration sample device session data into session vector path data. Specifically, the migration sample device session data can be processed using a codec structure that converts it into a session vector path data consisting of k session link vectors. Each session link vector represents the characteristics of a specific state session event. For example, in a wind power generation system, parameters such as wind speed, turbine speed, and generator output power in the session data can be represented by vectors, and then these vectors can be assembled into a session vector path data along the time axis. In this way, the characteristics of each state session event can be captured and useful information can be extracted from it.

[0106] Step S1212, performing data obfuscation processing on the session vector path data, generating an obfuscated processing vector as the migration sample industrial control data after the data obfuscation processing, wherein the obfuscated processing vector includes x hidden session link vectors, where x<k and x is a positive integer.

[0107] For example, in order to make the neural network model more generalizable, it is necessary to perform data obfuscation processing on the session vector path data. Specifically, a part of the session link vectors in the session vector path data can be selected to be hidden, that is, their characteristic values ​​are randomly reset to zero or set to the average value. The final, obfuscated session vector path data is called an obfuscated processing vector. In this case, the obfuscated processing vector contains kx visible session link vectors and x hidden session link vectors. For example, in a wind power generation system, parameters such as turbine speed and wind speed at certain time points in the session vector path data can be selected for obfuscation processing. For example, the values ​​of these parameters can be replaced with their average values ​​or random values ​​within a certain time window to generate data noise. Through this data obfuscation processing, the model can have better generalization ability when facing different data distributions.

[0108] Step S131 may include:

[0109] Step S1311, loading the confusion graph self-attention features into the confusion restoration network, performing industrial equipment diagnosis on the state session events corresponding to the x hidden session link vectors, and generating x derived state session events.

[0110] For example, in the feature learning and processing of the sample industrial control data, the confusion restoration network is needed to restore the information lost due to data obfuscation. Specifically, the self-attention features of the obfuscated graph are input into the confusion restoration network, which will learn how to reconstruct the original session vector path data.

[0111] In the confusion restoration network, the learned knowledge and patterns are used to estimate or predict the session link vectors lost in the self-attention features of the confusion graph. In this process, the state session events corresponding to the hidden session link vectors are mainly diagnosed to retrieve these key feature information. According to the output of the confusion restoration network, x derived state session events can be obtained. These derived state session events can fill in the information lost due to data obfuscation processing and help to understand the operation of the equipment more accurately. For example, in the application scenario of wind power generation system, the confusion restoration network may adopt a similar graph neural network structure, which performs equipment diagnosis by learning the complex correlation between wind speed, turbine speed, etc. and equipment failure. When the self-attention features of the confusion graph are input, the confusion restoration network will try to restore the hidden key parameter values ​​such as wind speed and turbine speed, thereby generating x derived state session events.

[0112] Step S1312: obtaining reconstructed device session data corresponding to the migration sample device session data according to the x derived state session events, and using the reconstructed device session data as obfuscated restoration data.

[0113] After obtaining x derived state session events, it is necessary to integrate these derived results together to reconstruct the original migration sample device session data. Specifically, these x derived state session events will be used to fill in the hidden part of the obfuscation processing vector, thereby generating a complete and reconstructed device session data. This reconstructed device session data can be considered as a kind of "clean" data after obfuscation restoration, which reflects the original characteristics and distribution of the migration sample industrial control data before obfuscation processing. For example, in the application scenario of a wind power generation system, the migration sample device session data can be reconstructed based on x derived state session events. For example, the derived parameter values ​​such as wind speed and turbine speed can be filled into the corresponding time points in the original session data. In this way, a reconstructed device session data can be obtained, which retains the characteristics and details of the original migration sample industrial control data as much as possible.

[0114] In a possible implementation, step S120 may further include:

[0115] Step S123, performing feature expansion on the migration sample industrial control data to generate expanded migration sample industrial control data as the migration sample conversion data, wherein the expanded migration sample industrial control data includes valid training data and invalid training data corresponding to the migration sample industrial control data.

[0116] Step S124, loading the migration sample industrial control data, the valid training data and the invalid training data into the first graph self-attention network for graph self-attention feature analysis, generating migration graph self-attention features corresponding to the migration sample industrial control data, valid graph self-attention features corresponding to the valid training data and invalid graph self-attention features corresponding to the invalid training data.

[0117] For example, in a scenario of a wind power generation system, it is necessary to perform feature expansion on the migrated sample industrial control data. This can be done by combining the migrated sample industrial control data with its corresponding invalid training data and valid training data. Valid training data generally includes features that are relevant to the target task and helpful for learning the neural network model, while invalid training data may include features that are irrelevant to the target task or not conducive to learning the neural network model. For example, in this embodiment, it may be desirable to establish a neural network model to predict the failure of wind turbine equipment. Then the valid training data may include the operating parameters, fault records, etc. of the equipment, while the invalid training data may include irrelevant parameters or noise data of the equipment under normal operation. By combining these data, a richer and more challenging training data set can be obtained.

[0118] Therefore, in this process, the first graph self-attention network will be used to perform feature analysis on the expanded migration sample industrial control data. Specifically, each feature in the expanded migration sample industrial control data is regarded as a node in the graph, and the association between the features is regarded as an edge, so as to construct a graph structure. Then the graph self-attention network is applied to capture the dependencies and interactions between these nodes and edges. In this way, corresponding graph self-attention features can be generated for the migration sample industrial control data, valid training data, and invalid training data respectively. These features will help the subsequent training process, enabling the neural network model to learn deeper knowledge and patterns from complex industrial control data.

[0119] Step S130 may also include: performing parameter learning on the first graph self-attention network with the training direction of minimizing the feature difference between the transition graph self-attention feature and the valid graph self-attention feature and maximizing the feature difference between the transition graph self-attention feature and the invalid graph self-attention feature to generate the second graph self-attention network.

[0120] For example, during the training process, the goal is to enable the model to better focus on features related to the target task and ignore features that may be irrelevant or interfering with the task when processing the transfer sample industrial control data. To achieve this goal, the training direction is to minimize the feature difference between the transfer graph self-attention features and the effective graph self-attention features (thereby retaining the advantages of the effective features) while maximizing the feature difference between the transfer graph self-attention features and the invalid graph self-attention features (so that the model learns to maintain stability or robustness in the face of invalid features). In this way, the parameters of the model can be optimized for complex industrial scenarios, giving it better performance and application potential.

[0121] For example, in the scenario of wind power generation systems, the model can be made more resistant to interference when processing variable data in actual industrial scenarios, thereby improving the accuracy of predicting wind turbine failures by minimizing the differences in parameters such as wind speed and turbine speed between the transition graph self-attention features and the effective graph self-attention features (because these parameters are critical to predicting wind turbine failures), while maximizing the differences in features such as background noise and irrelevant equipment parameters between the transition graph self-attention features and the invalid graph self-attention features.

[0122] In a possible implementation, in step S150, based on the benchmark sample industrial control data and the fuzzy diagnosis result data generated by the guidance model through the migration sample industrial control data, parameter learning is performed on the guided model to generate a second industrial equipment diagnosis model, including:

[0123] Step S151, in the polling optimization of the qth training stage, the guided model optimized and generated in the q-1th training stage is optimized based on the fuzzy diagnosis result data generated by the guidance model optimized in the q-1th training stage through the benchmark sample industrial control data and the migration sample industrial control data, so as to generate the guided model optimized and generated in the qth training stage and the guidance model optimized and generated in the qth training stage.

[0124] For example, in a scenario of a wind power generation system, it is assumed that the training of the q-1 training stage has been completed, and the guided model and the guiding model optimized and generated in the q-1 training stage have been obtained. Now we are going to enter the training of the q-1 training stage. In this stage, the benchmark sample industrial control data and the migration sample industrial control data are input into the guiding model optimized and generated in the q-1 training stage to obtain the fuzzy diagnosis result data. Then these fuzzy diagnosis result data are used to optimize the guided model to generate the guided model optimized and generated in the q-1 training stage. In this way, at the end of each training stage, the guided model and the guiding model will be updated and optimized, thereby gradually improving the model performance.

[0125] Step S152, initializing the network parameter information of the graph self-attention network in the guided model generated by optimization in the qth training stage to the network parameter information of the second graph self-attention network, and generating the guided model initialized in the qth training stage.

[0126] In this process, the network parameter information of the graph self-attention network in the guided model generated by the optimization in the qth training stage will be used to initialize the second graph self-attention network. The advantage of this is that the useful information obtained from the previous training is retained and used to initialize the second graph self-attention network, thereby accelerating the convergence of the model and improving the accuracy of the model in industrial equipment diagnosis tasks.

[0127] Step S153, based on the fuzzy diagnosis result data generated by the guidance model optimized and generated according to the benchmark sample industrial control data and the migration sample industrial control data in the qth training stage, parameter learning is performed on the guided model initialized in the qth training stage to generate the second industrial equipment diagnosis model.

[0128] For example, the benchmark sample industrial control data and the transfer sample industrial control data and the fuzzy diagnosis result data generated by the guidance model optimized and generated in the qth training stage are used as training sets to learn the parameters of the guided model initialized in the qth training stage. This means that the model parameters should be optimized based on the previously learned network parameters to achieve higher accuracy and robustness. Through this process, the model can be continuously adjusted and improved to better meet the actual needs of industrial equipment diagnosis tasks. Combining the above steps, in the entire training process, by gradually optimizing the model parameters and using methods such as multiple rounds of training and parameter initialization, the guided model can fully learn the patterns and features in the industrial equipment diagnosis task. The final second industrial equipment diagnosis model will have higher accuracy, robustness and generalization ability, and can be used to solve practical industrial equipment diagnosis problems.

[0129] In a possible implementation, step S151 may include:

[0130] Step S1511, after obtaining the guided model optimized and generated in the qth training stage, obtaining first network parameter information of the guided model optimized and generated in the qth training stage and second network parameter information of the guiding model optimized and generated in the q-1th training stage.

[0131] For example, in a wind power generation system scenario, after obtaining the guided model generated by optimization in the qth training phase, the first network parameter information of the model needs to be obtained. In addition, the second network parameter information of the guiding model generated by optimization in the q-1th training phase also needs to be obtained. These network parameter information include weights and biases of each layer and node in the model.

[0132] Step S1512: perform parameter integration on the first network parameter information and the second network parameter information based on the set optimization rule to generate integrated network parameter information.

[0133] After obtaining the first network parameter information and the second network parameter information, it is necessary to use the set optimization rules to integrate their parameters to generate integrated network parameter information. There are many ways to set the optimization rules, such as weighted averaging, parameter sharing, etc. The main purpose of these methods is to combine the network parameter information of the two models so as to fully utilize their advantages in the subsequent model optimization process. For example, in the scenario of a wind power generation system, the weight ratio of the first network parameter information and the second network parameter information can be determined based on experience or experimental results, and then weighted averaging is performed to obtain the integrated network parameter information. The purpose of this is to introduce the advantages of the guiding model while retaining some of the advantages of the guided model, so that the model performs better.

[0134] Step S1513, optimizing the guidance model generated by optimization in the q-1th training stage according to the integrated network parameter information, and generating the guidance model generated by optimization in the qth training stage.

[0135] In this step, the integrated network parameter information obtained previously will be used to optimize the guidance model generated by the optimization in the q-1 training stage. The optimization process can be achieved by adjusting parameters such as weights and biases in the model to improve the accuracy and generalization ability of the model. The optimization algorithm can use methods such as gradient descent and adaptive learning rate. For example, in the scenario of a wind power generation system, the integrated network parameter information can be applied to the layers and nodes of the guidance model to find the optimal parameter update through the gradient descent method. In this process, some optimization constraints may be set to ensure the performance and stability of the model. After the optimization is completed, a new guidance model will be obtained, that is, the guidance model generated by the optimization in the q-1 training stage. This model integrates the advantages of the guided model and the guidance model in terms of network parameters, and can provide more accurate and stable prediction results for industrial equipment diagnosis tasks.

[0136] In a possible implementation, the step S150 performs parameter learning on the guided model based on the benchmark sample industrial control data and the migration sample industrial control data through the fuzzy diagnosis result data generated by the guiding model to generate a second industrial equipment diagnosis model, including:

[0137] Step A110, respectively enhance and extend the benchmark sample industrial control data and the migration sample industrial control data to generate enhanced and extended benchmark sample data and enhanced and extended migration sample data.

[0138] For example, in a wind power generation system scenario, the benchmark sample industrial control data and the migration sample industrial control data need to be enhanced and expanded first. Enhanced expansion is a data enhancement technology that creates new samples by deeply expanding and transforming the features of the original data to increase the diversity of the model's training data. Different feature expansion methods will be used for the benchmark sample industrial control data and the migration sample industrial control data, such as feature transformation from multiple aspects such as time domain, frequency domain, and phase space, and considering the combination of multiple scenarios such as different working conditions and different equipment to generate enhanced extended benchmark sample data and enhanced extended migration sample data. This can make the expanded data more challenging and diverse, which helps to improve the performance of the model.

[0139] Step A120, performing industrial equipment diagnosis on the enhanced extended benchmark sample data and the enhanced extended migration sample data respectively according to the guided model to generate benchmark industrial equipment diagnostic data and migration industrial equipment diagnostic data, wherein the migration industrial equipment diagnostic data reflects the industrial equipment diagnostic data of the migration sample industrial control data in the first industrial equipment diagnostic process, and the benchmark industrial equipment diagnostic data reflects the industrial equipment diagnostic data of the benchmark sample industrial control data in the first industrial equipment diagnostic process.

[0140] In this step, the guided model will be used to perform industrial equipment diagnosis on the enhanced and extended baseline sample data and the migrated sample data. Specifically, the enhanced and extended baseline sample data and the enhanced and extended migrated sample data will be input into the guided model respectively to let it learn the characteristics and patterns of industrial equipment under various working conditions. Then, the baseline industrial equipment diagnosis data and the migrated industrial equipment diagnosis data are output, which can reflect the diagnosis results obtained by the guided model when processing these two types of data.

[0141] Step A130, obtaining benchmark industrial equipment diagnosis annotation data, wherein the benchmark industrial equipment diagnosis annotation data reflects reference diagnosis data of the benchmark sample industrial control data in the first industrial equipment diagnosis process.

[0142] For example, in order to train the guided model and the subsequent second industrial equipment diagnosis model, it is first necessary to obtain reference diagnosis data of the benchmark sample industrial control data, that is, the benchmark industrial equipment diagnosis annotation data. This data is obtained through traditional diagnosis methods or the first industrial equipment diagnosis model without using transfer learning. The benchmark industrial equipment diagnosis annotation data can be used to evaluate the performance of the guided model when processing the benchmark sample data.

[0143] Step A140: lightweight-extend the migration sample industrial control data to generate lightweight-extended migration sample data.

[0144] For example, after obtaining the benchmark industrial equipment diagnosis annotated data, it is necessary to perform lightweight expansion on the migration sample industrial control data. The purpose of lightweight expansion is to generate a new set of data that is similar to the migration sample data but easier to process, thereby reducing the difficulty of the transfer learning process. Lightweight expansion can be achieved by reducing the amount of data, reducing the feature dimension, or reducing noise. The lightweight expansion of the migration sample data can be used to initially introduce the idea of ​​transfer learning when training the second industrial equipment diagnosis model.

[0145] Step A150, fuzzy diagnosis prediction is performed on the lightweight extension migration sample data according to the guidance model to generate fuzzy diagnosis result data, wherein the feature extension depth of the enhanced extension to the data is greater than the feature extension depth of the lightweight extension to the data.

[0146] In this step, the guidance model will be used to make fuzzy diagnosis predictions on the lightweight extension migration sample data. Since the feature extension depth of the enhanced extension migration sample data is greater than that of the lightweight extension migration sample data, the guidance model may capture more fuzziness and interaction information between devices from the lightweight extension migration sample data. Then, the guidance model will generate fuzzy diagnosis result data, which can reflect the more fuzzy diagnosis results obtained by the guidance model when processing the migration sample data.

[0147] Step A160, determining a benchmark training error parameter based on feature differences between the benchmark industrial equipment diagnosis data and the benchmark industrial equipment diagnosis annotation data.

[0148] In this step, the prediction error of the guided model when processing the benchmark industrial equipment diagnostic data needs to be calculated. Specifically, the feature differences between the benchmark industrial equipment diagnostic data output by the guided model and the actual benchmark industrial equipment diagnostic annotated data are compared. These feature differences can be manifested as errors, missing or overfitting of the model prediction results. By analyzing these feature differences, a benchmark training error parameter can be determined, which can measure the performance of the guided model when processing the benchmark sample data.

[0149] Step A170, determining a migration training error parameter based on the feature difference between the migrated industrial equipment diagnosis data and the fuzzy diagnosis result data.

[0150] Similar to the calculation of the baseline training error parameter, a migration training error parameter needs to be determined. Specifically, the prediction error generated by the guided model when processing the migrated industrial equipment diagnostic data will be analyzed. This can be achieved by comparing the feature differences between the migrated industrial equipment diagnostic data and the fuzzy diagnostic result data. These feature differences can reflect the challenges and characteristics encountered by the model when processing the migrated data. By analyzing these feature differences, a

[0151] Step A180: Perform parameter learning on the guided model according to the benchmark training error parameter and the transfer training error parameter to generate the second industrial equipment diagnostic model.

[0152] For example, after obtaining the baseline training error parameter and the transfer training error parameter, these two parameters can be used to adjust the weights and biases of the guided model to obtain a better model. Specifically, various optimization algorithms (such as gradient descent, stochastic gradient descent, etc.) can be used to minimize the sum of the baseline training error parameter and the transfer training error parameter, so as to ensure that the model can better process the transfer sample data while processing the baseline sample data. Ultimately, through parameter learning, a more accurate and more suitable second industrial equipment diagnosis model for actual industrial equipment diagnosis tasks can be obtained.

[0153] For example, in the above embodiments, in order to better explain the above description, a detailed process description can be made through a calculation formula. The following is the corresponding calculation formula.

[0154] 1. Data preparation

[0155] a. Enhanced extended benchmark sample data X_base_aug = AUG (X_base, α_base), where α_base is the depth parameter of enhanced extension and AUG represents the data enhancement method.

[0156] b. Enhanced extended migration sample data X_migrate_aug = AUG (X_migrate, α_migrate), where α_migrate is the depth parameter of enhanced extension, and AUG represents the data enhancement method.

[0157] c. Lightweight extension migration sample data X_migrate_light = LITE (X_migrate), where LITE represents a lightweight extension method.

[0158] 2. Guide model prediction

[0159] a. Baseline industrial equipment diagnostic data Y_base_pred = guidance model (X_base_aug).

[0160] b. Migrate industrial equipment diagnostic data Y_migrate_pred = guidance model (X_migrate_aug).

[0161] 3. Calculate the baseline training error parameters

[0162] The baseline training error parameter E_base can be calculated by the following formula:

[0163] E_base=loss_func(Y_base_pred, Y_base_label), where loss_func represents the loss function (eg, mean square error, cross entropy loss, etc.) between the benchmark industrial equipment diagnostic data prediction and the benchmark industrial equipment diagnostic label.

[0164] 4. Calculate the migration training error parameters

[0165] In this step, it is first necessary to generate reference fuzzy diagnosis result data Y_migrate_fuzzy through the guidance model and calculate the migration training error parameter E_migrate.

[0166] a. The reference fuzzy diagnosis result data Y_migrate_fuzzy contains the probability values ​​of w fault categories, which can be expressed as:

[0167] Y_migrate_fuzzy=softmax(guide_model(X_migrate_aug)).

[0168] Among them, the softmax function converts the probability value output by the guidance model into a probability distribution.

[0169] I apologize for the interruption in my answer. Now continue with the instructions for calculating the transfer training error parameter in step 4.

[0170] b. Calculate the threshold probability value thr_r corresponding to the rth industrial equipment fault label:

[0171] thr_r=average probability value (Y_migrate_fuzzy[r], r=1,...,w), where Y_migrate_fuzzy[r] represents the probability value of the rth fault category in the reference fuzzy diagnosis result data, and w represents the total number of fault categories. The average probability value is a threshold value obtained by aggregating the probability values ​​of each fault category.

[0172] c. Filter the migrated industrial equipment diagnostic data according to the threshold probability value to generate the filtered migrated industrial equipment diagnostic data Y_migrate_filtered:

[0173] If Y_migrate_fuzzy[r]>thr_r, then Y_migrate_filtered[r]=Y_migrate_pred[r], otherwise Y_migrate_filtered[r]=0. Finally, Y_migrate_filtered represents the filtered migrated industrial equipment diagnostic data.

[0174] d. Calculate the migration training error parameter E_migrate:

[0175] E_migrate=loss_func(Y_migrate_filtered,Y_base_label), where loss_func represents the loss function (eg, mean square error, cross entropy loss, etc.) between the filtered migrated industrial equipment diagnostic data and the benchmark industrial equipment diagnostic label.

[0176] 5. Comprehensive training error

[0177] According to the previous calculation, the global error parameter E_global can be calculated by combining the baseline training error parameter and the migration training error parameter.

[0178] E_global = α*E_base+(1-α)*E_migrate, where α is a weight coefficient between 0 and 1, which is used to balance the importance of the baseline training error parameter and the migration training error parameter. When α is close to 1, the model pays more attention to the training error of the baseline sample data; when α is close to 0, the model relies more on the training error of the migration sample data.

[0179] 6. Model parameter optimization

[0180] The global error parameter E_global is used to optimize the parameters of the guided model M_guide. The optimization process can be achieved through optimization methods such as back propagation algorithm and gradient descent method. The optimized parameters generate the second industrial equipment diagnosis model M_sec.

[0181] In a possible implementation, the method further includes: step B110, when performing industrial equipment diagnosis on the enhanced extended benchmark sample data and the enhanced extended migration sample data according to the guided model, performing domain feature consistency analysis on the enhanced extended benchmark sample data and the enhanced extended migration sample data to generate a domain feature consistency error.

[0182] For example, in the process of training the guided model, in order to evaluate the performance of the guided model when processing data from different sources, domain feature consistency analysis is also required. The purpose of this step is to determine the difference in feature representation between the enhanced extended benchmark sample data and the enhanced extended migration sample data, that is, the domain feature consistency error. Specifically, the feature consistency between the two data sets can be measured by calculating metrics such as distance and similarity in the feature space. A larger domain feature consistency error may indicate that the guided model encounters greater difficulties when processing data from different sources, which helps to further understand the performance and characteristics of the model.

[0183] In the above embodiments, most of the calculation processes and formulas have been described in detail. Now, a supplementary description will be given for the above further embodiments.

[0184] The domain feature consistency error calculation mainly includes the following steps:

[0185] a. Calculate the feature representation of the enhanced extended baseline sample data X_base_aug and the enhanced extended migration sample data X_migrate_aug:

[0186] F_base=extract_features(X_base_aug), F_migrate=extract_features(X_migrate_aug)

[0187] Among them, the extract_features() function extracts the feature representation of the data set.

[0188] b. Compute domain feature consistency error Domain_CE:

[0189] Domain_CE=loss_func(F_base, F_migrate), where loss_func represents the loss function between the calculated feature representations (such as mean square error, KL divergence, etc.).

[0190] By integrating and detailing the formulas and calculation processes in the above steps

[0191] Step A180 may include: fusing the reference training error parameter, the transfer training error parameter and the domain feature consistency error to generate a global error parameter, and then performing parameter learning on the guided model according to the global error parameter to generate the second industrial equipment diagnostic model.

[0192] For example, after obtaining the baseline training error parameter, the transfer training error parameter, and the domain feature consistency error, these error metrics need to be integrated into a global error parameter. This can be achieved by weighted averaging or other fusion strategies for the three error parameters. The global error parameter comprehensively considers the performance of the model when processing the baseline sample data and the transfer sample data, as well as the feature consistency of the model between different data sources. By using the global error parameter, the parameters of the guided model can be more comprehensively evaluated and adjusted.

[0193] Finally, the global error parameter is used to adjust the parameters of the guided model to generate a second industrial equipment diagnosis model. Specifically, various optimization algorithms (such as gradient descent, stochastic gradient descent, etc.) can be used to minimize the global error parameter to optimize the weights and biases of the model. In this way, a second industrial equipment diagnosis model that is more accurate, more suitable for actual industrial equipment diagnosis tasks, and has stronger generalization capabilities can be obtained. The second industrial equipment diagnosis model can achieve better performance and accuracy when processing industrial equipment diagnosis tasks from different sources.

[0194] For example, according to the previous calculation, the base training error parameter E_base, the migration training error parameter E_migrate and the domain feature consistency error Domain_CE have been obtained. Now these error parameters are fused to generate the global error parameter E_global.

[0195] Fusion baseline training error parameters and migration training error parameters:

[0196] E_combined = α_1*E_base+α_2*E_migrate+α_3*Domain_CE, where α_1, α_2 and α_3 are three weight coefficients between 0 and 1, which are used to control the influence of each error parameter in the fusion process. Here, the weight coefficient can be determined by methods such as learning rate adjustment and adaptive learning.

[0197] b. Generate global error parameter E_global:

[0198] By adjusting and combining α_1, α_2 and α_3, a global error parameter E_global is generated. This global error parameter will be used in the subsequent guided model parameter learning process to comprehensively optimize the model performance in multiple aspects and generate a second industrial equipment diagnosis model. Therefore, by introducing domain feature consistency errors and fusing multiple error parameters, a more powerful second industrial equipment diagnosis model can be further optimized and generated based on the guided model. This enables the model to better handle industrial equipment diagnosis tasks and provide more accurate and efficient services for industrial equipment in practical applications.

[0199] Wherein, the first industrial equipment diagnosis process includes a fault diagnosis task.

[0200] The step A150 may include:

[0201] Step A151, performing fuzzy diagnosis prediction on the lightweight extended migration sample data according to the guidance model, generating multiple reference fuzzy diagnosis result data and diagnosis probability values ​​corresponding to the multiple reference fuzzy diagnosis result data, wherein the multiple reference fuzzy diagnosis result data reflect the fault diagnosis category generated after fault diagnosis is performed on the lightweight extended migration sample data, and the diagnosis probability value reflects the probability value of the fault diagnosis category.

[0202] In this step, the guidance model is used to perform fuzzy diagnosis prediction on the lightweight extension migration sample data. Specifically, the guidance model outputs multiple reference fuzzy diagnosis result data and the diagnosis probability value corresponding to each reference fuzzy diagnosis result data. These reference fuzzy diagnosis result data represent the fault diagnosis categories predicted by the model when processing the lightweight extension migration sample data, and the diagnosis probability value reflects the model's confidence in these fault diagnosis categories.

[0203] Step A152: outputting the reference fuzzy diagnosis result data whose diagnosis probability value is greater than the set probability value among the plurality of reference fuzzy diagnosis result data as the fuzzy diagnosis result data.

[0204] After obtaining multiple reference fuzzy diagnosis result data and their corresponding diagnosis probability values, it is necessary to filter out data with a diagnosis probability value greater than a set probability value and output these data as fuzzy diagnosis result data. The set probability value is a pre-set threshold used to filter out results with low diagnosis probability and high uncertainty. By setting the probability value, it can be ensured that the output fuzzy diagnosis result data has high credibility and stability, thereby providing more valuable information for model training in subsequent steps.

[0205] Through the above two steps, the guidance model is used to perform fuzzy diagnosis prediction on the lightweight expansion migration sample data, and fuzzy diagnosis result data with high credibility are obtained. These data will serve as an important basis for model training and help improve the accuracy and robustness of the secondary industrial equipment diagnosis model in fault diagnosis tasks.

[0206] In a possible implementation, the method further includes:

[0207] Step C110, performing fuzzy diagnostic prediction on the benchmark sample industrial control data according to the guidance model, generating multiple benchmark industrial equipment diagnostic annotation data and diagnostic probability values ​​corresponding to the multiple benchmark industrial equipment diagnostic annotation data, wherein the multiple diagnostic probability values ​​correspond to w industrial equipment fault labels, and w is a positive integer.

[0208] In this step, the guidance model is first used to perform fuzzy diagnosis prediction on the benchmark sample industrial control data. Specifically, the guidance model will output multiple benchmark industrial equipment diagnostic annotation data and the diagnostic probability value corresponding to each benchmark industrial equipment diagnostic annotation data. These benchmark industrial equipment diagnostic annotation data represent the possible fault types of industrial equipment predicted by the model when processing the benchmark sample industrial control data, and the diagnostic probability value reflects the degree of credibility of the guidance model for these fault types. The diagnostic probability value corresponds to w industrial equipment fault labels, where w represents the number of all possible fault types.

[0209] Step C120 , calculating an average probability value of multiple diagnosis probability values ​​corresponding to the r th industrial equipment fault label as a threshold probability value corresponding to the r th industrial equipment fault label, where r≤w and r is a positive integer.

[0210] In this step, it is necessary to calculate the average probability value of the diagnostic probability value corresponding to each industrial equipment fault label. Specifically, the multiple diagnostic probability values ​​corresponding to the rth industrial equipment fault label are aggregated, and their average value is calculated to obtain the threshold probability value corresponding to the label. The threshold probability value is a threshold value used in the subsequent screening of the reference fuzzy diagnosis result data. By calculating the average probability value, the credibility of each fault label can be better grasped, and those fault labels with lower probability values ​​can be eliminated, so as to retain relatively credible fault labels in the fuzzy diagnosis result data.

[0211] In step A152, if the reference fuzzy diagnosis result data corresponding to the r-th industrial equipment fault label among the multiple reference fuzzy diagnosis result data is greater than the threshold probability value corresponding to the r-th industrial equipment fault label, the reference fuzzy diagnosis result data corresponding to the r-th industrial equipment fault label is output as the fuzzy diagnosis result data.

[0212] Finally, the reference fuzzy diagnosis result data needs to be screened according to the threshold probability value calculated previously. Specifically, for each industrial equipment fault label, it will be checked whether the predicted value corresponding to the label in the reference fuzzy diagnosis result data is greater than the corresponding threshold probability value. If so, this reference fuzzy diagnosis result data is output as the fuzzy diagnosis result data. In this way, more reliable and accurate fuzzy diagnosis result data can be obtained, which can be further used to optimize the performance of the model and provide a more reliable basis for subsequent industrial equipment fault diagnosis tasks.

[0213] In summary, by performing fuzzy diagnosis prediction on the benchmark sample industrial control data and calculating the diagnosis probability value and average probability value, more reliable fuzzy diagnosis result data is obtained. These data will serve as important information for the optimization model and help generate an industrial equipment fault diagnosis model.

[0214] Figure 2 The hardware structure of the industrial Internet system 100 provided in the embodiment of the present application for implementing the above-mentioned big data processing method applied to the industrial Internet is shown, as shown in FIG. Figure 2 As shown, the industrial Internet system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0215] In an alternative embodiment, the industrial Internet system 100 can be a single server or a server group. The server group can be centralized or distributed (for example, the industrial Internet system 100 can be a distributed system). In an alternative embodiment, the industrial Internet system 100 can be local or remote. For example, the industrial Internet system 100 can access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the industrial Internet system 100 can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In an alternative embodiment, the industrial Internet system 100 can be implemented on a cloud platform. By way of example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any combination thereof.

[0216] The machine-readable storage medium 120 can store data and / or instructions. In an alternative embodiment, the machine-readable storage medium 120 can store data obtained from an external terminal. In an alternative embodiment, the machine-readable storage medium 120 can store data and / or instructions used by the Industrial Internet system 100 to execute or use to complete the exemplary methods described in this application. In an alternative embodiment, the machine-readable storage medium 120 may include a large-capacity memory, a removable memory, a volatile read-write memory, a read-only memory, etc. or any combination thereof. Exemplary large-capacity memory may include a magnetic disk, an optical disk, a solid-state disk, etc. Exemplary removable memory may include a flash drive, a floppy disk, an optical disk, a memory card, a compressed disk, a magnetic tape, etc.

[0217] During the specific implementation process, multiple processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the big data processing method applied to the industrial Internet in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0218] The specific implementation process of the processor 110 can be found in the various method embodiments executed by the above-mentioned industrial Internet system 100. The implementation principles and technical effects are similar and will not be repeated here in this embodiment.

[0219] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the big data processing method applied to the industrial Internet as above is implemented.

[0220] Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A big data processing method applied to industrial Internet, characterized in that: The method comprises: Acquire a first industrial equipment diagnosis model, wherein the first industrial equipment diagnosis model is a neural network model generated by learning network parameter information using benchmark sample industrial control data in a first industrial equipment diagnosis process, wherein the first industrial equipment diagnosis model includes a first graph self-attention network; Performing feature conversion on the migration sample industrial control data to generate migration sample conversion data, and performing graph self-attention feature analysis on the migration sample conversion data according to the first graph self-attention network to generate graph self-attention features; Performing parameter learning on the first image self-attention network according to the image self-attention features and the migration sample industrial control data to generate a second image self-attention network; Optimizing the first industrial equipment diagnosis model according to the second graph self-attention network to generate a guiding model and a guided model; According to the fuzzy diagnosis result data generated by the guiding model using the benchmark sample industrial control data and the migration sample industrial control data, the guided model is subjected to parameter learning to generate a second industrial equipment diagnosis model, wherein the second industrial equipment diagnosis model is used to perform industrial equipment diagnosis on the benchmark industrial control data or the migration industrial equipment data in the target industrial Internet big data during the first industrial equipment diagnosis process; The fuzzy diagnosis result data generated by the guiding model based on the benchmark sample industrial control data and the migration sample industrial control data, performing parameter learning on the guided model to generate a second industrial equipment diagnosis model, includes: Strengthening and extending the benchmark sample industrial control data and the migration sample industrial control data respectively to generate strengthened and extended benchmark sample data and strengthened and extended migration sample data; Performing industrial equipment diagnosis on the enhanced extended benchmark sample data and the enhanced extended migration sample data respectively according to the guided model to generate benchmark industrial equipment diagnostic data and migration industrial equipment diagnostic data, wherein the migration industrial equipment diagnostic data reflects the industrial equipment diagnostic data of the migration sample industrial control data in the first industrial equipment diagnostic process, and the benchmark industrial equipment diagnostic data reflects the industrial equipment diagnostic data of the benchmark sample industrial control data in the first industrial equipment diagnostic process; Acquire benchmark industrial equipment diagnostic annotation data, wherein the benchmark industrial equipment diagnostic annotation data reflects reference diagnostic data of the benchmark sample industrial control data in the first industrial equipment diagnostic process; Performing lightweight extension on the migration sample industrial control data to generate lightweight extended migration sample data; Performing fuzzy diagnosis prediction on the lightweight extension migration sample data according to the guidance model to generate fuzzy diagnosis result data; wherein the feature extension depth of the enhanced extension to the data is greater than the feature extension depth of the lightweight extension to the data; Determining a benchmark training error parameter based on feature differences between the benchmark industrial equipment diagnostic data and the benchmark industrial equipment diagnostic annotated data; Determining a migration training error parameter according to a feature difference between the migration industrial equipment diagnosis data and the fuzzy diagnosis result data; Parameter learning is performed on the guided model according to the benchmark training error parameter and the transfer training error parameter to generate the second industrial equipment diagnostic model.

2. The big data processing method applied to industrial Internet according to claim 1 is characterized in that: The step of performing feature conversion on the migration sample industrial control data to generate migration sample conversion data, and performing graph self-attention feature analysis on the migration sample conversion data according to the first graph self-attention network includes: Performing data obfuscation processing on the migration sample industrial control data to generate the migration sample industrial control data after the data obfuscation processing as the migration sample conversion data; Loading the migrated sample industrial control data after the data obfuscation processing into the first graph self-attention network to perform graph self-attention feature analysis to generate obfuscated graph self-attention features; The step of performing parameter learning on the first graph self-attention network according to the graph self-attention feature and the migration sample industrial control data to generate a second graph self-attention network includes: Loading the self-attention features of the confusion graph into a confusion restoration network for feature restoration processing to generate confusion restoration data, wherein the confusion restoration data reflects the data result generated by reconstructing the migration sample industrial control data after the data obfuscation processing; Parameters of the first graph self-attention network are learned according to feature differences between the obfuscated restored data and the migrated sample industrial control data to generate the second graph self-attention network.

3. The big data processing method applied to industrial Internet according to claim 2 is characterized in that: The migration sample industrial control data includes migration sample device session data, and the migration sample device session data includes k state session events, where k is an integer greater than 1; The performing data obfuscation processing on the migration sample industrial control data to generate the migration sample industrial control data after the data obfuscation processing as the migration sample conversion data includes: Performing session vector analysis on the migration sample device session data to generate session vector path data corresponding to the migration sample device session data, wherein the session vector path data includes session link vectors corresponding to the k state session events respectively; Performing data obfuscation processing on the session vector path data to generate an obfuscated processing vector as the migration sample industrial control data after the data obfuscation processing, wherein the obfuscated processing vector includes x hidden session link vectors, where x<k and x is a positive integer; The step of loading the self-attention features of the confusion map into a confusion restoration network for feature restoration processing to generate confusion restoration data includes: Loading the confusion graph self-attention features into the confusion restoration network, performing industrial equipment diagnosis on the state session events corresponding to the x hidden session link vectors, and generating x derived state session events; The reconstructed device session data corresponding to the migration sample device session data is obtained according to the x derived state session events, and the reconstructed device session data is used as the obfuscated restoration data.

4. The big data processing method applied to industrial Internet according to claim 1 is characterized in that: The step of performing feature conversion on the migration sample industrial control data to generate migration sample conversion data, and performing graph self-attention feature analysis on the migration sample conversion data according to the first graph self-attention network includes: Performing feature expansion on the migration sample industrial control data to generate expanded migration sample industrial control data as the migration sample conversion data, wherein the expanded migration sample industrial control data includes valid training data and invalid training data corresponding to the migration sample industrial control data; Loading the migration sample industrial control data, the valid training data and the invalid training data into the first graph self-attention network for graph self-attention feature analysis, generating migration graph self-attention features corresponding to the migration sample industrial control data, valid graph self-attention features corresponding to the valid training data and invalid graph self-attention features corresponding to the invalid training data; The step of performing parameter learning on the first graph self-attention network according to the graph self-attention feature and the migration sample industrial control data to generate a second graph self-attention network includes: The first graph self-attention network is trained in a direction of minimizing the feature difference between the transition graph self-attention feature and the valid graph self-attention feature and maximizing the feature difference between the transition graph self-attention feature and the invalid graph self-attention feature, and parameter learning is performed on the first graph self-attention network to generate the second graph self-attention network.

5. The big data processing method applied to industrial Internet according to any one of claims 1 to 4, characterized in that: The fuzzy diagnosis result data generated by the guiding model based on the benchmark sample industrial control data and the migration sample industrial control data, performing parameter learning on the guided model to generate a second industrial equipment diagnosis model, includes: In the polling optimization of the qth training stage, the guided model optimized and generated in the q-1th training stage is optimized according to the fuzzy diagnosis result data generated by the guidance model optimized and generated in the q-1th training stage through the reference sample industrial control data and the migration sample industrial control data, so as to generate the guided model optimized and generated in the qth training stage and the guidance model optimized and generated in the qth training stage; Initializing the network parameter information of the graph self-attention network in the guided model generated by optimization in the qth training stage as the network parameter information of the second graph self-attention network, to generate the guided model initialized in the qth training stage; According to the fuzzy diagnosis result data generated by the guidance model optimized and generated according to the benchmark sample industrial control data and the migration sample industrial control data in the qth training stage, parameter learning is performed on the guided model initialized in the qth training stage to generate the second industrial equipment diagnosis model.

6. The big data processing method applied to industrial Internet according to claim 5 is characterized in that: The step of generating the guided model generated by optimization in the qth training stage and the guiding model generated by optimization in the qth training stage comprises: After obtaining the guided model generated by optimization in the qth training stage, obtaining first network parameter information of the guided model generated by optimization in the qth training stage and second network parameter information of the guiding model generated by optimization in the q-1th training stage; Performing parameter integration on the first network parameter information and the second network parameter information based on a set optimization rule to generate integrated network parameter information; The guidance model generated by optimization in the q-1th training stage is optimized according to the integrated network parameter information to generate the guidance model generated by optimization in the qth training stage.

7. The big data processing method applied to industrial Internet according to claim 1 is characterized in that: The method further comprises: When industrial equipment diagnosis is performed on the enhanced extended benchmark sample data and the enhanced extended migration sample data according to the guided model, domain feature consistency analysis is performed on the enhanced extended benchmark sample data and the enhanced extended migration sample data to generate a domain feature consistency error; The step of performing parameter learning on the guided model according to the reference training error parameter and the migration training error parameter to generate the second industrial equipment diagnosis model includes: The reference training error parameter, the migration training error parameter and the domain feature consistency error are integrated to generate a global error parameter; performing parameter learning on the guided model according to the global error parameter to generate the second industrial equipment diagnostic model; Wherein, the first industrial equipment diagnosis process includes a fault diagnosis task; The step of performing fuzzy diagnosis prediction on the lightweight extended migration sample data according to the guidance model to generate fuzzy diagnosis result data includes: Performing fuzzy diagnosis prediction on the lightweight extended migration sample data according to the guidance model, generating a plurality of reference fuzzy diagnosis result data and diagnosis probability values ​​respectively corresponding to the plurality of reference fuzzy diagnosis result data, wherein the plurality of reference fuzzy diagnosis result data reflect the fault diagnosis category generated after the fault diagnosis is performed on the lightweight extended migration sample data, and the diagnosis probability value reflects the probability value of the fault diagnosis category; The reference fuzzy diagnosis result data having a diagnosis probability value greater than a set probability value among the plurality of reference fuzzy diagnosis result data are output as the fuzzy diagnosis result data.

8. The big data processing method applied to industrial Internet according to claim 7 is characterized in that: The method further comprises: Performing fuzzy diagnosis prediction on the benchmark sample industrial control data according to the guidance model, generating a plurality of benchmark industrial equipment diagnosis annotation data and diagnosis probability values ​​corresponding to the plurality of benchmark industrial equipment diagnosis annotation data, wherein the plurality of diagnosis probability values ​​correspond to w industrial equipment fault labels, and w is a positive integer; Calculate the average probability value of multiple diagnosis probability values ​​corresponding to the r-th industrial equipment fault label as the threshold probability value corresponding to the r-th industrial equipment fault label, r≤w and r is a positive integer; The step of outputting the reference fuzzy diagnosis result data having a diagnosis probability value greater than a set probability value among the plurality of reference fuzzy diagnosis result data as the fuzzy diagnosis result data comprises: If the reference fuzzy diagnosis result data corresponding to the r-th industrial equipment fault label among the multiple reference fuzzy diagnosis result data is greater than the threshold probability value corresponding to the r-th industrial equipment fault label, the reference fuzzy diagnosis result data corresponding to the r-th industrial equipment fault label is output as the fuzzy diagnosis result data.

9. An industrial Internet system, characterized in that: The industrial Internet system includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the big data processing method applied to the industrial Internet according to any one of claims 1 to 8.

Citation Information

Patent Citations

  • Intelligent mechanical fault diagnosis method based on deep transfer learning

    CN112906644A

  • Method for obtaining distribution line fault classification, positioning and line selection deep learning model based on transfer learning

    CN114169231A

  • Method and device for intelligent fault diagnosis and operation and maintenance guidance of thermal power plant equipment, medium and equipment

    CN114169545A

  • Motor rolling bearing fault diagnosis method of dynamic multi-adversarial adaptive network based on transfer learning

    CN116429417A

  • Small sample fault classification method and system for gearbox

    CN116578907A