A method for diagnosing wind turbine gearbox faults using graph convolutional neural networks

Abstract

The present invention discloses a method for diagnosing wind turbine gearbox faults using a graph convolutional neural network. This method belongs to the field of renewable energy equipment status monitoring and fault diagnosis. The method mainly constructs a gearbox system diagram by firstly combining the operating mechanism and data correlation, then inputs the data into the graph convolution model in the form of a graph to obtain the predicted value of each node in the graph. The component status is quantified based on the Euclidean distance between the predicted value and the measured value of the key parameter according to the fault mechanism, and finally the fault is warned by whether the calculated status indicator exceeds the threshold. The beneficial effect of the present invention is that it proposes a method for constructing gearbox system graph data, utilizes a graph convolutional neural network to mine the causal relationship between the data of different measuring points in the gearbox, extracts the spatial position information between the measuring points in the gearbox, improves the rationality of the prediction of related measuring points in the gearbox system, and makes the model well applicable to different fault types.

Description

Technical Field

[0001] The present invention relates to the field of power generation equipment status monitoring and fault diagnosis, and in particular to a method for diagnosing wind turbine gearbox faults using a graph convolutional neural network. Background Art

[0002] The speed-increasing gearbox is one of the most critical systems in a doubly-fed asynchronous wind turbine. Its use significantly reduces the size of the generator and inverter. However, due to the technical difficulties in gearbox design, manufacturing, and assembly, as well as the harsh operating environment, gearbox system failures are inevitable. A gearbox failure can lead to reduced unit efficiency, power transmission interruption, and even system downtime, resulting in significant economic losses. The Wind Energy Committee of the China Renewable Energy Society, in its survey of wind power equipment operating quality, found that the gearbox, generator, and blade systems had the highest failure rates. Unplanned gearbox downtime accounted for as high as 21%, and the resulting repair costs accounted for 18.61% of total maintenance costs. This demonstrates the economic value of providing early warning of gearbox system failures.

[0003] Traditionally, fault diagnosis of gearbox systems is mainly based on an additionally installed vibration condition monitoring system. The high data acquisition frequency and large data volume make real-time diagnosis difficult, and the system relies on certain expert knowledge, making it less applicable to components other than rotating parts.

[0004] Patent document CN115935737A discloses a method and system for offshore wind farm operation and maintenance based on digital twins, which belongs to the field of intelligent operation and maintenance technology of offshore wind farms. The signals required for establishing a digital twin system are collected at the data perception level, and data fusion is realized through edge computing. A digital twin model driven by both mechanism and data is established, and the signal activation model collected in the physical space is used to realize functions such as power prediction, state evaluation, fault warning, fault diagnosis and life prediction based on digital twins, and can realize adaptive correction of the digital twin model as the operating state of the equipment changes. Through the practical application of the digital twin system, this invention can effectively reduce on-site personnel, reduce maintenance time, reduce the average annual downtime, and reduce the failure rate of units. It can serve the safe, efficient and intelligent development and utilization of large-scale offshore wind power, and effectively improve the technical level of offshore wind power operation and maintenance.

[0005] Patent document CN115495505A discloses a method for association analysis based on new energy airport cluster data. When using input and output data sets to search for association rules, it is first necessary to preprocess the data set. After preprocessing, the new energy airport cluster data is subjected to association rule mechanism analysis, association rule mining, and association rule function analysis, and finally the relationship between the expected power of wind power and wind speed, the influence of dust factors on the output characteristics of photovoltaic arrays, the Apriori identification method for mining association rules of new energy data, and the association rule function between photovoltaic active output and light intensity are obtained. This invention is a method for association analysis based on new energy airport cluster data that can complete the repair of missing and bad data and the mining of association rule methods for airport cluster data, while performing reasonable dimensionality reduction processing on new energy airport cluster data, reducing computational complexity, and improving the efficiency of association rule search.

[0006] Currently, most grid-connected power generation units are equipped with SCADA systems for wind turbine operation monitoring and control. However, existing technologies often only extract information from the time domain, ignoring the relationship between various measurement points within the gearbox, and cannot fully extract information in the spatial domain. In addition, a single model can often only warn of one type of fault, making it less applicable. Summary of the Invention

[0007] To address these issues and the shortcomings of existing technologies, this paper proposes a method for diagnosing wind turbine gearbox faults using a graph convolutional neural network. This method uses a graph convolutional neural network to diagnose wind turbine gearbox faults. This method primarily uses the Euclidean distance between the model's predicted fault representation and the SCADA measurement data to provide fault warnings.

[0008] The present invention comprises the following steps:

[0009] A method for diagnosing wind turbine gearbox faults using a graph convolutional neural network, the method specifically comprising the following steps:

[0010] Step 1: Based on the physical structure and operating principle of the wind turbine, analyze the fault mechanism of typical faults and construct a gearbox system fault tree. Construct a preliminary graph structure based on the relationship between the measurement point location and energy transfer.

[0011] Step 2: Use random forest to calculate the importance of the operating parameters to the parameters in the gearbox system, and connect the two most important parameters on the preliminary graph structure. The direction of the edge is determined according to the direction of their energy transfer to obtain the complete graph structure of the gearbox system.

[0012] Step 3: Extract SCADA data from the wind turbine's healthy operation period; filter out data on abnormal power generation and other known faults, and clean the wind speed-power scatter plot using the quartile method;

[0013] Step 4: Extract the fan operating condition data and the corresponding gearbox measurement point data from the cleaned data, construct the input samples using the sliding window method, and output the predicted value of each node at the next time point. The input samples are input into the graph neural network in the form of a graph for training;

[0014] Step 5: Based on the fault tree, the Euclidean distance between the predicted value and the measured value of the fault characterization is used as the component status indicator. After smoothing using the moving average method, the threshold is calculated from the healthy period status indicator using the 3σ rule;

[0015] Step 6: The SCADA data to be tested is cleaned according to steps 3 and 4 above, abnormal working data and outlier data are deleted, and input samples are constructed. The cleaned data to be tested is put into the trained graph convolutional neural network prediction model to obtain the prediction results of each node.

[0016] Step 7: Query the gearbox system fault tree based on different components and select the corresponding fault representation. Calculate the Euclidean distance between the predicted result and the measured value of the representation parameter as the status indicator. Smooth the obtained status indicator and then compare it with the threshold to determine whether the gearbox is abnormal.

[0017] Furthermore, the gearbox system fault tree is based on the analysis of the gearbox lubrication oil circuit structure and its circulation process, and is composed of the changes in the values ​​of each measuring point on the oil circuit when a typical fault occurs.

[0018] Furthermore, the operating parameters related to the operation of the wind turbine that have no direct physical connection with the measuring points in the gearbox are added to the preliminary graph structure by calculating the importance between the operating parameters and the measuring points in the gearbox.

[0019] Furthermore, in step 2, random forest is used to calculate the importance, specifically, the importance between the working condition parameters and each measuring point in the gearbox is calculated based on out-of-bag detection.

[0020] Furthermore, the obtained importance is used to normalize the measurement points in each gearbox, and the operating parameters with greater importance are selected and added to the graph structure. Then, the direction of the edges between the nodes is determined according to the energy transfer relationship to obtain the complete graph structure of the gearbox system.

[0021] Furthermore, the quartile method in step 3 is used to clean the wind speed-power scatter plot data. The specific steps are as follows: divide the data into multiple small intervals according to the wind speed interval of each 0.1m / s, calculate the first quantile Q1 and the third quantile Q3 of the statistical power sequence for each wind speed interval, and calculate the quantile range I = Q3-Q1. According to the calculated Q1, Q3 and I, remove the data with power less than Q1-1.5I and greater than Q3+1.5I.

[0022] Furthermore, in step 4, the graph neural network structure consists of a graph convolution layer, a graph attention layer, and a fully connected layer. The graph convolution block is a stack of graph convolution layers and graph attention layers. The data is output by the graph convolution layer and the graph attention layer in turn. The calculation results of each layer are concatenated and then passed to the next layer. The last graph attention layer is followed by a fully connected layer. The fully connected block is a 5-layer linear layer. The input dimension of the first layer is the same as the dimension of the flattened output of the graph attention layer. The output of the last layer is 1-dimensional. The remaining intermediate layers are all 128-dimensional, and each layer is activated with a PReLU function.

[0023] Furthermore, the loss function formula of the graph neural network structure in step 4 is: Where y is the original measurement value of the node, is the model's predicted output value, and n is the number of samples.

[0024] Furthermore, in step 5, the Euclidean distance between the predicted value and the measured value of the fault characterization is used as the calculation formula of the component status index x: Where K is the number of fault representations.

[0025] Furthermore, the calculation formula for the state indicator threshold is: T = μ + 3σ, where μ and σ represent the mean and standard deviation of the state indicator after smoothing during the healthy operation period, respectively.

[0026] The beneficial effect of the present invention is that a method for constructing graph data of a gearbox system is proposed, and the causal relationship between the data of different measuring points in the gearbox is mined by using a graph convolutional neural network, and the spatial position information between the measuring points in the gearbox is extracted, thereby improving the rationality of the prediction of related measuring points in the gearbox system and making the model have good applicability in different fault types. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Flowchart showing a wind turbine gearbox system fault early warning method based on graph convolutional neural networks.

[0028] Figure 2 Represents the gearbox system fault tree.

[0029] Figure 3 Represents the graph structure that is initially constructed based on the physical structure inside the gearbox.

[0030] Figure 4 Represents the graph structure after adding operating condition parameters.

[0031] Figure 5 Represents the specific structure and parameters of the graph neural network.

[0032] Figure 6 The figure shows the trend of the predicted and measured values ​​of the gearbox oil pool temperature of unit 36 ​​in a certain wind farm.

[0033] Figure 7 The figure shows the trend of the predicted and measured values ​​of the gearbox inlet oil temperature of unit 36 ​​in a certain wind farm.

[0034] Figure 8 The figure shows the trend of the predicted and measured values ​​of the gearbox cooling water temperature of unit 36 ​​in a certain wind farm.

[0035] Figure 9 Indicates the changes in the health status indicators of the temperature control valve of unit 36 ​​in a certain wind farm.

[0036] Figure 10 Indicates the changes in the filter status indicators of unit 43 in a certain wind farm. DETAILED DESCRIPTION

[0037] Two preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings, so that the advantages and characteristics of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more specific definition of the protection scope of the present invention.

[0038] like Figure 1 Figure 2 shows a flow chart of a wind turbine gearbox system fault warning method based on a graph convolutional neural network. The data for Example 1 comes from SCADA data for wind turbine unit 36 ​​at a 2.0 MW wind farm, covering the period from April 2018 to January 2019. A maintenance work order indicated an overtemperature in the main shaft front bearing of unit 36's gearbox on January 12, 2019. On-site inspection determined the temperature control valve to be abnormal.

[0039] The data in Example 2 comes from the SCADA data of wind turbine unit 43 in an actual 2.0MW wind farm, from January 2020 to August 2020. The maintenance work order shows that on August 12, 2020, the working surface of the gear ring in the gear box of unit 43 was severely peeled off, causing the filter element to be blocked.

[0040] Example 1: The implementation process of the present invention for temperature control valve failure includes the following steps:

[0041] S1: Based on the physical structure and operating principle of the wind turbine, analyze the fault mechanism of typical faults and construct the gearbox system fault tree, such as Figure 2 As shown in the figure, according to the relationship between the measurement point position and energy transfer, the graph structure is initially constructed, as shown in the figure. Figure 3 As shown;

[0042] S2: Use random forest to calculate the importance of working condition parameters to each measuring point in the gearbox, and normalize the obtained importance results for each measuring point in the gearbox. The results are as follows:

[0043]

[0044] As shown in , the working condition parameters with importance greater than 0.2 are selected as adjacent nodes on the graph structure and added to the graph structure. Then, the direction of the edge between the nodes is determined according to the energy transfer relationship to obtain a complete graph structure, as shown in Figure 4 As shown;

[0045] S3: Extract SCADA data from the healthy operation period of wind turbines. The training data is from April 2018 to November 2018, including the following parameters:

[0046] 1) Corresponding measuring points of the gearbox system: main shaft front and rear bearing temperatures, high-speed shaft front and rear end temperatures, gearbox oil pool temperature, gearbox cooling water temperature, gearbox inlet oil temperature, and oil pressure before and after the filter element;

[0047] 2) Working condition related parameters: wind speed, wind direction, active power, pitch angle, rotor speed, generator speed and ambient temperature;

[0048] S4: Data cleaning:

[0049] 1) Filter out data of abnormal power generation, that is, data with active power, generator speed or wind speed of 0;

[0050] 2) Eliminate wind curtailment and power rationing data and select data with active power less than 0.9 times the rated power and a pitch angle greater than 2°;

[0051] 3) Eliminate data with other known faults;

[0052] 4) Use the quartile method to clean the wind speed-power scatter plot data. Divide the data into multiple small intervals at wind speed intervals of 0.1 m / s. For each wind speed interval, calculate the first quantile Q1 and the third quantile Q3 of the statistical power series, and calculate the quantile range I = Q3-Q1. Based on the calculated Q1, Q3, and I, remove data with power less than Q1-1.5I and greater than Q3+1.5I.

[0053] S5: After data cleaning, the input sample is constructed using the sliding window method with a sliding window length of 6. Each node feature is then vertically concatenated with an embedding vector of length 8 to obtain the node input feature. The embedding vector is concatenated to each node sliding window sequence and used to distinguish the spatial positions of different nodes. The embedding vector is obtained using the Embedding function in PyTorch. The input features of each node are horizontally concatenated to obtain the input sample. The model output value is the 10-minute value under each node parameter.

[0054] S6: Input samples are input into the graph neural network in the form of graphs for training. The model network structure is as follows: Figure 5 As shown, the loss function is the MSE mean square error loss function, and the formula is:

[0055] S7: According to the fault tree, the Euclidean distance between the predicted value and the measured value of the fault characterization is used as the component status indicator x:

[0056] Where K is the number of fault representations. First, the original state indicators are averaged and downsampled every hour, and then smoothed by moving average: T is 5 as the moving window length, x t is the original state indicator, z t is the smoothed state indicator. After smoothing, the threshold of the healthy period state indicator is calculated. The threshold calculation formula is: T = μ + 3σ, where μ and σ represent the mean and standard deviation of the smoothed state indicator z during the healthy operation period, respectively. The calculated threshold is 7.17.

[0057] S8: During the test, the test data is taken from April 2018 to February 2019. The parameters are selected as S3. The SCADA data to be tested is cleaned according to S4 and S5, and abnormal working data and outlier data are deleted to construct the input sample;

[0058] S9: During testing, the test data cleaned in S8 is put into the trained graph convolutional neural network prediction model to obtain the prediction results of each node;

[0059] S10: During the test, according to the temperature control valve fault described in the work order, the gearbox system fault tree is queried to select the fault characterization corresponding to the temperature control valve fault, and the predicted values ​​of the gearbox oil pool temperature, gearbox cooling water temperature and gearbox inlet oil temperature are selected. The prediction results are as follows: Figures 6 to 8 As shown, the Euclidean distance between the predicted result and the measured value is calculated as the state indicator;

[0060] S11: During the test, the obtained status indicator is smoothed in the same way as in S7, and then compared with the threshold value under the healthy state obtained in S7. Figure 9 This is the change in the health status indicators of the unit's temperature control valve. After December 2018, the status indicators obviously exceeded the threshold. The model issued an early warning and discovered the fault 40 days in advance.

[0061] Example 2: The implementation process of the present invention for filter element blockage failure includes the following steps:

[0062] S1: Based on the physical structure and operating principle of the wind turbine, analyze the fault mechanism of typical faults and construct the gearbox system fault tree, such as Figure 2 As shown in the figure, according to the relationship between the measurement point position and energy transfer, the graph structure is initially constructed, as shown in the figure. Figure 3 As shown;

[0063] S2: The importance of the operating parameters and the measuring points in the gearbox is calculated using the random forest method. The obtained importance results are normalized for each measuring point in the gearbox. The results are shown in Table 1. Table 1 shows the importance of the operating parameters to each measuring point in the gearbox after normalization.

[0064]

[0065] Table 1

[0066] Select the working condition parameters with importance greater than 0.2 as adjacent nodes on the graph structure and add them to the graph structure. Then determine the direction of the edge between the nodes according to the energy transfer relationship to obtain a complete graph structure, such as Figure 4 As shown;

[0067] S3: Extract SCADA data from the healthy operation period of wind turbines. The training data is from January 2020 to June 2020, including the following parameters:

[0068] 1) Corresponding measuring points of the gearbox system: main shaft front and rear bearing temperatures, high-speed shaft front and rear end temperatures, gearbox oil pool temperature, gearbox cooling water temperature, gearbox inlet oil temperature, and oil pressure before and after the filter element;

[0069] 2) Working condition related parameters: wind speed, wind direction, active power, pitch angle, rotor speed, generator speed and ambient temperature;

[0070] S4: Data cleaning:

[0071] 1) Filter out data that is not generating power normally, i.e. data where active power, generator speed or wind speed is zero;

[0072] 2) Eliminate wind curtailment and power rationing data and select data with active power less than 0.9 times the rated power and a pitch angle greater than 2°;

[0073] 3) Eliminate data with other known faults;

[0074] 4) Use the quartile method to clean the wind speed-power scatter plot data. Divide the data into multiple small intervals at wind speed intervals of 0.1 m / s. For each wind speed interval, calculate the first quantile Q1 and the third quantile Q3 of the statistical power series, and calculate the quantile range I = Q3-Q1. Based on the calculated Q1, Q3, and I, remove data with power less than Q1-1.5I and greater than Q3+1.5I.

[0075] S5: After data cleaning, the input sample is constructed using the sliding window method with a sliding window length of 6. Each node feature is then vertically concatenated with an embedding vector of length 8 to obtain the node input feature. The embedding vector is concatenated to each node sliding window sequence and used to distinguish the spatial positions of different nodes. The embedding vector is obtained using the Embedding function in PyTorch. The input features of each node are horizontally concatenated to obtain the input sample. The model output value is the 10-minute value under each node parameter.

[0076] S6: Input samples are input into the graph neural network in the form of graphs for training. The graph neural network structure is as follows: Figure 5 As shown, the loss function is the MSE mean square error loss function, and the formula is:

[0077] S7: According to the fault tree, the Euclidean distance between the predicted value and the measured value of the fault characterization is used as the component status indicator x: Where K is the number of fault representations. First, the original state indicators are averaged and downsampled every hour, and then smoothed by moving average: T is 5 as the moving window length, x t is the original state indicator, z t is the smoothed state indicator. After smoothing, the threshold of the healthy period state indicator is calculated. The threshold calculation formula is: T = μ + 3σ, where μ and σ represent the mean and standard deviation of the smoothed state indicator z during the healthy operation period, respectively. The calculated threshold is 0.73.

[0078] S8: During the test, the test data is taken from January to August 2020. The parameters are selected the same as S3. The SCADA data to be tested is cleaned according to S4 and S5, abnormal working data and outlier data are deleted, and the input sample is constructed;

[0079] S9: During testing, the test data cleaned in S8 is put into the trained graph convolutional neural network prediction model to obtain the prediction results of each node;

[0080] S10: During the test, based on the filter element fault described in the work order, the gearbox system fault tree is queried to select the fault representation corresponding to the filter element fault, the gearbox oil pump outlet pressure and the gearbox inlet pressure are selected, and the Euclidean distance between the predicted result and the measured value is calculated as the status indicator;

[0081] S11: During the test, the obtained status indicator is smoothed in the same way as in S7, and then compared with the threshold value under the healthy state obtained in S7. Figure 10 This shows the changes in the health status indicators of the filter element of unit 43. After July 10, 2020, the status indicators obviously exceeded the threshold. The model issued an early warning and discovered the fault 32 days in advance.

[0082] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for diagnosing wind turbine gearbox faults using a graph convolutional neural network, characterized in that: The method specifically comprises the following steps: Step 1: Based on the physical structure and operating principle of the wind turbine, analyze the fault mechanism of typical faults and construct a gearbox system fault tree. Construct a preliminary graph structure based on the relationship between the measurement point location and energy transfer. Step 2: Use random forest to calculate the importance of the operating parameters to the parameters in the gearbox system, and connect the two most important parameters on the preliminary graph structure. The direction of the edge is determined according to the direction of their energy transfer to obtain the complete graph structure of the gearbox system. Step 3: Extract SCADA data from the wind turbine's healthy operation period; filter out data on abnormal power generation and other known faults, and clean the wind speed-power scatter plot using the quartile method; Step 4: Extract the fan operating condition data and the corresponding gearbox measurement point data from the cleaned data, construct the input samples using the sliding window method, and output the predicted value of each node at the next time point. The input samples are input into the graph neural network in the form of a graph for training; Step 5: Based on the fault tree, the Euclidean distance between the predicted value and the measured value of the fault characterization is used as the component status indicator. After smoothing using the moving average method, the threshold is calculated from the component status indicator using the 3σ rule. Step 6: The SCADA data to be tested is cleaned according to steps 3 and 4 above, abnormal working data and outlier data are deleted, and input samples are constructed. The cleaned data to be tested is put into the trained graph convolutional neural network prediction model to obtain the prediction results of each node. Step 7: Query the gearbox system fault tree based on different components and select the corresponding fault representation. Calculate the Euclidean distance between the predicted result and the measured value of the representation parameter as the status indicator. Smooth the obtained status indicator and then compare it with the threshold to determine whether the gearbox is abnormal.

2. The method for diagnosing wind turbine gearbox faults using a graph convolutional neural network according to claim 1, characterized in that: The gearbox system fault tree is based on the analysis of the gearbox lubrication oil circuit structure and its circulation process, and is composed of the changes in the values ​​of each measuring point on the oil circuit when a typical fault occurs.

3. The method for diagnosing wind turbine gearbox faults using a graph convolutional neural network according to claim 1, characterized in that: The operating parameters related to the operation of the wind turbine that have no direct physical connection with the measuring points in the gearbox are added to the preliminary graph structure by calculating the importance between the operating parameters and the measuring points in the gearbox.

4. The method for diagnosing wind turbine gearbox faults using a graph convolutional neural network according to claim 1, characterized in that: In step 2, random forest is used to calculate the importance, specifically, the importance between the working condition parameters and each measuring point in the gearbox is calculated based on out-of-bag detection.

5. The method for diagnosing wind turbine gearbox faults using a graph convolutional neural network according to claim 3, characterized in that: The importance is used to normalize each measuring point in the gearbox, and the operating parameters with greater importance are selected and added to the graph structure. The direction of the edges between the nodes is determined according to the energy transfer relationship to obtain the complete graph structure of the gearbox system.

6. The method for diagnosing wind turbine gearbox faults using a graph convolutional neural network according to claim 1, characterized in that: In step 3, the quartile method is used to clean the wind speed-power scatter plot data. The specific steps are as follows: divide the data into multiple small intervals according to the wind speed interval of 0.1m / s. For each wind speed interval, calculate the first quantile Q1 and the third quantile Q3 of the statistical power sequence, and calculate the quantile range I = Q3-Q1. According to the calculated Q1, Q3 and I, remove the data with power less than Q1-1.5I and greater than Q3+1.5I.

7. The method for diagnosing wind turbine gearbox faults using a graph convolutional neural network according to claim 1, characterized in that: In step 4, the graph neural network structure consists of a graph convolution layer, a graph attention layer, and a fully connected layer. The graph convolution block is a stack of graph convolution layers and graph attention layers. The data is output by the graph convolution layer and the graph attention layer in turn. The calculation results of each layer are concatenated and then passed to the next layer. The last graph attention layer is followed by a fully connected layer. The fully connected block is a 5-layer linear layer. The input dimension of the first layer is the same as the dimension of the flattened output of the graph attention layer. The output of the last layer is 1-dimensional. The remaining intermediate layers are all 128-dimensional, and each layer is activated with the PReLU function.

8. The method for diagnosing wind turbine gearbox faults using a graph convolutional neural network according to claim 3, characterized in that: The loss function formula of the graph neural network structure in step 4 is: Where y is the original measurement value of the node, is the model's predicted output value, and n is the number of samples.

9. The method for diagnosing wind turbine gearbox faults using a graph convolutional neural network according to claim 1, characterized in that: In step 5, the Euclidean distance between the predicted value and the measured value of the fault characterization is used as the calculation formula of the component status index x: Where K is the number of fault representations.

10. The method for diagnosing wind turbine gearbox faults using a graph convolutional neural network according to claim 1, characterized in that: The calculation formula for the state indicator threshold is: T = μ + 3σ, where μ and σ represent the mean and standard deviation of the state indicator after smoothing during the healthy operation period, respectively.

Images