Data processing method and electronic equipment suitable for training and assessment system

Through vectorized representation and recursive neural network modeling, combined with heat map and graph structure feature extraction, the problem that traditional fault diagnosis methods are difficult to describe fault association relationships in complex systems is solved, and efficient fault diagnosis of training and assessment systems is realized.

CN118761688BActive Publication Date: 2024-11-19STATE GRID SICHUAN ELECTRIC POWER CO +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411253934.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-09
Publication Date
2024-11-19
Estimated Expiration
2044-09-09

AI Technical Summary

Technical Problem

Traditional fault diagnosis methods are difficult to accurately describe the complex relationship between fault knowledge points in complex systems, resulting in limited diagnostic efficiency and accuracy.

Method used

By vectorizing the target system program operation data and referring to system fault knowledge points, generating program operation vector data and fault knowledge points vector data, using recursive neural network to model the connection vector between fault knowledge points, and through heat map and graph structure Feature extraction improves fault matching accuracy.

Benefits of technology

It significantly improves the accuracy and efficiency of system fault diagnosis, realizes the rapid and accurate identification of complex system faults, and provides guarantees for the stable operation of the training and assessment system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118761688B_ABST
    Figure CN118761688B_ABST
Patent Text Reader

Abstract

The present application provides a data processing method and electronic equipment suitable for a training and assessment system. By vectorizing the target system program operation data and the reference system program fault knowledge points, the key features of the data can be efficiently captured. The heat map generated by the program operation vector data can intuitively display the degree of match between the target data and the set fault label. Furthermore, the connectivity vectors between the fault knowledge points are modeled in combination with a recursive neural network, which deeply explores the complex associations between the fault knowledge points and enhances the depth and breadth of fault diagnosis. The target fault knowledge point vector data generated by graph structure feature extraction refines the representation of fault features and improves the accuracy of fault matching. Finally, the target system program fault knowledge points are determined based on the estimated thermal value, which realizes the rapid and accurate identification of system faults and provides a strong guarantee for the stable operation of the training and assessment system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a data processing method and electronic equipment suitable for a training and assessment system. Background Art

[0002] In the actual application of training and assessment systems, the stability and reliability of system program operation are crucial to ensuring the effectiveness of training. However, with the increase in system complexity, various faults occasionally occur during program operation, which not only affect the progress of training, but may also cause irreversible damage to the system. Traditional fault diagnosis methods often rely on manual experience and static rule bases, which are difficult to cope with complex and changeable system fault conditions, and have limited diagnostic efficiency and accuracy.

[0003] In order to improve the fault diagnosis capability of the training and assessment system, with the rapid development of big data processing technology and artificial intelligence algorithms, fault diagnosis methods based on data-driven and machine learning have gradually become a research hotspot. These methods can automatically identify and locate system faults through in-depth mining and analysis of system program operation data, providing strong support for system maintenance and management.

[0004] However, related technologies still face some challenges when applied to training and assessment systems. For example, there are often complex correlations between different fault knowledge points, and traditional linear models are difficult to accurately describe such relationships. Therefore, more advanced nonlinear models need to be introduced for modeling. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a data processing method applicable to a training assessment system, the method comprising:

[0006] Obtain target system program operation data and multiple reference system program fault knowledge points of the training and assessment system, and respectively vectorize the target system program operation data and each reference system program fault knowledge point to generate program operation vector data of the target system program operation data and fault knowledge point vector data of each reference system program fault knowledge point;

[0007] Based on the program running vector data, estimating a heat map of the target system program running data matching each set program fault label;

[0008] Obtaining a recursive neural network corresponding to each set program fault label, wherein the recursive neural network is used to reflect the connectivity vectors between each reference system program fault knowledge point under the set program fault label;

[0009] According to the heat map and the recursive neural network corresponding to each set program fault label, the fault knowledge point vector data of each reference system program fault knowledge point is extracted with graph structure features to generate target fault knowledge point vector data corresponding to each reference system program fault knowledge point;

[0010] Determine, based on the target fault knowledge point vector data corresponding to each reference system program fault knowledge point, an estimated thermal value of each reference system program fault knowledge point matching the target system program running data;

[0011] Based on the estimated thermal value, a target system program fault knowledge point corresponding to the target system program operation data is determined from the respective reference system program fault knowledge points.

[0012] On the other hand, an embodiment of the present application also provides an electronic device, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0013] Based on the above aspects, the present application significantly improves the accuracy and efficiency of system fault diagnosis. By vectorizing the target system program operation data and the reference system program fault knowledge points, the key features of the data can be efficiently captured, laying a solid foundation for subsequent analysis. The heat map generated by the program operation vector data can intuitively display the degree of match between the target data and the set fault label, providing strong support for the initial fault location. Furthermore, the connectivity vectors between the fault knowledge points are modeled in combination with the recursive neural network, which deeply explores the complex associations between the fault knowledge points and enhances the depth and breadth of fault diagnosis. The target fault knowledge point vector data generated by graph structure feature extraction refines the representation of fault features and improves the accuracy of fault matching. Finally, the target system program fault knowledge points are determined based on the estimated thermal value, which realizes the rapid and accurate identification of system faults, and provides a strong guarantee for the stable operation of the training and assessment system. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the execution flow of the data processing method applicable to the training and assessment system provided in the embodiment of the present application.

[0015] Figure 2 It is a schematic diagram of the hardware architecture of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0016] The present application will be described in detail below with reference to the accompanying drawings. Figure 1It is a flow chart of a data processing method applicable to a training and assessment system provided by an embodiment of the present application. The data processing method applicable to a training and assessment system is introduced in detail below.

[0017] Step S110, obtain the target system program operation data and multiple reference system program fault knowledge points of the training and assessment system, and vectorize the target system program operation data and each reference system program fault knowledge point respectively to generate program operation vector data of the target system program operation data and fault knowledge point vector data of each reference system program fault knowledge point.

[0018] In detail, the target system program running data refers to the data collected from the real-time running environment of the training and assessment system, which reflects the current status and performance of the training and assessment system. For example, it may include: CPU usage: indicating the current usage of the system CPU. High CPU usage may mean that the system is performing a large number of computing tasks or there is a performance bottleneck; memory usage: reflecting the usage of the system memory. High memory usage may cause the system to respond slowly or crash; disk I / O: indicating the frequency of disk read and write operations. High disk I / O may mean that the system is frequently reading and writing data; network request response time: recording the time required for the system to process network requests. High response time may indicate network delays or electronic equipment performance problems; log output of specific modules: contains the operation logs of specific modules of the system. These logs may contain error messages, warnings or other information useful for diagnosis.

[0019] The reference system program fault knowledge points refer to knowledge base entries about various possible faults of the system and related information compiled and verified by experts. Each fault knowledge point describes the fault phenomenon, possible causes, impact scope and solution in detail.

[0020] For example, fault knowledge point 1: memory leak

[0021] Symptom: System memory usage continues to increase, but the allocated memory is not released.

[0022] Cause: There is a code segment in the program that does not manage memory correctly.

[0023] Scope of impact: May cause the system to respond slowly, crash, or fail to start.

[0024] Solution: Find and fix the memory leak code and optimize the memory management strategy.

[0025] Troubleshooting knowledge point 2: Network delay

[0026] Phenomenon: The time it takes for the system to process network requests increases, and the response time becomes longer.

[0027] Causes: Insufficient network bandwidth, network equipment failure, network congestion, etc.

[0028] Scope of impact: Impact on user experience and may cause business interruption.

[0029] Solution: Optimize network configuration, increase bandwidth, check network device status, etc.

[0030] The program running vector data refers to the vector form of the target system program running data in high-dimensional space after vectorization, which contains the key feature information of the training and assessment system operation status. For example, assuming that the target system program running data contains two features, CPU usage and memory occupancy, after being processed by the embedding layer, these two features are mapped into two high-dimensional vectors. These two vectors are concatenated to form the program running vector data for subsequent analysis and processing.

[0031] The fault knowledge point vector data refers to the vector form of the reference system program fault knowledge point in high-dimensional space after vectorization, which contains the key feature information and semantic information of the fault knowledge point. For example, for the fault knowledge point of "memory leakage", the electronic device may use word embedding technology to convert its description text into a high-dimensional vector. This vector is the fault knowledge point vector data, which is used to represent the key features and semantic information of the "memory leakage" fault.

[0032] That is, in this embodiment, a training and assessment system may be composed of multiple interdependent modules for performing online training and assessment tasks for employees. Due to the complexity and variability of the training and assessment system, program operation anomalies may occasionally occur, affecting the smooth progress of the training and assessment. In order to quickly locate and solve these problems, the electronic device needs to obtain and analyze the program operation data of the training and assessment system in real time, and match it with known fault knowledge points.

[0033] Specifically, the electronic device can collect the target system program running data in real time through the monitoring tools deployed at each key node of the system. Among them, the collected original target system program running data may contain noise or abnormal values. The electronic device will clean and pre-process these target system program running data to ensure the accuracy of subsequent analysis. For example, data points beyond the reasonable range are eliminated, and missing values ​​are interpolated.

[0034] At the same time, the electronic device can access the pre-built system program fault knowledge base. The system program fault knowledge base contains multiple reference system program fault knowledge points collated and verified by experts. Each reference system program fault knowledge point describes the fault phenomenon, possible causes, impact scope and solution in detail. The electronic device extracts all relevant reference system program fault knowledge points from the system program fault knowledge base and organizes them in a certain format for subsequent processing and analysis.

[0035] On this basis, the electronic device divides the target system program running data into multiple feature spaces according to the characteristics of the target system program running data. Each feature space corresponds to a type of system program running characteristics, such as performance feature space, log feature space, etc. For the system program running characteristics under each feature space, the electronic device uses the embedding layer in the fault diagnosis network model to embed and represent them. The embedding layer maps each feature to a vector (knowledge structure vector) in a high-dimensional space, which can capture the distribution characteristics and correlation relationships of the system program running characteristics in the data set. The electronic device assigns weights to the knowledge structure vectors according to the importance of each feature space. The weight assignment process may involve multiple iterations to ensure that the data structure under each feature space can be reasonably represented.

[0036] In addition, the electronic device performs preprocessing operations such as word segmentation and stop word removal on the description text of the fault knowledge point to extract key information. For example, using word embedding technology (such as Word2Vec, GloVe, etc.), the electronic device maps each word in the fault knowledge point to a vector in a high-dimensional space (fault knowledge point representation vector). Then, the vectors of all words in the fault knowledge point are aggregated (such as average pooling, maximum pooling, etc.) to generate the fault knowledge point vector data of the fault knowledge point. At the same time, the electronic device will also consider the prior associated knowledge point vector data and construct a target knowledge point representation vector set to enrich the representation of the fault knowledge point.

[0037] Step S120 , based on the program running vector data, estimating a heat map of the target system program running data matching each set program fault label.

[0038] In detail, a heat map is a data visualization technology that is used to show the degree of match between the target system program running data and each set program fault label. Each grid in the heat map represents a fault label, and the color of the grid indicates the degree of match (the darker the color, the higher the match). For example, suppose there are three set program fault labels: "memory leak", "network delay" and "module abnormality". The electronic device calculates the matching probability of the target system program running data and these three labels, and displays these probability values ​​in the form of a heat map. If the target data has the highest probability of matching "memory leak", the grid corresponding to "memory leak" has the darkest color.

[0039] That is, in this embodiment, the electronic device processes the program running vector data using the classification layer in the fault diagnosis network model. The classification layer usually includes a fully connected layer and a Softmax activation function, which is used to calculate the probability that the program running vector data matches each set program fault label. Through the processing of the classification layer, the electronic device obtains the probability value of the program running vector data matching each set program fault label, and these probability values ​​reflect the matching degree between the target system program running data and each fault label.

[0040] As a result, the electronic device visualizes the calculated probability value in the form of a heat map. Each grid in the heat map represents a set program fault label, and the color of the grid indicates the degree of matching (the darker the color, the higher the matching degree). By observing the heat map, the electronic device can quickly identify the several fault labels with the highest matching degree with the target system program running data. These labels provide important clues for the subsequent graph structure feature extraction and fault knowledge point determination.

[0041] Step S130, obtaining a recursive neural network corresponding to each set program fault label, wherein the recursive neural network is used to reflect the connectivity vectors between each reference system program fault knowledge point under the set program fault label.

[0042] In detail, the recursive neural network is a special neural network structure used to process sequence data. In fault diagnosis, RNN can capture the connectivity vectors and association relationships between each reference system program fault knowledge point under different fault labels. The connectivity vector refers to the weight vector between two adjacent network members in the recursive neural network, which reflects the strength of association and mutual influence between them.

[0043] That is, in this embodiment, for each set program fault label, the electronic device pre-builds and trains a corresponding recursive neural network (RNN) model, which is trained based on a large amount of historical fault data and expert knowledge, and can capture the connectivity vectors and association relationships between the reference system program fault knowledge points under a specific fault label. The trained RNN model is stored in a local or remote database of the electronic device for subsequent retrieval and loading.

[0044] Each RNN model consists of multiple network members (nodes), each of which represents a reference system program fault knowledge point. Network members are connected through links, and the weight on the link reflects the connectivity vector and association strength between two adjacent network members.

[0045] Step S140, based on the heat map and the recursive neural network corresponding to each set program fault label, the fault knowledge point vector data of each reference system program fault knowledge point is extracted with graph structure features to generate target fault knowledge point vector data corresponding to each reference system program fault knowledge point.

[0046] In this embodiment, for each reference system program fault knowledge point, the electronic device determines the target network member in the RNN model corresponding to each set program fault label. In detail, the electronic device extracts the adjacent vector data of the target network member based on the connectivity vector between the target network member and other network members in the RNN model. These adjacent vector data reflect the association relationship between the target knowledge point and other knowledge points under a specific fault label. The electronic device performs weighted aggregation on the adjacent vector data under each set program fault label based on the thermal value information in the thermal map generated in step S120. The adjacent vector data corresponding to the fault label with a higher matching degree has a greater weight in the aggregation process.

[0047] The electronic device fuses the aggregated adjacent vector data with the original fault knowledge point vector data (such as weighted summation, splicing, etc.) to generate target fault knowledge point vector data corresponding to the reference system program fault knowledge points. These target fault knowledge point vector data contain richer graph structure features and associated information.

[0048] Step S150, according to the target fault knowledge point vector data corresponding to each reference system program fault knowledge point, determine the estimated thermal value of each reference system program fault knowledge point matching the target system program operation data, and based on the estimated thermal value, determine the target system program fault knowledge point corresponding to the target system program operation data from the each reference system program fault knowledge point.

[0049] In this embodiment, for each reference system program fault knowledge point, the electronic device aggregates its fault knowledge point vector data and the target fault knowledge point vector data (such as cosine similarity calculation, dot product operation, etc.) to generate an estimated thermal value of the knowledge point matching the target system program running data.

[0050] In detail, the electronic device sorts the estimated thermal values ​​of all reference system program fault knowledge points, and selects several knowledge points with the highest thermal values ​​as candidate target system program fault knowledge points. If the number of candidate knowledge points is large or further confirmation is required, the electronic device can combine other information (such as fault history records, expert opinions, etc.) for comprehensive analysis and judgment, and finally determine the target system program fault knowledge point.

[0051] Thus, the electronic device outputs the determined target system program fault knowledge points as fault diagnosis results to relevant personnel or systems for processing and resolution. At the same time, the electronic device can also record the process and results of this fault diagnosis in a log for subsequent analysis and audit.

[0052] Based on the above steps, the embodiment of the present application significantly improves the accuracy and efficiency of system fault diagnosis. By vectorizing the target system program operation data and the reference system program fault knowledge points, the key features of the data can be efficiently captured, laying a solid foundation for subsequent analysis. The heat map generated by the program operation vector data can intuitively display the degree of match between the target data and the set fault label, providing strong support for the initial fault location. Furthermore, the connectivity vectors between the fault knowledge points are modeled in combination with the recursive neural network, and the complex associations between the fault knowledge points are deeply excavated, which enhances the depth and breadth of fault diagnosis. The target fault knowledge point vector data generated by graph structure feature extraction refines the representation of fault features and improves the accuracy of fault matching. Finally, the target system program fault knowledge points are determined based on the estimated thermal value, which realizes the rapid and accurate identification of system faults, and provides a strong guarantee for the stable operation of the training and assessment system.

[0053] In a possible implementation, step S110 includes:

[0054] Step S111 , embedding and representing the system program running features of the target system program running data in at least one feature space, and generating a knowledge structure vector of the target system program running data in at least one feature space.

[0055] In this embodiment, the electronic device first divides the target system program running data into multiple feature spaces according to the characteristics of the target system program running data. These feature spaces may include performance feature space (such as CPU usage, memory usage), log feature space (including log output of specific modules), network feature space (such as network request response time), etc. Each feature space focuses on a specific type of system program running characteristics.

[0056] For each system program running feature in the feature space, the electronic device uses the embedding layer in the fault diagnosis network model to embed the representation. The embedding layer is a pre-trained neural network layer that can map each feature to a vector (knowledge structure vector) in a high-dimensional space. For example, for the feature of CPU usage, the embedding layer generates a vector containing the distribution characteristics and association relationships of the feature in the data set.

[0057] Since the importance of data structures in different feature spaces to fault diagnosis may be different, the electronic device needs to assign importance weights to the knowledge structure vectors in each feature space. This process may involve multiple iterations to ensure that the data structure in each feature space can be reasonably represented. In each iteration, the electronic device adjusts the weight distribution in each feature space based on the current feature importance evaluation results.

[0058] Step S112, embedding and representing each reference system program fault knowledge point respectively, and generating a fault knowledge point representation vector of each reference system program fault knowledge point.

[0059] In this embodiment, the electronic device performs preprocessing operations such as word segmentation and stop word removal on the description text of the reference system program fault knowledge point to extract key information. This step helps to remove noise in the text and retain information useful for fault diagnosis.

[0060] Then, using word embedding technology (such as Word2Vec, GloVe, etc.), the electronic device maps each word in the reference system program fault knowledge point to a vector in a high-dimensional space (fault knowledge point representation vector), which can capture the semantic information and contextual relationship of each word in the text. In order to generate the vector representation of the entire reference system program fault knowledge point, the electronic device aggregates the vectors of all words in the reference system program fault knowledge point (such as average pooling, maximum pooling, etc.). The aggregated vector (fault knowledge point vector data) can reflect the key information and features of the entire reference system program fault knowledge point.

[0061] Step S113, constructing a target knowledge point representation vector set according to the fault knowledge point representation vectors of each reference system program fault knowledge point and the prior associated knowledge point vector data.

[0062] In addition to the vector representation of the fault knowledge point itself, the electronic device also considers the prior associated knowledge point vector data. These prior associated knowledge points can be predetermined through historical data or expert knowledge and have some association or influence with the current fault knowledge point. The electronic device combines these prior associated knowledge point vectors with the fault knowledge point vectors to construct a target knowledge point representation vector set to enrich the representation of the fault knowledge point.

[0063] Step S114: determining the knowledge structure vector in the target feature space to be currently weighted from the knowledge structure vector in the at least one feature space.

[0064] Step S115, assigning importance weights to the knowledge structure vector in the target feature space and the target knowledge point representation vector set, generating importance weight distribution information of the target system program running data and importance weight distribution information of each reference system program fault knowledge point.

[0065] Step S116, determining an iterative target knowledge point representation vector set according to the importance weight distribution information of the target system program running data and the importance weight distribution information of each reference system program fault knowledge point.

[0066] Step S117, return to execute step S114 until the knowledge structure vectors in all feature spaces complete the importance weight assignment, and obtain the program running vector data of the target system program running data and the fault knowledge point vector data of each reference system program fault knowledge point based on the current iteration target knowledge point representation vector set.

[0067] In this embodiment, the electronic device determines the target feature space to be currently weighted from the knowledge structure vector under at least one feature space. This process may be performed in a certain order (such as descending order of importance) or randomly selected.

[0068] For the knowledge structure vector under the selected target feature space, the electronic device assigns importance weights to it and the target knowledge point representation vector set. This process aims to determine which data structures under the feature space are more important for the representation of the current fault knowledge point and adjust the weight distribution accordingly.

[0069] Through the weight allocation process, the electronic device generates the importance weight distribution information of the target system program running data and the importance weight distribution information of each reference system program fault knowledge point, which will be used in the subsequent analysis and matching process.

[0070] The electronic device returns to the step of determining the target feature space to be weighted from the knowledge structure vectors in the feature space until the importance weights of the knowledge structure vectors in all feature spaces are assigned. In each iteration, the electronic device updates the target knowledge point representation vector set according to the current importance weight distribution information.

[0071] After multiple iterations, the electronic device obtains the program running vector data of the target system program running data and the fault knowledge point vector data of each reference system program fault knowledge point based on the vector set of the current iteration target knowledge point representation. These vector data will be used for subsequent steps such as thermal map estimation, graph structure feature extraction and fault knowledge point matching.

[0072] In a possible implementation, the recursive neural network includes network members corresponding to each reference system program fault knowledge point, and links between network members, where the links are used to reflect the connectivity vectors between two adjacent network members.

[0073] Step S140 includes:

[0074] Step S141, for each reference system program fault knowledge point, determine the target network member corresponding to the reference system program fault knowledge point in the recursive neural network corresponding to each set program fault label.

[0075] Step S142, for each recursive neural network corresponding to the set program fault label, determine the adjacent vector data of the target network member under the set program fault label according to the connectivity vector between the target network member and other network members in the recursive neural network.

[0076] Step S143 , according to the heat map, aggregate the adjacent vector data of the target network members under each set program fault label to generate aggregated adjacent vector data of the reference system program fault knowledge point.

[0077] Step S144, updating the fault knowledge point vector data of the reference system program fault knowledge point according to the aggregated adjacent vector data, and generating target fault knowledge point vector data corresponding to the reference system program fault knowledge point.

[0078] In the fault diagnosis process of the training and assessment system, electronic equipment uses recursive neural networks (RNNs) to capture the complex correlations between the fault knowledge points of each reference system program under different fault labels. These RNN models are trained with historical data and expert knowledge, and can reflect the connectivity vectors and mutual influences between fault knowledge points.

[0079] In this embodiment, the RNN model corresponding to each set program fault label is composed of multiple network members, each of which represents a reference system program fault knowledge point. The network members are connected through links, and the weights on the links reflect the connectivity vectors between two adjacent network members. These connectivity vectors are automatically learned by the model during the training process, and they capture the potential associations and mutual influences between fault knowledge points.

[0080] The trained RNN model is stored in the local or remote database of the electronic device. When fault diagnosis is required, the electronic device will retrieve and load the corresponding RNN model based on the current target system program running data and fault labels.

[0081] For each reference system program fault knowledge point, the electronic device first determines its target network member in the RNN model corresponding to each set program fault label. This step is achieved by matching the unique identifier (such as ID or name) of the fault knowledge point with the network member in the RNN model.

[0082] Assuming there is a reference system program failure knowledge point named "memory leak", the electronic device will search for all RNN models corresponding to the set program failure labels and find the network member representing "memory leak" in each model.

[0083] For each RNN model corresponding to a set program fault label, the electronic device determines the adjacency vector data of the target network member under the fault label based on the connectivity vector between the target network member and other network members in the model. The adjacency vector data contains the weight information of other network members directly connected to the target network member, reflecting the strength of the association between them.

[0084] For example, in the RNN model corresponding to the fault label "system performance degradation", the electronic device will extract the connectivity vectors between the "memory leakage" network member and other network members such as "high CPU usage" and "disk I / O delay" to generate the adjacency vector data of "memory leakage".

[0085] Next, the electronic device performs weighted aggregation on the adjacent vector data under each set program fault label based on the thermal value information in the thermal map generated in step S120. The thermal value reflects the degree of match between the target system program running data and each fault label, so the adjacent vector data corresponding to the fault label with a higher matching degree has a greater weight in the aggregation process. For example, if the thermal map shows that "system performance degradation" has the highest matching degree with the current target system program running data, then the adjacent vector data of "memory leak" under the "system performance degradation" fault label will obtain a greater weight in the aggregation process.

[0086] Finally, the electronic device fuses the aggregated adjacency vector data with the original fault knowledge point vector data (such as weighted summation, splicing, etc.) to generate target fault knowledge point vector data corresponding to the reference system program fault knowledge point. These data contain richer graph structure features and associated information, which helps to locate the cause of the fault more accurately.

[0087] For example, the electronic device performs weighted summation on the original fault knowledge point vector data of "memory leakage" and the aggregated adjacent vector data to generate new target fault knowledge point vector data. The new vector data not only contains the information of "memory leakage" itself, but also incorporates the associated information of other related fault knowledge points.

[0088] Through the above steps, recursive neural network and graph structure feature extraction technology are used to conduct in-depth analysis of the reference system program fault knowledge points, and generate more accurate fault knowledge point vector data to provide support for subsequent fault diagnosis.

[0089] In a possible implementation, the thermal map includes thermal value information of the target system program running data matching each set program fault label.

[0090] Step S143 includes:

[0091] Step S1431, for each set program fault label, the target system program running data is matched with the thermal value information of the set program fault label and the adjacent vector data of the target network member under the set program fault label, and the initial aggregated adjacent vector of the reference system program fault knowledge point under the set program fault label is generated.

[0092] Step S1432 , determining aggregated adjacent vector data of the reference system program fault knowledge point according to the initial aggregated adjacent vectors of the reference system program fault knowledge point under each set program fault label.

[0093] In this embodiment, in the fault diagnosis process of the training and assessment system, the electronic device has generated a heat map of the target system program running data matching each set program fault label through classification layer processing, and the heat map contains the thermal value information between the target system program running data and each fault label, reflecting the degree of matching between them. Next, the electronic device will use this thermal value information, combined with the adjacency vector data of the target network member in the recursive neural network, to perform further aggregation operations.

[0094] In detail, the electronic device first reads the thermal value information in the thermal map, which indicates the matching degree between the target system program running data and each set program fault label, and the higher the value, the stronger the matching degree. For example, suppose there are three set program fault labels in the thermal map: "memory leak", "network delay" and "module abnormality", and their corresponding thermal values ​​are 0.8, 0.5 and 0.3 respectively, which indicates that the target system program running data has the highest matching degree with "memory leak", followed by "network delay" and finally "module abnormality".

[0095] For each programmed fault tag, the electronics does the following:

[0096] From the corresponding recursive neural network, the adjacency vector data between the target network member (i.e., the reference system program fault knowledge point currently being analyzed) and other network members are extracted.

[0097] The thermal value of the target system program running data matching the set program fault label is used as the weight to perform weighted processing on the adjacent vector data. In this way, the adjacent vector data corresponding to the fault label with a high matching degree will have a greater impact in the aggregation process.

[0098] The weighted adjacency vector data is the initial aggregated adjacency vector under the set program fault label.

[0099] For example, taking "memory leak" as an example, the electronic device extracts the adjacency vector data between the "memory leak" network member and other network members (such as "high CPU usage" and "disk I / O delay") from the recursive neural network corresponding to "memory leak". Then, these adjacency vector data are weighted using a heat value of 0.8 to generate the initial aggregated adjacency vector of "memory leak" under the "memory leak" fault label.

[0100] After the initial aggregated adjacency vectors under all set program fault labels are generated, the electronic device needs to further aggregate these initial vectors to generate the final aggregated adjacency vector data of the reference system program fault knowledge points. This process may involve multiple aggregation strategies, such as average pooling, maximum pooling, or weighted summation.

[0101] For example, the electronic device may choose weighted sum as the aggregation strategy, and the weights are still the thermal values ​​corresponding to each set program fault label. Therefore, the final aggregated neighbor vector of "memory leak" will be the weighted sum of its initial aggregated neighbor vectors under the three fault labels of "memory leak", "network delay" and "module abnormality", with weights of 0.8, 0.5 and 0.3 respectively.

[0102] The generated aggregated adjacency vector data will be used to update the fault knowledge point vector data of the reference system program fault knowledge point and generate target fault knowledge point vector data, which will contain richer graph structure features and associated information, and help to diagnose the cause of system failure more accurately.

[0103] For example, the electronic device fuses the aggregated adjacent vector data of "memory leakage" with its original fault knowledge point vector data (such as weighted summation, splicing, etc.) to generate the target fault knowledge point vector data corresponding to "memory leakage". The new vector data will serve as an important basis for the subsequent fault diagnosis process.

[0104] In a possible implementation, step S150 includes:

[0105] Step S151 : for each reference system program fault knowledge point, aggregate the fault knowledge point vector data of the reference system program fault knowledge point and the target fault knowledge point vector data to generate the target fault knowledge point vector data of the reference system program fault knowledge point.

[0106] Step S152, determining the estimated thermal value of each reference system program fault knowledge point matching the target system program running data based on the target fault knowledge point vector data of each reference system program fault knowledge point.

[0107] In this embodiment, in the fault diagnosis process of the training and assessment system, the electronic device has generated corresponding target fault knowledge point vector data for each reference system program fault knowledge point through graph structure feature extraction technology. These data not only contain the feature information of the fault knowledge point itself, but also incorporate the association information with other fault knowledge points, which helps to diagnose system faults more accurately. Next, the electronic device will use these target fault knowledge point vector data to calculate the estimated thermal value of each reference system program fault knowledge point matching the target system program running data.

[0108] In detail, for each reference system program fault knowledge point, the electronic device first aggregates its original fault knowledge point vector data with the target fault knowledge point vector data. The aggregation method can be weighted summation, splicing or other customized methods, depending on the design of the fault diagnosis network model and data characteristics.

[0109] For example, suppose there is a reference system program fault knowledge point "memory leak", whose original fault knowledge point vector data is V1, and the target fault knowledge point vector data generated by graph structure feature extraction is V2. The electronic device selects weighted summation as the aggregation method, and the weights α and (1-α) are assigned to V1 and V2 respectively (the specific value of α is determined according to model training). The aggregated target fault knowledge point vector data is V = α*V1 + (1-α)*V2.

[0110] After aggregating the target fault knowledge point vector data of each reference system program fault knowledge point, the electronic device will use these vector data to calculate their estimated thermal values ​​that match the target system program running data. The estimated thermal values ​​reflect the similarity or matching degree between the reference system program fault knowledge point and the target system program running data.

[0111] In detail, the electronic device can use cosine similarity, Euclidean distance, dot product operation or other similarity measurement methods to calculate the estimated thermal value. The specific method to be selected depends on the requirements of the fault diagnosis network model and the data characteristics.

[0112] For example, suppose that an electronic device uses cosine similarity to calculate the estimated thermal value. First, the electronic device also represents the target system program running data in vector form (if it has not yet been represented, it will be vectorized accordingly), recorded as D. Then, for each reference system program fault knowledge point (taking "memory leak" as an example), the cosine similarity between its target fault knowledge point vector data V and D is calculated as the estimated thermal value. The calculation formula of cosine similarity is:

[0113] Cosine Similarity=(V·D) / (||V||×||D||)

[0114] Wherein, (V·D) represents the dot product of V and D, and ||V|| and ||D|| respectively represent the modulus lengths of V and D. The calculated cosine similarity value Cosine Similarity is the estimated thermal value of the "memory leak" matching target system program running data.

[0115] Finally, the electronic device sorts the estimated thermal values ​​of each reference system program fault knowledge point and selects the knowledge points with the highest thermal values ​​as candidate fault knowledge points. If there are a large number of candidate knowledge points or further confirmation is required, the electronic device can conduct a comprehensive analysis and judgment based on other information (such as fault history records, expert opinions, etc.) to finally determine the target system program fault knowledge point corresponding to the target system program operation data.

[0116] For example, suppose that the electronic device calculates the estimated thermal values ​​of all reference system program fault knowledge points and sorts them in order from high to low. The top reference system program fault knowledge points include "memory leak", "high CPU usage" and "network delay". The electronic device uses these three knowledge points as candidate fault knowledge points and conducts further analysis in combination with fault history records and expert opinions. Finally, "memory leak" is determined to be the most likely cause of the fault, and it is output as a fault diagnosis result to relevant personnel or systems for processing and resolution.

[0117] In a possible implementation, respectively vectorizing the target system program operation data and each reference system program fault knowledge point to generate program operation vector data of the target system program operation data and fault knowledge point vector data of each reference system program fault knowledge point, includes: utilizing a fault diagnosis network model to respectively vectorize the target system program operation data and each reference system program fault knowledge point to generate program operation vector data of the target system program operation data and fault knowledge point vector data of each reference system program fault knowledge point.

[0118] In the fault diagnosis process of the training and assessment system in this embodiment, the electronic device needs to first convert the target system program running data and each reference system program fault knowledge point into a vector form for subsequent efficient matching and analysis. To achieve this goal, the electronic device uses a pre-trained fault diagnosis network model that can automatically map the input data into a high-dimensional vector space.

[0119] Before performing vectorized representation, the electronic device first ensures that the fault diagnosis network model has been loaded into the memory and is in an available state. The fault diagnosis network model can be a complex deep learning network that includes multiple types of network layers such as embedding layers, convolutional layers, and loop layers, which are used to extract features of input data and generate corresponding vector representations.

[0120] In detail, the electronic device collects the target system program running data in real time through the monitoring tools deployed at each key node of the system. These data may include CPU usage, memory usage, disk I / O, network request response time, log output of specific modules, etc. The collected raw data may contain noise or outliers, so the electronic device first cleans and preprocesses these data, such as removing data points beyond a reasonable range and interpolating missing values.

[0121] The preprocessed target system program running data is input into the fault diagnosis network model. The embedding layer or similar mechanism in the model automatically maps each feature to a vector in a high-dimensional space (i.e., the program running vector), which can capture the distribution characteristics and association relationships of the feature in the data set.

[0122] After being processed by the network model, the electronic device finally obtains program running vector data of the target system program running data. The program running vector data is a multidimensional array or tensor that contains key feature information of the target system program running status.

[0123] For each reference system program fault knowledge point, the electronic device first performs preprocessing operations such as word segmentation and stop word removal on its description text to extract key information. This step helps to remove noise in the text and retain information useful for fault diagnosis. The preprocessed fault knowledge point text is input into the corresponding part of the fault diagnosis network model (which can be another embedding layer or text processing module). The model uses word embedding technology (such as Word2Vec, GloVe, etc.) to map each word in the fault knowledge point to a vector in a high-dimensional space (i.e., the fault knowledge point representation vector). Then, the model may combine these word vectors into a single fault knowledge point vector data through aggregation operations (such as average pooling, maximum pooling, etc.).

[0124] In addition to the vector representation of the fault knowledge point itself, the electronic device may also consider the prior associated knowledge point vector data, which can be predetermined by historical data or expert knowledge and have some association or influence with the current fault knowledge point. The electronic device inputs these prior associated knowledge point vectors into the model together with the fault knowledge point vector (if the model supports it) to generate a richer set of target knowledge point representation vectors.

[0125] However, it should be noted that not all fault diagnosis network models directly consider prior knowledge points. In some cases, electronic devices may need to perform this step outside the model, that is, first generate a preliminary vector representation of the fault knowledge points, and then adjust or optimize them based on the prior knowledge points.

[0126] After being processed by the network model (which may include internal aggregation and external adjustment), the electronic device finally obtains the fault knowledge point vector data of each reference system program fault knowledge point. These fault knowledge point vector data are also a multidimensional array or tensor, which contains the key feature information of the fault knowledge point and the potential correlation with other fault knowledge points.

[0127] The generated program running vector data and fault knowledge point vector data will be used in subsequent steps such as heat map estimation, graph structure feature extraction and fault knowledge point matching. Through the processing and analysis of these steps, the electronic device can accurately diagnose the fault knowledge points corresponding to the program running data of the target system and provide processing suggestions for relevant personnel or systems.

[0128] In a possible implementation manner, the use of the fault diagnosis network model to vectorize the target system program operation data and each reference system program fault knowledge point, and before generating the program operation vector data of the target system program operation data and the fault knowledge point vector data of each reference system program fault knowledge point, further includes:

[0129] Step S101, obtaining a sample program operation learning data set and a recursive neural network corresponding to each set program fault label, wherein the sample program operation learning data set includes a plurality of template system program operation data and a priori system program fault knowledge points of the template system program operation data.

[0130] In this embodiment, before constructing a network model for training and assessment system fault diagnosis, the electronic device needs to collect and process a large amount of sample data in order to train a network model that can accurately identify system program operation failures. The process includes obtaining a sample program operation learning data set, building a recursive neural network, vectorizing the template system program operation data, estimating the thermal map, extracting graph structure features, calculating estimated thermal values, and finally using these data to train the network model.

[0131] In detail, the electronic device retrieves a sample program operation learning data set from a database or an external data source, and the sample program operation learning data set includes multiple template system program operation data and prior system program fault knowledge points corresponding to each template system program operation data. These template system program operation data represent various fault conditions that may occur in the system under different operating conditions.

[0132] For each set program fault label, the electronic device has pre-built and trained the corresponding recursive neural network (RNN) model. These RNN models can capture the connectivity vectors and association relationships between the reference system program fault knowledge points under the specific fault label.

[0133] Step S102, using a fault diagnosis network model, respectively vectorizes the template system program operation data and each reference system program fault knowledge point to generate template program operation vector data of the template system program operation data and template fault knowledge point vector data of each reference system program fault knowledge point.

[0134] In this embodiment, the electronic device uses an embedding layer or a similar mechanism in a fault diagnosis network model to vectorize the template system program running data. This process maps the features of each template system program running data to a vector in a high-dimensional space (template program running vector data) for subsequent analysis and processing.

[0135] Similarly, the electronic device vectorizes each reference system program fault knowledge point and generates template fault knowledge point vector data, which captures the key features and semantic information of the fault knowledge point.

[0136] Step S103 , based on the template program running vector data, estimating a heat map of the template system program running data matching each set program fault label.

[0137] In this embodiment, the electronic device uses a classification layer or similarity mechanism to process the template program running vector data and estimates the probability that they match each set program fault label. These probability values ​​are displayed in the form of a heat map, reflecting the degree of matching between the template system program running data and different fault labels.

[0138] Step S104, based on the heat map and the recursive neural network corresponding to each set program fault label, the template fault knowledge point vector data of each reference system program fault knowledge point is extracted with graph structure features to generate target template fault knowledge point vector data corresponding to each reference system program fault knowledge point.

[0139] For each reference system program fault knowledge point, the electronic device extracts the graph structure features of its template fault knowledge point vector data based on the thermal map and the corresponding recursive neural network. This process takes into account the connectivity vectors and association relationships between the fault knowledge points and generates the target template fault knowledge point vector data.

[0140] Step S105 , determining the estimated thermal value of each reference system program fault knowledge point matching the template system program running data based on the target template fault knowledge point vector data corresponding to each reference system program fault knowledge point.

[0141] The electronic device calculates an estimated thermal value of each reference system program fault knowledge point matching the template system program operation data, and the estimated thermal value reflects the similarity or matching degree between the fault knowledge point and the template system program operation data.

[0142] Step S106, training the fault diagnosis network model based on the estimated thermal value and the priori system program fault knowledge points of the template system program operation data to generate a fault diagnosis network model that has completed model learning.

[0143] The electronic device calculates the error information of the fault diagnosis network model based on the estimated thermal values ​​and the prior system program fault knowledge points of the template system program operation data. The error information may include two parts: fault knowledge point error and fault label error, which respectively reflect the accuracy of the model in identifying fault knowledge points and fault labels.

[0144] If the error information does not meet the set model convergence requirements, the electronic device will train and adjust the fault diagnosis network model based on the error information. This process may include updating the model's weights and bias parameters, adjusting the network structure, or introducing regularization terms to improve the model's generalization ability and accuracy.

[0145] During the training process, the electronic device may also adjust the recursive neural network corresponding to each set program fault label based on the target template fault knowledge point vector data and estimated thermal value. This process aims to optimize the connectivity vectors and association relationships between network members in order to more accurately reflect the interaction between fault knowledge points.

[0146] The electronic device repeatedly performs the above steps (vectorization, heat map estimation, graph structure feature extraction, estimated heat value calculation, and model training and adjustment) until the error information of the fault diagnosis network model meets the set convergence requirements. At this point, the electronic device obtains a fault diagnosis network model that has completed model learning and can accurately identify program operation faults in the training and assessment system.

[0147] The trained fault diagnosis network model will be used in actual application scenarios to perform real-time analysis and fault diagnosis on the target system program running data. Through the processing and analysis of steps such as vector representation, thermal map estimation, graph structure feature extraction, and estimated thermal value calculation, the electronic equipment can quickly locate and solve system fault problems to ensure the normal operation of the training and assessment system.

[0148] In a possible implementation, step S106 includes:

[0149] Step S1061, determining fault diagnosis error information of the fault diagnosis network model based on the estimated thermal value and the priori system program fault knowledge points of the template system program operation data.

[0150] In the process of building a complete network model for fault diagnosis in the training and assessment system, the electronic equipment needs to continuously optimize the performance of the model through iterative training. This process involves using the template system program running data and its prior fault knowledge points, combined with estimated thermal values ​​and recursive neural networks, to fine-tune the fault diagnosis network model.

[0151] In this embodiment, the electronic device first calculates the fault diagnosis error information of the fault diagnosis network model based on the estimated thermal value and the prior system program fault knowledge points of the template system program operation data. The fault diagnosis error information may include multiple aspects, such as the recognition accuracy of the fault knowledge points, the classification accuracy of the fault labels, etc.

[0152] For example, electronic equipment calculates the error of each sample by comparing the estimated thermal value output by the fault diagnosis network model with the actual fault knowledge point of the template system program running data. Then, the errors of all samples are summarized to obtain the overall fault diagnosis error rate.

[0153] Step S1062: When the fault diagnosis error information does not meet the set model convergence requirement, the fault diagnosis network model is trained according to the fault diagnosis error information.

[0154] In this embodiment, if the fault diagnosis error information does not meet the pre-set model convergence requirements (such as the error rate is higher than a certain threshold), the electronic device will train and adjust the fault diagnosis network model based on the error information. For example, the electronic device uses a back propagation algorithm to update the weights and bias parameters in the network model through methods such as gradient descent. In each iteration, the electronic device calculates the gradient of the loss function with respect to the model parameters and adjusts the parameter values ​​according to the gradient information to minimize the loss function (i.e., minimize the error).

[0155] Step S1063 , based on the heat map of the template system program running data matching each set program fault label, determine the target program fault label matched by the template system program running data from the each set program fault label.

[0156] In this embodiment, the electronic device matches the heat map of each set program fault label based on the template system program running data, and determines the target program fault label matched by each template system program running data from all the set program fault labels.

[0157] For example, the electronic device observes the thermal map, finds the fault label with the highest thermal value corresponding to each template system program running data, and uses it as the target program fault label of the template system program running data.

[0158] Step S1064 , clustering each template system program running data according to the target program fault label matched by each template system program running data, and generating a template system program running data sequence corresponding to each set program fault label.

[0159] In this embodiment, the electronic device clusters the template system program running data according to the target program fault labels matched by each template system program running data, and generates template system program running data sequences corresponding to each set program fault label.

[0160] For example, the electronic device groups all template system program running data matching the "memory leak" fault label into one group to form a template system program running data sequence corresponding to the "memory leak" fault label. Similarly, corresponding template system program running data sequences are also generated for other fault labels.

[0161] Step S1065, for each set program fault label, based on the prior system program fault knowledge point of the template system program operation data in the template system program operation data sequence corresponding to the set program fault label, determine the template system program operation data subsequence of each reference system program fault knowledge point in the template system program operation data sequence corresponding to the set program fault label, and the template system program operation data subsequence of each reference system program fault knowledge point includes the template system program operation data corresponding to the corresponding reference system program fault knowledge point.

[0162] For each set program fault label, the electronic device determines the template system program operation data subsequences of each reference system program fault knowledge point in the template system program operation data sequence corresponding to the set program fault label according to the prior system program fault knowledge point. For example, in the template system program operation data sequence corresponding to the "memory leak" fault label, the electronic device screens out the template system program operation data containing "high CPU occupancy rate" as a common fault point according to the prior fault knowledge point of each template system program operation data, and forms the template system program operation data subsequence of "high CPU occupancy rate".

[0163] Step S1066, for each reference system program fault knowledge point, count the first data scale of the template system program operation data in the template system program operation data subsequence of the reference system program fault knowledge point, determine the second data scale of the template system program operation data that simultaneously matches other reference system program fault knowledge points in the template system program operation data subsequence of the reference system program fault knowledge point, and determine, based on the first data scale and the second data scale, that when the template system program operation data matches the reference system program fault knowledge point under the set program fault label, the template system program operation data simultaneously matches the target thermal value information of other reference system program fault knowledge points, and the target thermal value information is used to reflect the connectivity vector between the reference system program fault knowledge point and the other reference system program fault knowledge points.

[0164] In this embodiment, for each reference system program fault knowledge point, the electronic device counts the first data scale in its template system program operation data subsequence (i.e., the total number of template system program operation data in the subsequence), and determines the second data scale of the template system program operation data that matches other reference system program fault knowledge points at the same time. Then, the target thermal value information is calculated based on the two data scales.

[0165] For example, in the "CPU occupancy rate high" template system program running data subsequence, the electronic device counts a total of 100 template system program running data (first data scale). Among them, 30 data are also marked as "disk I / O delay" (second data scale). The electronic device calculates a target thermal value based on these two data scales, which reflects the connectivity strength between "CPU occupancy rate high" and "disk I / O delay".

[0166] Step S1067, based on the target thermal value information, the connectivity vectors of the reference system program fault knowledge points in the recursive neural network corresponding to the set program fault label are adjusted to generate an adjusted recursive neural network corresponding to the set program fault label.

[0167] The electronic device adjusts the connectivity vectors of each reference system program fault knowledge point in the recursive neural network corresponding to the set program fault label according to the target thermal value information.

[0168] For example, the electronic device increases the connectivity weights of “high CPU usage” and “disk I / O latency” in the recursive neural network under the “memory leak” fault label to reflect the stronger correlation between them.

[0169] Step S1068, based on the adjusted fault diagnosis network model and the adjusted recursive neural network corresponding to each set program fault label, return to execute the step of using the fault diagnosis network model to vectorize the template system program running data and each reference system program fault knowledge point, until the fault diagnosis error information of the fault diagnosis network model meets the set model convergence requirements, and a fault diagnosis network model that completes model learning is generated.

[0170] Based on the adjusted fault diagnosis network model and recursive neural network, the electronic device returns to execute steps such as vector representation, heat map estimation, and graph structure feature extraction, and starts a new round of iterative training. This process will be repeated until the fault diagnosis error information of the fault diagnosis network model meets the set convergence requirements.

[0171] For example, the electronic device repeats all the above steps and recalculates the error information of the model after each iteration. As the iteration proceeds, the error rate of the model gradually decreases and tends to be stable. When the error rate is lower than the set threshold, the electronic device stops iterative training and outputs the fault diagnosis network model that has completed model learning.

[0172] In a possible implementation, step S101 includes:

[0173] Step S1011, obtain a sample program operation learning data set, and based on the prior system program fault knowledge points of the template system program operation data in the sample program operation learning data set, determine the template system program operation data sequence of each reference system program fault knowledge point in the sample program operation learning data set, the template system program operation data sequence of each reference system program fault knowledge point includes the template system program operation data of the corresponding reference system program fault knowledge point in the sample program operation learning data set.

[0174] In the preparation stage, electronic devices need to collect and organize a large amount of sample program operation learning data, which will be used to train and optimize the fault diagnosis network model. At the same time, electronic devices also need to build or adjust recursive neural networks (RNNs) based on these sample program operation learning data to better capture the complex correlation between the knowledge points of each reference system program fault under different fault labels.

[0175] In detail, the electronic device retrieves a sample program operation learning data set from a storage medium (such as a database, a file system or an external data source), and the sample program operation learning data set includes multiple template system program operation data, and each template system program operation data is accompanied by prior system program fault knowledge point information, that is, the fault type or fault point represented by the template system program operation data that has been confirmed by experts.

[0176] For example, the electronic device retrieves a set of 1,000 template system program running data from the database. Each template system program running data records various performance indicators (such as CPU usage, memory usage, etc.) of the system during a certain operation and the corresponding prior fault knowledge points (such as "memory leak", "network delay", etc.).

[0177] Step S1012: for each reference system program fault knowledge point, counting the first data size of the template system program running data in the template system program running data sequence of the reference system program fault knowledge point.

[0178] In this embodiment, the electronic device traverses the sample program operation learning data set, and based on the prior fault knowledge point information of each template system program operation data, classifies the template system program operation data with the same or similar fault knowledge points into the corresponding reference system program fault knowledge points, thereby forming a template system program operation data sequence for each reference system program fault knowledge point.

[0179] For example, the electronic device identifies all template system program operation data marked as "memory leak" and classifies them under the "memory leak" reference system program failure knowledge point, forming a sequence containing multiple template system program operation data. Similarly, the electronic device also creates corresponding template system program operation data sequences for other reference system program failure knowledge points such as "network delay" and "disk I / O error".

[0180] Step S1013, determining a second data scale of template system program operation data that simultaneously matches other reference system program failure knowledge points in the template system program operation data sequence of the reference system program failure knowledge point.

[0181] For each reference system program fault knowledge point, the electronic device counts the total number of template system program operation data in its template system program operation data sequence (first data scale), and the number of template system program operation data that simultaneously matches other reference system program fault knowledge points (second data scale).

[0182] For example, in the template system program running data sequence of "memory leakage", the electronic device counted a total of 200 template system program running data (first data scale). Further analysis found that 50 of them were also marked as "high CPU usage" (second data scale), which shows that there is a certain correlation between "memory leakage" and "high CPU usage".

[0183] Step S1014, based on the first data scale and the second data scale, determines that when the template system program operation data matches the reference system program fault knowledge point, the template system program operation data also matches the thermal value information of other reference system program fault knowledge points, and the thermal value information is used to reflect the connectivity vector between the reference system program fault knowledge point and the other reference system program fault knowledge points.

[0184] In this embodiment, the electronic device calculates a thermal value information based on the first data scale and the second data scale, which is used to reflect the degree to which the reference system program fault knowledge point matches other reference system program fault knowledge points when matching the template system program running data. The thermal value can be regarded as a quantitative representation of the connectivity strength between knowledge points.

[0185] For example, the electronic device uses a certain calculation formula (such as ratio, weighted sum, etc.) to calculate a thermal value based on the first data scale (200) and the second data scale (50) of "memory leakage", indicating the connectivity strength between "memory leakage" and "high CPU usage". Similarly, the electronic device also calculates corresponding thermal values ​​for all other possible fault knowledge point pairs.

[0186] Step S1015, according to the thermal value information, the connectivity vectors between the network members corresponding to the reference system program fault knowledge points in the recursive neural network of each setting program fault label are configured to generate the recursive neural network corresponding to each setting program fault label after configuration.

[0187] Finally, based on the calculated thermal value information, the electronic device configures the connectivity vectors between the network members corresponding to the reference system program fault knowledge points in the recursive neural network corresponding to each set program fault label. This process may involve adjusting the weights between network members, adding or deleting links, and other operations, so as to better reflect the actual correlation between fault knowledge points.

[0188] For example, the electronic device checks the recursive neural network corresponding to the "memory leak" fault label and finds the network members representing "memory leak" and "high CPU usage". Based on the thermal value calculated previously, the electronic device adjusts the connection weights between these two network members to more accurately reflect the actual connection strength between them. Similarly, the electronic device also makes corresponding configuration adjustments to the recursive neural networks under all other fault labels.

[0189] After the above steps, the electronic device not only obtains a rich set of sample program running learning data, but also configures a more accurate and efficient recursive neural network model based on these data, laying a solid foundation for subsequent fault diagnosis work.

[0190] In a possible implementation, step S1061 includes:

[0191] Step S1061 - 1 , calculating fault knowledge point error information based on the estimated thermal value and the a priori system program fault knowledge points of the template system program running data.

[0192] In this embodiment, during the process of training the fault diagnosis network model, the electronic device needs to continuously evaluate the performance of the model so as to adjust and optimize the model according to the evaluation results. The evaluation process involves calculating fault diagnosis error information, which reflects the accuracy of the model in identifying fault knowledge points and classifying fault labels.

[0193] In detail, the electronic device first calculates the fault knowledge point error information based on the estimated thermal value and the prior system program fault knowledge point of the template system program operation data, and the error information measures the accuracy of the model in identifying a single fault knowledge point.

[0194] For example, suppose there is a template system program running data, and its prior fault knowledge point is "memory leak". After the model performs a series of processing on the data, such as vector representation, thermal map estimation, and graph structure feature extraction, it gives a list of estimated thermal values, which lists the possibility that the data matches each reference system program fault knowledge point. The electronic device compares the thermal value corresponding to "memory leak" in the estimated thermal value list with a preset threshold. If the thermal value is lower than the threshold, it is considered that the model has failed to correctly identify the fault knowledge point, resulting in a certain error. By performing similar processing on all template system program running data, the electronic device can calculate the overall fault knowledge point error rate.

[0195] Step S1061 - 2 , matching the heat map of each set program fault label with the template system program running data, and calculating the fault label uncertainty of the template system program running data.

[0196] Next, the electronic device matches the heat map of each set program fault label based on the template system program operation data, and calculates the fault label uncertainty of each template system program operation data. The uncertainty reflects the model's hesitation in classifying the template system program operation data into a specific fault label.

[0197] For example, for the same template system program running data, its heat map may show heat values ​​associated with multiple set program fault labels, and the difference between these heat values ​​is not large, which indicates that the model has a certain uncertainty in classifying the data into a specific fault label. Electronic devices can quantify this uncertainty by calculating the entropy value of the heat map or other similar indicators.

[0198] Step S1061-3, based on the heat map of the template system program running data matching each set program fault label, determine the target program fault label matched by the template system program running data, and determine the template data scale information corresponding to each set program fault label based on the target program fault label matched by each template system program running data.

[0199] In this embodiment, based on the heat map, the electronic device determines a target program fault label for each template system program running data, that is, the fault label with the highest heat value. Then, the electronic device counts the template data scale information corresponding to each set program fault label, that is, how many template system program running data are assigned to each fault label.

[0200] For example, if there are 100 pieces of template system program running data, after the heat map analysis, 30 pieces of data are assigned to the "memory leak" fault label, 20 pieces of data are assigned to the "network delay" fault label, and the rest of the data are assigned to other fault labels. The electronic device will record this information for subsequent calculation of fault label error information.

[0201] Step S1061 - 4 , calculating fault label error information based on the fault label uncertainty of each template system program running data and the template data scale information corresponding to each setting program fault label.

[0202] In this embodiment, the electronic device combines the fault label uncertainty of the template system program running data and the template data scale information corresponding to each set program fault label to calculate the fault label error information. The fault label error information measures the accuracy of the model in correctly classifying the template system program running data into fault labels.

[0203] For example, the electronic device may consider those template system program operation data with higher uncertainty of fault labels and give them more weight. At the same time, the electronic device will also pay attention to those fault labels with larger scale of template system program operation data, because the classification error will have a greater impact on the overall performance. By comprehensively considering these factors, the electronic device can calculate the error rate of each fault label and further obtain the overall fault label error rate.

[0204] Step S1061 - 5 , determining the fault diagnosis error information of the fault diagnosis network model according to the fault knowledge point error information and the fault label error information.

[0205] Finally, the electronic equipment comprehensively determines the fault diagnosis error information of the fault diagnosis network model based on the fault knowledge point error information and the fault label error information. The error information comprehensively reflects the performance of the model in identifying fault knowledge points and classifying fault labels.

[0206] For example, electronic equipment will perform weighted averaging of the fault knowledge point error rate and the fault label error rate according to certain weights (the selection of weights may be based on the needs of the actual application scenario) to obtain the final fault diagnosis error rate. This error rate will be used as an important indicator to evaluate model performance and to guide subsequent model training and optimization work.

[0207] In a possible implementation, step S102 includes:

[0208] Step S1021, using the fault diagnosis network model, embedding the template system program running data to generate a template knowledge structure vector of the template system program running data.

[0209] Step S1022: using the fault diagnosis network model, embedding each reference system program fault knowledge point to generate a template fault knowledge point representation vector for each reference system program fault knowledge point. Based on the template fault knowledge point representation vector of each reference system program fault knowledge point and the learnable prior associated knowledge point vector data, a target knowledge point representation vector set is constructed.

[0210] Step S1023, assigning importance weights to the template knowledge structure vector and the target knowledge point representation vector set to generate template program operation vector data of the template system program operation data and template fault knowledge point vector data of each reference system program fault knowledge point.

[0211] Step S106 includes: based on the estimated thermal value and the prior system program fault knowledge points of the template system program running data, updating the weight and bias information in the fault diagnosis network model and the prior associated knowledge point vector data to generate a fault diagnosis network model that completes model learning.

[0212] In this embodiment, the electronic device uses the fault diagnosis network model to embed and represent the template system program running data to generate a template knowledge structure vector.

[0213] For example, an electronic device receives a batch of template system program operation data, which contains the system's operating indicators in different states, such as CPU usage, memory occupancy, etc. The electronic device inputs this data into the embedding layer in the fault diagnosis network model. The embedding layer maps the original operating data into a high-dimensional vector space through a series of nonlinear transformations to generate a template knowledge structure vector that captures the key features of the system's operating state.

[0214] The electronic device uses the fault diagnosis network model to embed the fault knowledge points of each reference system program and generate a template fault knowledge point representation vector. Then, the target knowledge point representation vector set is constructed by combining the learnable prior associated knowledge point vector data.

[0215] For example, an electronic device has a knowledge base containing multiple reference system program fault knowledge points, which describe various possible fault conditions of the system. For each reference system program fault knowledge point, the electronic device first uses an embedding layer to embed it and generate a template fault knowledge point representation vector, which captures the semantic features and key information of the fault knowledge point. Next, the electronic device considers the prior associations between knowledge points, which can be predetermined based on historical data or expert knowledge. The electronic device represents these prior associated knowledge points in vector form (if not already represented), and together with the template fault knowledge point representation vector, constructs a set of target knowledge point representation vectors.

[0216] The electronic device assigns importance weights to the template knowledge structure vector and the target knowledge point representation vector set to generate template program operation vector data of the template system program operation data and template fault knowledge point vector data of each reference system program fault knowledge point. For example, the electronic device assigns importance weights to each vector in the template knowledge structure vector and the target knowledge point representation vector set according to the design and training objectives of the model. This process may involve multiple iterations and optimizations to ensure that the weight assignment can accurately reflect the association and importance between the data. After assigning the weights, the electronic device fuses or splices the weighted vectors to generate the final template program operation vector data and template fault knowledge point vector data, which will be used in subsequent steps such as heat map estimation, graph structure feature extraction, and model training.

[0217] Based on the estimated thermal values ​​and the prior system program fault knowledge points of the template system program running data, the electronic device updates the weight and bias information in the fault diagnosis network model and the prior associated knowledge point vector data to generate a fault diagnosis network model that completes model learning. For example, based on the results of the error calculation, the electronic device determines the model parameters that need to be adjusted. These parameters include the weight and bias information in the network layers such as the embedding layer and the fully connected layer, as well as the prior associated knowledge point vector data. The electronic device uses optimization techniques such as the back propagation algorithm to iteratively update the model parameters according to the error gradient. In each iteration, the electronic device calculates the gradient of the loss function with respect to the model parameters, and adjusts the parameter values ​​according to the gradient information to minimize the loss function (i.e., minimize the error).

[0218] After multiple iterations, the performance of the model gradually improves and the error rate gradually decreases. When the error rate is lower than the preset threshold or meets other convergence conditions, the electronic device stops iterative training and outputs a fault diagnosis network model that has completed model learning. The fault diagnosis network model has good fault diagnosis capabilities and can be used in actual application scenarios.

[0219] Figure 2 The hardware structure of the electronic device 100 provided in the embodiment of the present application for implementing the above-mentioned data processing method applicable to the training and assessment system is shown. Figure 2 As shown, the electronic device 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0220] In one possible design, the electronic device 100 may be a single electronic device or a group of electronic devices. The electronic device group may be centralized or distributed (for example, the electronic device 100 may be a distributed system). In some embodiments, the electronic device 100 may be local or remote. For example, the electronic device 100 may access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the electronic device 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the electronic device 100 may be implemented on an electronic device. By way of example only, the electronic device may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc., or any combination thereof.

[0221] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions used by the electronic device 100 to execute or use to complete the exemplary methods described in this application.

[0222] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the data processing method applicable to the training and assessment system 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.

[0223] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned electronic device 100. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0224] In addition, an embodiment of the present application further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the data processing method applicable to the training and assessment system as described above is implemented.

[0225] It should be noted that in order to simplify the description disclosed in this application and thus help 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 and thus help 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 data processing method suitable for a training and assessment system, characterized in that: The method comprises: Obtain target system program operation data and multiple reference system program fault knowledge points of the training and assessment system, and respectively vectorize the target system program operation data and each reference system program fault knowledge point to generate program operation vector data of the target system program operation data and fault knowledge point vector data of each reference system program fault knowledge point; Based on the program running vector data, estimating a heat map of the target system program running data matching each set program fault label; Obtaining a recursive neural network corresponding to each set program fault label, wherein the recursive neural network is used to reflect the connectivity vectors between each reference system program fault knowledge point under the set program fault label; According to the heat map and the recursive neural network corresponding to each set program fault label, the fault knowledge point vector data of each reference system program fault knowledge point is extracted with graph structure features to generate target fault knowledge point vector data corresponding to each reference system program fault knowledge point; Determine, based on the target fault knowledge point vector data corresponding to each reference system program fault knowledge point, an estimated thermal value of each reference system program fault knowledge point matching the target system program running data; Based on the estimated thermal value, determining the target system program fault knowledge point corresponding to the target system program operation data from the respective reference system program fault knowledge points; The recursive neural network includes network members corresponding to each reference system program fault knowledge point, and links between network members, wherein the links are used to reflect the connectivity vector between two adjacent network members; The recursive neural network corresponding to the heat map and each set program fault label performs graph structure feature extraction on the fault knowledge point vector data of each reference system program fault knowledge point to generate target fault knowledge point vector data corresponding to each reference system program fault knowledge point, including: For each reference system program fault knowledge point, determining a target network member corresponding to the reference system program fault knowledge point in a recursive neural network corresponding to each set program fault label; For each recursive neural network corresponding to a set program fault label, determining the adjacency vector data of the target network member under the set program fault label according to the connectivity vector between the target network member and other network members in the recursive neural network; According to the heat map, the adjacent vector data of the target network members under each set program fault label are aggregated to generate aggregated adjacent vector data of the reference system program fault knowledge point; The fault knowledge point vector data of the reference system program fault knowledge point is updated according to the aggregated adjacent vector data to generate target fault knowledge point vector data corresponding to the reference system program fault knowledge point.

2. The data processing method applicable to the training and assessment system according to claim 1 is characterized in that: The step of respectively vectorizing the target system program running data and each reference system program fault knowledge point to generate program running vector data of the target system program running data and fault knowledge point vector data of each reference system program fault knowledge point includes: Embedding and representing the system program running features of the target system program running data in at least one feature space to generate a knowledge structure vector of the target system program running data in at least one feature space; Embedding each reference system program fault knowledge point to generate a fault knowledge point representation vector for each reference system program fault knowledge point; Construct a target knowledge point representation vector set based on the fault knowledge point representation vectors of each reference system program fault knowledge point and prior associated knowledge point vector data; Determine the knowledge structure vector in the target feature space to be currently weighted from the knowledge structure vector in the at least one feature space; Performing importance weight distribution on the knowledge structure vector under the target feature space and the target knowledge point representation vector set to generate importance weight distribution information of the target system program running data and importance weight distribution information of each reference system program fault knowledge point; Determine an iterative target knowledge point representation vector set according to the importance weight distribution information of the target system program running data and the importance weight distribution information of each reference system program fault knowledge point; Return to the step of determining the knowledge structure vector under the target feature space to be currently weighted from the knowledge structure vector under the at least one feature space until the knowledge structure vectors under all feature spaces have completed importance weight allocation, and obtain the program running vector data of the target system program running data and the fault knowledge point vector data of each reference system program fault knowledge point based on the current iteration target knowledge point representation vector set.

3. The data processing method applicable to the training and assessment system according to claim 1 is characterized in that: The thermal map includes thermal value information of the target system program running data matching each set program fault label; The step of aggregating the adjacent vector data of the target network members under each set program fault label according to the heat map to generate aggregated adjacent vector data of the reference system program fault knowledge point includes: For each set program fault label, aggregating the thermal value information of the target system program running data matching the set program fault label with the adjacent vector data of the target network member under the set program fault label to generate an initial aggregated adjacent vector of the reference system program fault knowledge point under the set program fault label; According to the initial aggregated adjacency vectors of the reference system program fault knowledge points under each set program fault label, aggregated adjacency vector data of the reference system program fault knowledge points are determined.

4. The data processing method applicable to the training and assessment system according to claim 1 is characterized in that: The step of determining the estimated thermal value of each reference system program fault knowledge point matching the target system program running data based on the target fault knowledge point vector data corresponding to each reference system program fault knowledge point includes: For each reference system program fault knowledge point, aggregating the fault knowledge point vector data of the reference system program fault knowledge point and the target fault knowledge point vector data to generate the target fault knowledge point vector data of the reference system program fault knowledge point; According to the target fault knowledge point vector data of each reference system program fault knowledge point, an estimated thermal value of each reference system program fault knowledge point matching the target system program operation data is determined.

5. The data processing method applicable to the training and assessment system according to claim 1 is characterized in that: The step of respectively vectorizing the target system program running data and each reference system program fault knowledge point to generate program running vector data of the target system program running data and fault knowledge point vector data of each reference system program fault knowledge point includes: The target system program running data and each reference system program fault knowledge point are respectively vectorized by using a fault diagnosis network model to generate program running vector data of the target system program running data and fault knowledge point vector data of each reference system program fault knowledge point.

6. The data processing method applicable to the training and assessment system according to claim 5 is characterized in that: Before the target system program running data and each reference system program fault knowledge point are respectively vectorized using the fault diagnosis network model to generate program running vector data of the target system program running data and fault knowledge point vector data of each reference system program fault knowledge point, the method further includes: Obtaining a sample program running learning data set and a recursive neural network corresponding to each set program fault label, wherein the sample program running learning data set includes a plurality of template system program running data and a priori system program fault knowledge points of the template system program running data; Using a fault diagnosis network model, respectively vectorizing the template system program operation data and each reference system program fault knowledge point to generate template program operation vector data of the template system program operation data and template fault knowledge point vector data of each reference system program fault knowledge point; Based on the template program running vector data, estimating a heat map of the template system program running data matching each set program fault label; According to the heat map and the recursive neural network corresponding to each set program fault label, the template fault knowledge point vector data of each reference system program fault knowledge point is extracted with graph structure features to generate target template fault knowledge point vector data corresponding to each reference system program fault knowledge point; Determine the estimated thermal value of each reference system program fault knowledge point matching the template system program running data according to the target template fault knowledge point vector data corresponding to each reference system program fault knowledge point; Based on the estimated thermal value and the priori system program fault knowledge points of the template system program running data, the fault diagnosis network model is trained to generate a fault diagnosis network model that completes model learning.

7. The data processing method applicable to the training and assessment system according to claim 6 is characterized in that: The method of training the fault diagnosis network model based on the estimated thermal value and the priori system program fault knowledge points of the template system program operation data to generate a fault diagnosis network model that completes model learning includes: Determining fault diagnosis error information of the fault diagnosis network model based on the estimated thermal value and a priori system program fault knowledge points of the template system program operation data; When the fault diagnosis error information does not meet the set model convergence requirement, training the fault diagnosis network model according to the fault diagnosis error information; Based on a heat map of matching the template system program running data with various set program fault labels, determining a target program fault label matched by the template system program running data from the various set program fault labels; Clustering each template system program running data according to the target program fault label matched by each template system program running data to generate a template system program running data sequence corresponding to each set program fault label; For each set program fault label, according to a priori system program fault knowledge point of the template system program operation data in the template system program operation data sequence corresponding to the set program fault label, in the template system program operation data sequence corresponding to the set program fault label, a template system program operation data subsequence of each reference system program fault knowledge point is determined, and the template system program operation data subsequence of each reference system program fault knowledge point includes the template system program operation data corresponding to the corresponding reference system program fault knowledge point; For each reference system program fault knowledge point, a first data scale of the template system program operation data in the template system program operation data subsequence of the reference system program fault knowledge point is counted, and in the template system program operation data subsequence of the reference system program fault knowledge point, a second data scale of the template system program operation data that simultaneously matches other reference system program fault knowledge points is determined, and based on the first data scale and the second data scale, when the template system program operation data matches the reference system program fault knowledge point under the set program fault label, the target thermal value information of the template system program operation data that simultaneously matches other reference system program fault knowledge points is determined, and the target thermal value information is used to reflect the connectivity vector between the reference system program fault knowledge point and the other reference system program fault knowledge points; According to the target thermal value information, the connectivity vectors of the reference system program fault knowledge points in the recursive neural network corresponding to the set program fault label are adjusted to generate an adjusted recursive neural network corresponding to the set program fault label; Based on the adjusted fault diagnosis network model and the adjusted recursive neural network corresponding to each set program fault label, return to execute the step of using the fault diagnosis network model to vectorize the template system program running data and each reference system program fault knowledge point, until the fault diagnosis error information of the fault diagnosis network model meets the set model convergence requirements, and generate a fault diagnosis network model that completes model learning.

8. The data processing method applicable to the training and assessment system according to claim 6 is characterized in that: The step of obtaining a sample program running learning data set and a recursive neural network corresponding to each set program fault label includes: Acquire a sample program operation learning data set, and based on the prior system program fault knowledge points of the template system program operation data in the sample program operation learning data set, determine the template system program operation data sequence of each reference system program fault knowledge point in the sample program operation learning data set, wherein the template system program operation data sequence of each reference system program fault knowledge point includes the template system program operation data of the corresponding reference system program fault knowledge point in the sample program operation learning data set; For each reference system program fault knowledge point, counting a first data size of the template system program running data in the template system program running data sequence of the reference system program fault knowledge point; Determining a second data scale of template system program operation data that simultaneously matches other reference system program failure knowledge points in the template system program operation data sequence of the reference system program failure knowledge point; According to the first data scale and the second data scale, determining that when the template system program operation data matches the reference system program fault knowledge point, the template system program operation data also matches the thermal value information of other reference system program fault knowledge points, and the thermal value information is used to reflect the connectivity vector between the reference system program fault knowledge point and the other reference system program fault knowledge points; According to the thermal value information, the connectivity vectors between network members corresponding to the reference system program fault knowledge points in the recursive neural network of each set program fault label are configured to generate a recursive neural network corresponding to each set program fault label after configuration; The determining of the fault diagnosis error information of the fault diagnosis network model based on the estimated thermal value and the priori system program fault knowledge points of the template system program operation data includes: Calculating fault knowledge point error information based on the estimated thermal value and the priori system program fault knowledge points of the template system program operation data; According to the heat map of matching each set program fault label with the template system program running data, calculate the fault label uncertainty of the template system program running data; Based on the heat map of the template system program running data matching each set program fault label, determine the target program fault label matched by the template system program running data, and determine the template data scale information corresponding to each set program fault label according to the target program fault label matched by each template system program running data; Calculate the fault label error information based on the uncertainty of the fault labels of each template system program running data and the template data scale information corresponding to each set program fault label; Determining fault diagnosis error information of the fault diagnosis network model according to the fault knowledge point error information and the fault label error information; The method of using the fault diagnosis network model to vectorize the template system program operation data and each reference system program fault knowledge point to generate template program operation vector data of the template system program operation data and template fault knowledge point vector data of each reference system program fault knowledge point includes: Using a fault diagnosis network model, embedding the template system program running data to generate a template knowledge structure vector of the template system program running data; By using the fault diagnosis network model, each reference system program fault knowledge point is embedded and represented respectively, and a template fault knowledge point representation vector of each reference system program fault knowledge point is generated; based on the template fault knowledge point representation vector of each reference system program fault knowledge point and the learnable prior associated knowledge point vector data, a target knowledge point representation vector set is constructed; The template knowledge structure vector and the target knowledge point representation vector set are assigned importance weights to generate template program operation vector data of the template system program operation data and template fault knowledge point vector data of each reference system program fault knowledge point; The method of training the fault diagnosis network model based on the estimated thermal value and the priori system program fault knowledge points of the template system program operation data to generate a fault diagnosis network model that completes model learning includes: Based on the estimated thermal value and the prior system program fault knowledge points of the template system program running data, the weight and bias information in the fault diagnosis network model and the prior associated knowledge point vector data are updated to generate a fault diagnosis network model that completes model learning.

9. An electronic device, characterized in that: The electronic device includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the data processing method applicable to the training and assessment system as described in any one of claims 1 to 8 above.