Mechanical fault diagnosis method based on GCN network and knowledge graph

Abstract

The invention discloses a mechanical fault diagnosis method based on a GCN network and a knowledge graph, and belongs to the field of mechanical equipment fault diagnosis. According to the method, information from different sensors and data sources can be effectively fused by constructing the knowledge graph and combining the GCN network, the problem that the fusion effect is poor when multi-source heterogeneous data is processed through a traditional method is solved, and therefore the accuracy of fault diagnosis is improved; and in combination with the knowledge graph, not only can a fault diagnosis result be given, but also fault reason analysis and maintenance suggestions can be provided, the interpretability of the diagnosis result is enhanced, and the defect that a big data algorithm lacks interpretation capability is overcome.

Description

Technical Field

[0001] The present invention relates to a mechanical fault diagnosis method based on a GCN network and a knowledge graph, and belongs to the field of mechanical equipment fault diagnosis. Background Art

[0002] With the advancement of intelligent manufacturing and Industry 4.0, mechanical equipment is increasingly used in modern industrial production. However, as the equipment runs longer, various faults inevitably occur, which not only reduces production efficiency but also may lead to serious safety hazards. In order to ensure the continuity and safety of production, fault diagnosis and predictive maintenance of mechanical equipment have become key research directions in the industrial field. Traditional mechanical fault diagnosis methods often rely on expert experience or algorithms based on signal processing. These methods are insufficient in diagnostic efficiency and accuracy when faced with complex working conditions and various fault types, and are difficult to meet the needs of modern industry.

[0003] In recent years, with the rise of big data technology, mechanical fault diagnosis algorithms based on big data have gradually become a research hotspot. These algorithms usually use advanced technologies such as deep learning to automatically extract complex fault modes from a large amount of equipment operation data to improve the accuracy of diagnosis. For example, algorithms such as convolutional neural networks (CNN) and long short-term memory networks (LSTM) have been widely used in mechanical fault diagnosis. Compared with traditional algorithms, algorithms based on big data can handle more complex fault modes, and to a certain extent reduce the dependence on artificial feature engineering, thereby improving the intelligent level of fault diagnosis.

[0004] Existing big data mechanical fault diagnosis algorithms still have many shortcomings. First, although these algorithms can learn complex patterns from massive data, they still face the problem of poor data fusion and inaccurate fault detection when processing multi-source heterogeneous data. Secondly, in actual industrial applications, big data algorithms are often limited by computing resources and it is difficult to achieve real-time fault diagnosis. In addition, these algorithms usually lack in-depth understanding and interpretation of fault mechanisms, which limits their application value in engineering practice. In contrast, although traditional mechanical fault diagnosis algorithms perform well in certain specific scenarios, they have poor adaptability to complex faults and are highly dependent on manual experience, making it difficult to comprehensively and accurately diagnose various types of faults in mechanical equipment.

[0005] Based on the above analysis, the existing traditional and big data mechanical fault diagnosis algorithms are difficult to fully meet the industrial site's demand for efficient, real-time, and accurate fault diagnosis. Therefore, there is an urgent need for an innovative diagnostic method that can comprehensively utilize multi-source heterogeneous data and domain knowledge, while improving diagnostic accuracy and having good real-time and interpretability. Summary of the invention

[0006] In order to solve the problem that the existing fault diagnosis methods cannot effectively utilize multi-source heterogeneous data, the present invention provides a mechanical fault diagnosis method based on a GCN network and a knowledge graph, the method comprising:

[0007] Step 1, acquiring and preprocessing data, wherein the data includes historical data and data in a public data set of mechanical failures, and the historical data is multi-source heterogeneous data;

[0008] Step 2: Build a knowledge graph based on the data obtained in step 1;

[0009] Step 3, constructing a GCN network model and training the GCN network model;

[0010] Step 4: collect multi-source heterogeneous data of mechanical equipment in real time, and use the GCN network model trained in step 3 to perform fault diagnosis;

[0011] Step 5: extract information from the fault diagnosis result obtained in step 4 to match it with the knowledge graph constructed in step 2;

[0012] Step 6: Determine the fault type of the corresponding mechanical equipment based on the matching result obtained in step 5.

[0013] Optionally, the multi-source heterogeneous data includes mechanical equipment information, corresponding sensor information, mechanical equipment operation data and corresponding fault information.

[0014] Optionally, the step 2 includes:

[0015] Step 2.1, construct knowledge graph nodes;

[0016] The device node, fault type node, sensor data node and fault cause node are constructed respectively. The device node constructs a node data for each device, and the node data contains information such as device model, device operation time and maintenance history; the fault type node constructs node data for different types of faults, and the node data of the fault type node contains the fault name, fault description and associated device; the sensor data node constructs corresponding node data for each sensor, including sensor type, sensor measurement data and data timestamp; the fault cause node constructs node data for different fault causes, including cause description and associated fault type;

[0017] Step 2.2, construct the knowledge graph relationship edge;

[0018] Construct knowledge graph relationship edges based on whether there is a relationship between nodes;

[0019] If there is a relationship edge between the device node and the fault type node, the relationship edge includes the fault occurrence time and fault severity;

[0020] If there is a relationship edge between the device node and the sensor data node, the relationship edge includes the data collection time and the sensor installation location information;

[0021] If there is a relationship edge between the fault type node and the sensor data node, the relationship edge includes a correlation score and a data timestamp;

[0022] If there is a relationship edge between the fault type node and the fault cause node, the relationship edge includes the fault impact range and the cause confirmation degree, where the fault impact range is a text attribute that describes the range of devices that may be affected by the fault cause, and the cause confirmation degree is a numerical attribute that indicates the strength of association between the cause and the fault.

[0023] Optionally, the GCN model constructed in step 3 is composed of multiple graph convolutional layers, activation functions, pooling layers, and fully connected layers; during the training process of the GCN model, a back propagation algorithm is used to update the weights of the model:

[0024] First, forward propagation is performed through the graph convolutional layer and the fully connected layer to generate the prediction result; then, the loss value is calculated using the loss function based on the difference between the prediction result and the true label; then backpropagation is performed to calculate the gradient of the loss function for each layer parameter, and the chain rule is used to propagate the gradient from the output layer to the input layer layer by layer; finally, the optimizer is used to update the model weights based on the calculated gradient.

[0025] Optionally, in step 5, the fault diagnosis results and the knowledge graph are matched based on graph query technology: the correlation between the feature vector output by the GCN model and the nodes in the knowledge graph is calculated through a similarity measurement algorithm, and the node with the highest correlation is selected to extract historical fault information, equipment status and maintenance records related to the diagnosis results.

[0026] Optionally, the method further includes:

[0027] The knowledge graph constructed based on the fault type information determined in step 6 and the corresponding multi-source heterogeneous data is updated.

[0028] Optionally, step 6 further includes: performing semantic analysis and syntactic parsing on the extracted information based on a pre-trained Transformer model to generate a natural language description of the fault cause, impact scope, and maintenance suggestion.

[0029] Optionally, the step 1 includes:

[0030] Step 1.1, obtain historical data and public data sets;

[0031] Step 1.2: Check whether the acquired historical data and public data sets are continuous. If data is missing, linear interpolation is used to fill in the missing data to ensure data continuity.

[0032] Step 1.3, normalize the data;

[0033] Step 1.4: Integrate the normalized data and mark the fault information.

[0034] Optionally, the sensor information includes vibration signals, temperature, and rotation speed.

[0035] The beneficial effects of the present invention are:

[0036] 1) Solved the problem of effective utilization of multi-source heterogeneous data: By constructing a knowledge graph and combining it with the GCN network, it can effectively fuse information from different sensors and data sources, solving the problem of poor fusion effect of traditional methods when processing multi-source heterogeneous data, thereby improving the accuracy of fault diagnosis.

[0037] 2) Real-time and high efficiency are achieved: Thanks to the efficient computing power of the GCN model, the system can quickly output fault diagnosis results while processing complex data, meeting the real-time requirements of industrial sites.

[0038] 3) Enhanced interpretation of diagnostic results: By combining knowledge graphs, the system can not only give fault diagnosis results, but also provide fault cause analysis and maintenance suggestions, which enhances the interpretability of diagnostic results and solves the shortcoming of big data algorithms that lack explanatory power.

[0039] 4) Possess adaptive capabilities: Through the self-learning module, the system can continuously optimize the diagnostic model according to the actual operating status of the equipment, improving the adaptability and stability of the system, and has stronger practicality and reliability compared to existing technologies.

[0040] In summary, the present invention provides an innovative solution for diagnosing mechanical equipment failures efficiently, accurately and in real time by combining the GCN network and the knowledge graph, which significantly improves the intelligence level of mechanical fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 It is a flow chart of data collection and preprocessing.

[0043] Figure 2 It is a schematic diagram of knowledge graph construction.

[0044] Figure 3 It is a schematic diagram of knowledge graph construction.

[0045] Figure 4 This is the GCN model structure diagram.

[0046] Figure 5 It is a diagram of the graph database structure.

[0047] Figure 6 It is a fault diagnosis and explanation workflow diagram. DETAILED DESCRIPTION

[0048] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0049] Embodiment 1:

[0050] This embodiment provides a mechanical fault diagnosis method based on a GCN network and a knowledge graph, and takes the application of fault diagnosis in a blast furnace gas top pressure recovery turbine (TRT) device as an example for explanation. The method includes:

[0051] Step 1: Data acquisition and preprocessing;

[0052] like Figure 1 As shown in Figure 1, the data acquisition and preprocessing process includes:

[0053] Step 1.1, obtain historical data, public data sets, and real-time collected data;

[0054] Step 1.2: Check whether the acquired historical data, public data sets, and real-time collected data are continuous. If data is missing, linear interpolation is used to fill it in to ensure data continuity.

[0055] Step 1.3, normalize the data;

[0056] Step 1.4, integrate the normalized data and mark the faults;

[0057] Each step is further explained as follows:

[0058] (1) Data source:

[0059] The solution provided in this embodiment involves three types of data, namely historical data, public data sets, and real-time collected data, which are introduced as follows:

[0060] ①Historical data: Contains the long-term operating data of mechanical equipment. These data are organized in chronological order and cover the normal operating status of the equipment and known fault events.

[0061] The processing of historical data includes fault event marking to provide reference in subsequent model training and knowledge graph construction.

[0062] ② Public data sets: Use mechanical equipment operation data sets from public channels, which usually contain a wide range of fault types and operating conditions, which helps the generalization ability of the model. Public data sets can supplement the failure modes and equipment types that may be missing in historical data.

[0063] The public data sets used in this embodiment include the Case Western Reserve University Bearing Data Center Bearing Data Set (CWRU), the SUDA test bench data set, and the Shandong University of Science and Technology (SDUST) data set;

[0064] The CWRU dataset includes four bearings in different bearing conditions, namely normal state, inner ring fault, outer ring fault and rolling element fault, as well as vibration signals collected at four different loads of 0hp, 1hp, 2hp and 3hp and four different speeds of 1730r / min, 1750rmin, 1772r / min and 1797r / min.

[0065] The SUDA test bench dataset includes bearing fault data files of various loads collected and constructed at a sampling frequency of 10kHz under four motor load conditions: OKN, 1KN, 2KN, and 3KN. It contains normal bearing data and faulty bearing data. The bearing fault types include three single faults (IF, OF, BF) and four compound faults (IO, IB, OB, IOB).

[0066] The SDUST dataset is a bearing fault dataset collected under different speed and load conditions. The vibration signal of the bearing seat is recorded using an LMS data acquisition instrument with a vibration sensor. The sampling frequency is 25.6kHz and the engine speed range is 500 to 3000r / min. The dataset is based on N205EU cylindrical roller bearings and contains 10 data types: normal (N), inner ring fault (I) 0.2mm, 0.4mm, 0.6mm, outer ring fault (O) 0.2mm, 0.4mm, 0.6mm, rolling element fault (R) 0.2mm, 0.4mm, 0.6mm. The data of each state is collected under six speed conditions under four different loads, which is suitable for bearing fault diagnosis research under multiple working conditions.

[0067] ③ Real-time data collection: Through sensors installed on mechanical equipment, the operation data of the equipment is collected in real time, including vibration signals, temperature, speed, etc. Real-time data is used to dynamically monitor the status of the equipment and perform fault diagnosis in combination with historical data and public data sets.

[0068] (2) Data preprocessing: Vibration signal data is collected in real time from the sensors of TRT rotating equipment. The data format is time series, and historical vibration data accumulated after a certain period of time is used as the training and verification data set for the model.

[0069] For missing values ​​in the data, linear interpolation is used to fill them in to ensure the continuity of the data. At the same time, the box plot method is used to detect and eliminate outliers to prevent extreme data from having an adverse effect on model training.

[0070] In order to improve the numerical stability of the data during training, the Min-Max normalization method is used to scale the vibration signal data to the [0,1] interval. The formula is as follows:

[0071]

[0072] Among them, x is the original data, x min and x max are the minimum and maximum values ​​of the vibration signal data respectively.

[0073] (3) Data integration and labeling:

[0074] Data integration: Unify historical data, public data sets, and real-time data to ensure data format and feature consistency. During the integration process, the timestamps, sensor types, and feature data of all data sources are standardized to facilitate subsequent model training and knowledge graph construction.

[0075] Fault event labeling: Label known fault events in historical data and public data sets to ensure that the time point and status data of each fault are accurate. These labeled data are combined with real-time collected data to dynamically update the knowledge graph.

[0076] Labeled data expansion: Combine the label information in the historical data to generate supervised data sets for model training. These data sets include equipment status data before and after the failure, and are supplemented with relevant cases from public data sets to provide accurate references for nodes and edges in the knowledge graph.

[0077] Step 2: Build a knowledge graph based on the data obtained in step 1;

[0078] The knowledge graph is a network structure that combines the historical fault data of mechanical equipment, public data set information, and real-time collected data according to certain semantic relationships. The graph includes multiple entity nodes such as equipment, fault type, sensor data, fault cause, and the relationship edges between these nodes, including:

[0079] (1) Constructing knowledge graph nodes:

[0080] like Figure 2 As shown, device nodes, fault type nodes, sensor data nodes and fault cause nodes are constructed respectively, wherein the device node constructs a node data for each device, and the node data includes information such as device model, device operating time and maintenance history; the fault type node constructs node data for different types of faults, and the node data of the fault type node includes the fault name, fault description and associated devices; the sensor data node constructs corresponding node data for each sensor, including the sensor type, sensor measurement data and data timestamp; the fault cause node constructs node data for different fault causes, including cause description and associated fault type.

[0081] (2) Constructing knowledge graph relationship edges:

[0082] like Figure 3 As shown, the knowledge graph relationship edge refers to whether there is a relationship between nodes, such as between device nodes and fault type nodes, between device nodes and sensor data nodes, between fault type nodes and sensor data nodes, and between fault type nodes and fault cause nodes.

[0083] If there is a relationship edge between the device node and the fault type node, the relationship edge includes the fault occurrence time and the fault severity; if there is a relationship edge between the device node and the sensor data node, the relationship edge includes the data collection time and the sensor installation location information; if there is a relationship edge between the fault type node and the sensor data node, the relationship edge includes the correlation score and the data timestamp, where the correlation score can be determined by the technician based on experience; if there is a relationship edge between the fault type node and the fault cause node, the relationship edge includes the fault impact range and the cause confirmation degree, where the fault impact range is a text attribute that describes the range of devices that may be affected by the fault cause, and the cause confirmation degree is a numerical attribute that indicates the strength of the association between the cause and the fault, ranging from 0 to 1, which can be determined by the technician based on experience.

[0084] (3) Update and maintain the constructed knowledge graph:

[0085] Dynamic update: Based on real-time collected data, the sensor data nodes in the knowledge graph are regularly updated and the relevance scores with fault types are recalculated. The occurrence of new faults will automatically add the corresponding fault type nodes and device-fault type relationship edges to the knowledge graph.

[0086] Self-learning mechanism: Through the online learning mechanism, historical data and real-time data are used to continuously optimize the weights of each node and relationship in the graph. The self-learning module can dynamically adjust the relevance score of the fault cause node based on new data to ensure the accuracy of the knowledge graph.

[0087] (4) Graph storage format: Neo4j graph database is used for storage, and graph data is regularly exported and backed up in the format of CSV files.

[0088] The knowledge graph constructed is as follows Figure 5 As shown, the green node is the device node, which is the center of the associated device information, sensor information, device fault information, fault solution and other contents.

[0089] Step 3: Design and training of GCN model;

[0090] The design and training process of the GCN model based on graph convolutional network is as follows:

[0091] (1) Model structure design and training parameter configuration

[0092] like Figure 4 As shown in the figure, the constructed GCN model consists of multiple graph convolution layers, activation functions, pooling layers, and fully connected layers. The specific parameter configuration is shown in the following table:

[0093] Table 1 GCN specific structure and parameters

[0094]

[0095] Among them, N is the number of nodes; F is the feature dimension of the node; H1 and H2 are the number of hidden units in the corresponding layer. After building the model, configure the basic parameters of the model training, as shown in Table 2 below:

[0096] Table 2 Training parameter configuration

[0097]

[0098] During the training process of the GCN model, the back-propagation algorithm is used to update the model weights. First, forward propagation is performed through the graph convolution layer and the fully connected layer to generate the prediction result. Next, the loss value is calculated using a loss function (such as cross entropy or mean square error) based on the difference between the prediction result and the true label. Then back-propagation is performed to calculate the gradient of the loss function for each layer parameter, and the chain rule is used to propagate the gradient from the output layer to the input layer layer by layer. Finally, the model weights are updated based on the calculated gradients through an optimizer (such as Adam), thereby gradually optimizing the model performance. The update formula is as follows:

[0099]

[0100] Where: θ is the model parameter (weight); α is the learning rate; is the gradient of the loss function J(θ) with respect to the parameter θ.

[0101] In the GCN model, as the number of network layers increases, gradient vanishing or gradient exploding may occur. To solve these problems, the present invention uses gradient clipping to limit the size of the gradient:

[0102]

[0103] in: is the L2 norm of the gradient; clip_value is the set gradient clipping threshold, usually set to 10.

[0104] (2) Model training process

[0105] Use the preprocessed data and the node and relationship edge attributes determined in step 2 to construct node features and adjacency matrices (i.e. process the data and connections into a format that the model can recognize and process). Then divide the dataset into training set, validation set, and test set in a ratio of 8:1:1.

[0106] The training set data is input into the GCN model, and the model parameters are updated through the back propagation algorithm. After each round of training, the validation set is used to evaluate the model, and the validation set loss is monitored to prevent overfitting. After the training is completed, the final model is evaluated using the test set, and the model's accuracy, precision, recall rate, F1 score and other indicators are recorded, and the optimal model is saved.

[0107] (3) Model optimization and tuning

[0108] Since the hyperparameters determined at the beginning are not necessarily the optimal parameters, the Bayesian optimization method is used to tune the hyperparameters to find the optimal learning rate, number of hidden layer units and other parameters.

[0109] In addition, the solution of the present invention also adds newly collected real-time data to the training set and regularly retrains the GCN model to ensure the diagnostic accuracy of the model when the device status changes.

[0110] Through the above design and training process, the GCN-based mechanical fault diagnosis model constructed by the present invention is efficient, accurate, and real-time, and through reasonable hyperparameter configuration and optimization strategy, the reliability and practicability of the model in industrial applications are guaranteed.

[0111] The mechanical fault diagnosis method based on GCN network and knowledge graph provided by the present invention combines the GCN model and natural language processing (NLP) technology to diagnose mechanical faults. In addition, the system analyzes the data collected by the equipment in real time, outputs fault diagnosis results and explanations that are easy for the staff to understand, and incorporates the confirmed information into the knowledge graph, thereby expanding and optimizing the fault diagnosis capability. The following is a further explanation:

[0112] (1) Data input and fault diagnosis

[0113] The system first preprocesses the multi-source heterogeneous data (such as vibration signals, temperature, speed, etc.) collected by the equipment in real time, including normalization, noise reduction, and feature extraction, in order to meet the input requirements of the GCN model. The dimension of the input data is N×F, where N represents the number of nodes and F represents the feature dimension of each node. Next, the data is processed through multiple graph convolution layers to extract high-order structural features in the equipment data. The output dimensions of each graph convolution layer are set to H1 and H2 respectively, and the activation function uses ReLU. Subsequently, the output of the graph convolution layer is mapped to the fault category or parameter space through a fully connected layer to perform the fault diagnosis classification task, and the model outputs the fault prediction through a linear activation function.

[0114] (2) Information extraction and natural language processing

[0115] After the fault diagnosis results are generated, the system will enter the information extraction stage. To ensure the accuracy of information extraction, the system uses graph-based query technology to match the fault diagnosis results with the knowledge graph. Specifically, the system analyzes the feature vectors output by the GCN model and uses similarity measurement algorithms (such as cosine similarity or Euclidean distance) to determine the correlation with the nodes in the knowledge graph, thereby extracting historical fault information, equipment status, and maintenance records related to the diagnosis results.

[0116] The extracted data will enter the natural language processing (NLP) module to generate natural language descriptions suitable for workers to understand. To ensure efficient processing and accurate generation, the NLP module selects the BERT pre-trained model based on the Transformer architecture. This model has been pre-trained on a large number of industrial documents to enable it to have professional language understanding and generation capabilities. The specific processing steps include:

[0117] Text generation: Based on the pre-trained Transformer model, the extracted information is semantically analyzed and parsed to generate a natural language description of the fault cause, impact scope, and maintenance recommendations.

[0118] Terminology adaptation: To ensure the professionalism and accuracy of the generated text, the system introduces a domain-specific vocabulary and further optimizes the generation results of the pre-trained model through fine-tuning methods to ensure that it conforms to the language habits and terminology of industrial equipment maintenance.

[0119] (3) Continuous learning

[0120] After the staff confirms that the diagnostic information is accurate, the system incorporates the information into the knowledge graph. The process of adding new nodes and edges follows a clear set of algorithms and standards to ensure the structural rationality and accuracy of the graph.

[0121] First, when identifying and adding nodes and edges, the system analyzes the current knowledge graph to determine whether the newly added information already has corresponding nodes or relationships in the existing graph. The judgment criteria include:

[0122] 1) Node similarity: If the similarity between the feature vectors of the new information and the existing nodes (calculated based on cosine similarity or Euclidean distance) is lower than a preset threshold (e.g., 0.7), a new node is added.

[0123] 2) Relationship derivation: When the correlation score between two nodes (calculated based on edge weight or information entropy) exceeds a certain threshold (such as 0.6), the system will add a new edge.

[0124] Then, if it is judged to be new information, it starts to build new nodes and edges in the knowledge graph. The system uses the K-means clustering algorithm to determine the category of the new node and other nodes in the graph. The weight of the new edge is dynamically adjusted according to the strength of information association. For existing nodes and edges, the system dynamically adjusts their weights based on the importance and frequency of new data. The PageRank algorithm weight update algorithm is used to recursively update according to the relative importance of nodes to ensure that the graph retains valid structural information while it continues to expand.

[0125] Finally, to prevent the graph from becoming redundant or complex as data grows, the system regularly performs consistency checks and structural optimization. Based on the redundancy algorithm, low-frequency or no longer valid nodes and edges are automatically removed to maintain the simplicity and efficiency of the knowledge graph. Through the above process, the system can not only continuously expand the content of the knowledge graph, but also continuously optimize the structure and diagnostic capabilities of the graph through a feedback mechanism, thereby improving the overall intelligence level of the mechanical fault diagnosis system.

[0126] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0127] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A mechanical fault diagnosis method based on GCN network and knowledge graph, characterized in that: The method comprises: Step 1, acquiring and preprocessing data, wherein the data includes historical data and data in a public data set of mechanical failures, and the historical data is multi-source heterogeneous data; Step 2: Build a knowledge graph based on the data obtained in step 1; Step 3, constructing a GCN network model and training the GCN network model; Step 4: collect multi-source heterogeneous data of mechanical equipment in real time, and use the GCN network model trained in step 3 to perform fault diagnosis; Step 5: extract information from the fault diagnosis result obtained in step 4 to match it with the knowledge graph constructed in step 2; Step 6: Determine the fault type of the corresponding mechanical equipment based on the matching result obtained in step 5.

2. The method according to claim 1, characterized in that The multi-source heterogeneous data includes mechanical equipment information, corresponding sensor information, mechanical equipment operation data and corresponding fault information.

3. The method according to claim 2, characterized in that The step 2 comprises: Step 2.1, construct knowledge graph nodes; The device node, fault type node, sensor data node and fault cause node are constructed respectively. The device node constructs a node data for each device, and the node data contains information such as device model, device operation time and maintenance history; the fault type node constructs node data for different types of faults, and the node data of the fault type node contains the fault name, fault description and associated device; the sensor data node constructs corresponding node data for each sensor, including sensor type, sensor measurement data and data timestamp; the fault cause node constructs node data for different fault causes, including cause description and associated fault type; Step 2.2, construct the knowledge graph relationship edge; Construct knowledge graph relationship edges based on whether there is a relationship between nodes; If there is a relationship edge between the device node and the fault type node, the relationship edge includes the fault occurrence time and fault severity; If there is a relationship edge between the device node and the sensor data node, the relationship edge includes the data collection time and the sensor installation location information; If there is a relationship edge between the fault type node and the sensor data node, the relationship edge includes a correlation score and a data timestamp; If there is a relationship edge between the fault type node and the fault cause node, the relationship edge includes the fault impact range and the cause confirmation degree, where the fault impact range is a text attribute that describes the range of devices that may be affected by the fault cause, and the cause confirmation degree is a numerical attribute that indicates the strength of association between the cause and the fault.

4. The method according to claim 3, characterized in that The GCN model constructed in step 3 is composed of multiple graph convolutional layers, activation functions, pooling layers, and fully connected layers. During the training process of the GCN model, the back propagation algorithm is used to update the weights of the model: First, forward propagation is performed through the graph convolutional layer and the fully connected layer to generate the prediction result; then, the loss value is calculated using the loss function based on the difference between the prediction result and the true label; then backpropagation is performed to calculate the gradient of the loss function for each layer parameter, and the chain rule is used to propagate the gradient from the output layer to the input layer layer by layer; finally, the optimizer is used to update the model weights based on the calculated gradient.

5. The method according to claim 4, characterized in that In step 5, the fault diagnosis results and the knowledge graph are matched based on graph query technology: the correlation between the feature vector output by the GCN model and the nodes in the knowledge graph is calculated through a similarity measurement algorithm, and the node with the highest correlation is selected to extract historical fault information, equipment status and maintenance records related to the diagnosis results.

6. The method according to claim 5, characterized in that The method further comprises: The knowledge graph constructed based on the fault type information determined in step 6 and the corresponding multi-source heterogeneous data is updated.

7. The method according to claim 6, characterized in that The step 6 also includes: based on the pre-trained Transformer model, performing semantic analysis and syntactic parsing on the extracted information to generate a natural language description of the fault cause, impact scope and maintenance suggestion.

8. The method according to claim 7, characterized in that The step 1 comprises: Step 1.1, obtain historical data and public data sets; Step 1.2: Check whether the acquired historical data and public data sets are continuous. If data is missing, linear interpolation is used to fill in the missing data to ensure data continuity. Step 1.3, normalize the data; Step 1.4: Integrate the normalized data and mark the fault information.

9. The method according to claim 8, characterized in that The sensor information includes vibration signals, temperature, and rotation speed.

10. Application of the method according to any one of claims 1 to 9 in TRT equipment fault diagnosis.

Images