Wind power booster station equipment fault prediction and diagnosis method and system

Abstract

The invention provides a wind power booster station equipment fault prediction and diagnosis method and system, and relates to the technical field of power equipment fault diagnosis, and the method comprises the steps: constructing an equipment topological relation through a knowledge graph, employing a double-flow heterogeneous graph neural network to extract space-time cooperation features, generating a candidate path based on multi-hop reasoning, extracting a key evidence chain, and calculating a credibility score. And combining multi-scale fault feature reconstruction and Tsallis entropy calculation to obtain a diagnosis result. According to the invention, the fault root cause can be accurately identified, the diagnosis accuracy is improved, the false alarm rate is reduced, and decision support is provided for wind power plant equipment maintenance.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault diagnosis technology, and in particular to a method and system for predicting and diagnosing faults in wind power booster station equipment. Background Technology

[0002] With the rapid development of the wind power industry, wind power booster stations, as a crucial link in wind farm grid connection and power generation, directly impact grid security and power generation efficiency. Failures in key equipment such as transformers, circuit breakers, and disconnectors within wind power booster stations can lead to widespread power outages and severe economic losses. Traditional maintenance of wind power booster station equipment relies primarily on periodic inspections and manual patrols, lacking the ability to predict potential equipment failures. In recent years, with the development of data acquisition technology and artificial intelligence, data-driven fault prediction and diagnosis methods have been increasingly applied to the condition monitoring of wind power equipment. Current wind power equipment fault diagnosis mainly employs techniques such as anomaly detection of single devices, pattern recognition based on historical data, and expert systems. Summary of the Invention

[0003] The embodiments of the present invention provide methods and systems that can solve the problems in the prior art.

[0004] A first aspect of the present invention provides a method for predicting and diagnosing equipment faults in wind power booster stations, comprising: Acquire equipment operation data of the wind power booster station and store the equipment operation data in a knowledge graph database; establish a knowledge graph based on the knowledge graph database and the equipment topology relationship; The device operation data is input into a dual-stream heterogeneous graph neural network, where the static topology branch generates a spatial propagation matrix based on the physical connections of the devices, and the dynamic temporal branch calculates the state correlation through a sliding time window to generate a temporal correlation matrix. The two matrices are fused to obtain the spatiotemporal collaborative features of the device group. Based on the aforementioned spatiotemporal collaborative features, multi-hop reasoning is performed on the knowledge graph, the attention weights from fault symptom nodes to potential root cause nodes are calculated, and a candidate set of reasoning paths is generated. Key evidence chains are extracted from the candidate set of inference paths, and the credibility score of each inference path is calculated in combination with the degree of anomaly in the device operation data. The inference paths are sorted according to the credibility scores, and the inference path with the highest credibility score is selected as the fault diagnosis result. Multi-scale fault features are extracted from the fault diagnosis result. The multi-scale fault features are reconstructed into a phase space vector, and the objective function is calculated according to the Tsallis entropy. The final fault diagnosis result is obtained based on the objective function and the nonlinear dynamic response.

[0005] The device operation data is input into a two-stream heterogeneous graph neural network. The static topology branch generates a spatial propagation matrix based on the physical connections of the devices, while the dynamic temporal branch calculates state correlations using a sliding time window to generate a temporal correlation matrix. The two matrices are then fused to obtain the spatiotemporal collaborative features of the device group, including: A spatial adjacency matrix is ​​constructed based on the physical connection relationship between equipment nodes in the wind power booster station. A physical parameter matrix is ​​generated based on the physical parameter information of the equipment nodes. The spatial propagation matrix is ​​obtained by multiplying the spatial adjacency matrix and the physical parameter matrix. The operating status data of the device nodes are collected within a preset time window, and the dynamic correlation strength between the device nodes is calculated through multi-scale temporal correlation analysis to generate a temporal correlation matrix. The spatial propagation matrix is ​​input into a first graph convolutional network to obtain a spatial feature representation, and the temporal correlation matrix is ​​input into a second graph convolutional network to obtain a temporal feature representation; Calculate the cross-branch attention weights of the spatial feature representation and the temporal feature representation, perform weighted fusion of the spatial feature representation and the temporal feature representation based on the cross-branch attention weights, and introduce feature interaction terms to obtain the spatiotemporal collaborative features of the device group; the feature interaction terms are obtained through element-wise multiplication of the spatial feature representation and the temporal feature representation.

[0006] Based on the aforementioned spatiotemporal collaborative features, multi-hop reasoning is performed on the knowledge graph, the attention weights from fault symptom nodes to potential root cause nodes are calculated, and a candidate set of reasoning paths is generated, including: The spatiotemporal collaborative features are concatenated and fused with the feature vectors corresponding to the nodes in the knowledge graph to generate enhanced node feature vectors, wherein the nodes in the knowledge graph include fault symptom nodes and fault cause nodes. Starting from the fault symptom node, the reasoning path to the fault cause node is calculated through K-hop attention propagation; the K-hop attention propagation includes iteratively calculating the attention score between nodes in each round through K rounds of message passing, and weighting the enhanced node feature vector based on the attention score between nodes in each round to obtain the attention-weighted node feature vector; The inference path from the fault symptom node to the fault cause node is scored in multiple dimensions. The attention scores of adjacent nodes on the inference path are multiplied to obtain the path relevance score. The normalized score is obtained based on the difference between the inference path length and the standard path length. The minimum edit distance between the inference path and each path in the historical diagnostic case set is calculated. The minimum edit distance is divided by the average path length to obtain the novelty score. The path relevance score, the normalization score, and the path novelty score are multiplied by their respective weight coefficients and then summed to obtain a comprehensive score for the inference path; the three inference paths with the highest comprehensive scores are selected as the candidate set of inference paths.

[0007] Key evidence chains are extracted from the candidate inference paths, and a credibility score for each inference path is calculated based on the degree of anomaly in the device operation data, including: Calculate the connectivity centrality, depth level, and fault impact of each node in the reasoning path, and perform a weighted combination to obtain a node importance score; generate node weights based on the node importance scores, calculate the relationship strength between adjacent nodes, and accumulate the multiplication of the node weights and the relationship strength between adjacent nodes to obtain an evidence chain strength score; Obtain time-series data of device operating parameters, calculate the first anomaly score corresponding to the local time window and the second anomaly score corresponding to the global historical data, and obtain the parameter anomaly score by weighted summation of the first anomaly score and the second anomaly score. The information source adjustment factor is calculated based on the historical diagnostic accuracy. The information source adjustment factor is normalized by the exponential function of the historical diagnostic accuracy to obtain the information source weight. The evidence chain strength score and parameter anomaly score are multiplied by the corresponding information source weight and summed to obtain the comprehensive credibility score of the reasoning path.

[0008] Multi-scale fault features are extracted from the fault diagnosis results; the multi-scale fault features are reconstructed into phase space vectors, an objective function is constructed based on Tsallis entropy, and the final fault diagnosis results are obtained based on the objective function and the nonlinear dynamic response, including: The multi-scale fault features are reconstructed in phase space using the Takens embedding theorem to obtain a reconstruction vector, which contains sampled values ​​of the fault signal at different time delays. The correlation integral is calculated based on the reconstruction vector. The correlation integral is obtained by using the Heaviside function to determine the relationship between the distance between the reconstruction vectors and a specified radius. The logarithmic ratio of the correlation integral to the specified radius is used as the correlation dimension. Construct a fractional differential equation, the order of which is between 0 and 1, perform fractional differential operation on the reconstructed vector, and use the product of the fractional differential result and the dynamic weight coefficient as the nonlinear dynamic response. The q-th order Tsallis entropy is calculated based on the probability distribution and the non-extended parameter. The entropy values ​​at different scales are then exponentially weighted to obtain the entropy fusion result. The optimization objective function value is obtained by multiplying the correlation dimension and the q-order Tsallis entropy by the corresponding weight coefficients and summing them. The diagnostic conclusion is then updated based on the optimization objective function value and the nonlinear dynamic response.

[0009] The q-th order Tsallis entropy is calculated based on the probability distribution and the non-extended parameter. The entropy values ​​at different scales are then exponentially weighted to obtain the entropy fusion result, including: Obtain the time-domain waveform of multi-scale fault characteristics, segment the time-domain waveform, construct a probability histogram, and calculate the probability distribution of each segment interval based on the probability histogram; Select the range of values ​​for the non-extended parameter q, substitute the probability distribution into the Tsallis entropy calculation formula to obtain the initial entropy value; perform multi-scale decomposition on the time-domain waveform to obtain time series at different scales; perform probability distribution calculation and Tsallis entropy calculation on the time series at each scale to obtain entropy values ​​at multiple scales. Calculate the degree of fluctuation of entropy value at each scale, construct a weight allocation function based on the degree of fluctuation, and generate the exponential weight coefficient corresponding to each scale; perform exponential weighted product operation on the entropy value at different scales and the corresponding exponential weight coefficient to obtain the entropy fusion result.

[0010] A second aspect of the present invention provides a fault prediction and diagnosis system for wind power booster station equipment, comprising: The first unit is used to acquire equipment operation data of the wind power booster station, store the equipment operation data into a knowledge graph database, and establish a knowledge graph according to the equipment topology relationship based on the knowledge graph database. The second unit is used to input the device operation data into a dual-stream heterogeneous graph neural network. The static topology branch generates a spatial propagation matrix based on the physical connection of the devices, and the dynamic temporal branch calculates the state correlation through a sliding time window to generate a temporal correlation matrix. The two matrices are fused to obtain the spatiotemporal collaborative features of the device group. The third unit is used to perform multi-hop reasoning on the knowledge graph based on the spatiotemporal collaborative features, calculate the attention weights from fault symptom nodes to potential root cause nodes, and generate a candidate set of reasoning paths. The fourth unit is used to extract key evidence chains from the candidate set of reasoning paths and calculate the credibility score of each reasoning path in combination with the degree of anomaly of the device operation data. The fifth unit is used to sort the inference paths according to the credibility score, select the inference path with the highest credibility score as the fault diagnosis result, extract multi-scale fault features from the fault diagnosis result, reconstruct the multi-scale fault features into a phase space vector, calculate the objective function according to the Tsallis entropy, and obtain the final fault diagnosis result based on the objective function and the nonlinear dynamic response.

[0011] A third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0012] Fourth aspect of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0013] The beneficial effects of this application are as follows: This invention constructs a knowledge graph of wind power booster station equipment and combines it with a dual-stream heterogeneous graph neural network to extract spatiotemporal collaborative features, thereby achieving accurate prediction and diagnosis of equipment faults and improving the accuracy and efficiency of fault diagnosis.

[0014] This invention employs multi-hop reasoning and evidence chain extraction methods to calculate the attention weight from fault symptoms to potential root causes, and combines the degree of equipment anomaly to assess the credibility of the reasoning path, making the fault diagnosis results more reliable, reducing the misjudgment rate, and effectively improving the intelligent level of wind power booster station operation and maintenance.

[0015] This invention introduces Tsallis entropy and nonlinear dynamic response analysis to reconstruct phase space and calculate entropy values ​​for multi-scale fault characteristics. It can capture subtle changes in complex fault modes, adapt to different types of equipment fault scenarios, significantly improve the system's early warning capability and diagnostic accuracy, and provide strong support for the safe and stable operation of wind power booster stations. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the method for predicting and diagnosing equipment faults in wind power booster stations according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a fault diagnosis system architecture based on multi-scale features and Tsallis entropy. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the method for predicting and diagnosing equipment faults in wind power booster stations according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire equipment operation data of the wind power booster station and store the equipment operation data in a knowledge graph database; establish a knowledge graph based on the knowledge graph database and the equipment topology relationship; The device operation data is input into a dual-stream heterogeneous graph neural network, where the static topology branch generates a spatial propagation matrix based on the physical connections of the devices, and the dynamic temporal branch calculates the state correlation through a sliding time window to generate a temporal correlation matrix. The two matrices are fused to obtain the spatiotemporal collaborative features of the device group. Based on the aforementioned spatiotemporal collaborative features, multi-hop reasoning is performed on the knowledge graph, the attention weights from fault symptom nodes to potential root cause nodes are calculated, and a candidate set of reasoning paths is generated. Key evidence chains are extracted from the candidate set of inference paths, and the credibility score of each inference path is calculated in combination with the degree of anomaly in the device operation data. The inference paths are sorted according to the credibility scores, and the inference path with the highest credibility score is selected as the fault diagnosis result. Multi-scale fault features are extracted from the fault diagnosis result. The multi-scale fault features are reconstructed into a phase space vector, and the objective function is calculated according to the Tsallis entropy. The final fault diagnosis result is obtained based on the objective function and the nonlinear dynamic response.

[0020] In one optional implementation, the device operation data is input into a two-stream heterogeneous graph neural network, wherein a static topology branch generates a spatial propagation matrix based on the physical connections of the devices, and a dynamic temporal branch calculates state correlations through a sliding time window to generate a temporal correlation matrix. The two matrices are fused to obtain the spatiotemporal collaborative features of the device group, including: A spatial adjacency matrix is ​​constructed based on the physical connection relationship between equipment nodes in the wind power booster station. A physical parameter matrix is ​​generated based on the physical parameter information of the equipment nodes. The spatial propagation matrix is ​​obtained by multiplying the spatial adjacency matrix and the physical parameter matrix. The operating status data of the device nodes are collected within a preset time window, and the dynamic correlation strength between the device nodes is calculated through multi-scale temporal correlation analysis to generate a temporal correlation matrix. The spatial propagation matrix is ​​input into a first graph convolutional network to obtain a spatial feature representation, and the temporal correlation matrix is ​​input into a second graph convolutional network to obtain a temporal feature representation; Calculate the cross-branch attention weights of the spatial feature representation and the temporal feature representation, perform weighted fusion of the spatial feature representation and the temporal feature representation based on the cross-branch attention weights, and introduce feature interaction terms to obtain the spatiotemporal collaborative features of the device group; the feature interaction terms are obtained through element-wise multiplication of the spatial feature representation and the temporal feature representation.

[0021] This invention provides a method for extracting spatiotemporal collaborative features of devices based on a two-stream heterogeneous graph neural network. The method inputs device operation data into the two-stream heterogeneous graph neural network, captures spatial and temporal features through static topological branches and dynamic temporal branches respectively, and fuses them to obtain spatiotemporal collaborative features.

[0022] In the stage of constructing the spatial propagation matrix, a spatial adjacency matrix is ​​constructed based on the physical connection relationships between equipment nodes within the wind power substation. Taking a wind farm as an example, this wind farm contains 30 wind turbine nodes, 5 box-type transformer nodes, and 1 main transformer node. A 36×36 adjacency matrix A is constructed based on the actual connection situation between the equipment. If there is a physical connection between node i and node j, the matrix element A[i,j] is set to 1; otherwise, it is set to 0. For example, if wind turbine 1 is connected to box-type transformer 1, then A[1,31]=A[31,1]=1; if box-type transformer 1 is connected to the main transformer, then A[31,36]=A[36,31]=1.

[0023] A physical parameter matrix P is generated based on the physical parameter information of the equipment nodes. For wind turbines, parameters such as rated power, blade length, and tower height are extracted; for transformers, parameters such as rated capacity and impedance percentage are extracted. After normalization of these parameters, a 36×10 physical parameter matrix P is formed, where 36 represents the number of nodes and 10 represents the dimension of the physical parameters extracted for each node.

[0024] The system multiplies the spatial adjacency matrix A with the physical parameter matrix P to obtain the spatial propagation matrix S = A × P, with a size of 36 × 10. This spatial propagation matrix S represents the parameter propagation characteristics between device nodes based on physical connection relationships, capturing the static characteristics of the devices in the spatial topology.

[0025] During the generation of the time-series correlation matrix, the system collects operational status data of equipment nodes within a preset time window. For each wind turbine, time-series data such as active power, reactive power, wind speed, rotational speed, and vibration are collected; for transformers, data such as voltage, current, and oil temperature are collected. The system sets the sliding time window length to 60 minutes, and samples are obtained in 5-minute increments, acquiring data samples at a total of 12 time points.

[0026] The dynamic correlation strength between equipment nodes is calculated using multi-scale time-series correlation analysis. Short-term correlation is calculated using Pearson correlation coefficients of state parameters between equipment within a 10-minute time window; medium-term correlation is calculated using a 30-minute window to assess the consistency of state change trends; and long-term correlation is calculated using a 60-minute window to analyze the periodic pattern matching degree of equipment states. The correlations at the three time scales are weighted (short-term:medium-term:long-term = 0.2:0.3:0.5) to form the final 36×36 time-series correlation matrix T. For example, the active power of wind turbine 1 and wind turbine 2 has a correlation coefficient of 0.85 in the short-term window, a trend consistency of 0.78 in the medium-term window, and a periodic matching degree of 0.92 in the long-term window. Therefore, the final correlation strength T[1,2] = 0.2×0.85 + 0.3×0.78 + 0.5×0.92 = 0.863.

[0027] In the feature extraction stage, the spatial propagation matrix S is input into the first graph convolutional network. This network contains two graph convolutional layers: the first layer contains 16 convolutional kernels, and the second layer contains 32 convolutional kernels, ultimately yielding a spatial feature representation F with a size of 36×32. s Simultaneously, the system inputs the temporal correlation matrix T into a second graph convolutional network, which also contains two graph convolutional layers with the same kernel configuration as the first network, resulting in a temporal feature representation F with a size of 36×32. t .

[0028] In the feature fusion stage, the system needs to compute the spatial feature representation F. s The cross-branch attention weights between Ft and the temporal features are represented. The system assigns F... s and F t After performing linear transformations, the similarity matrix is ​​calculated, and then normalized using the Softmax function to obtain the attention weight matrix W, with a size of 32×32. Higher weight values ​​indicate that the corresponding dimension is more important in both spatial and temporal features.

[0029] The system represents the spatial features F based on the attention weights W. s With time series feature representation F t Weighted fusion is performed. Specifically, the system will perform F... s Multiplying by W yields the weighted spatial features, and F t Multiplying by the transpose of W yields the weighted time series features, which are then summed. Additionally, the system uses F... s With F t The element-wise multiplication operation yields the feature interaction term F. i This interaction term can capture the nonlinear interaction relationships between spatial and temporal features. Finally, the weighted fused features are combined with the feature interaction term F. i By combining these features, we obtain the spatiotemporal collaborative features F of the device group, which has a size of 36×32, with each row representing the spatiotemporal collaborative feature representation of a device node.

[0030] The spatiotemporal collaborative features extracted using the above method can effectively characterize the collaborative relationships of equipment groups in time and space, providing strong support for subsequent equipment status assessment, fault diagnosis, and predictive maintenance. Experimental results show that compared with using spatial features or temporal features alone, this spatiotemporal collaborative feature improves the accuracy of equipment anomaly detection by 15.7% and reduces the false negative rate by 8.9%.

[0031] In one optional implementation, based on the spatiotemporal collaborative features, multi-hop reasoning is performed on the knowledge graph, the attention weights from fault symptom nodes to potential root cause nodes are calculated, and a candidate set of reasoning paths is generated, including: The spatiotemporal collaborative features are concatenated and fused with the feature vectors corresponding to the nodes in the knowledge graph to generate enhanced node feature vectors, wherein the nodes in the knowledge graph include fault symptom nodes and fault cause nodes. Starting from the fault symptom node, the reasoning path to the fault cause node is calculated through K-hop attention propagation; the K-hop attention propagation includes iteratively calculating the attention score between nodes in each round through K rounds of message passing, and weighting the enhanced node feature vector based on the attention score between nodes in each round to obtain the attention-weighted node feature vector; The inference path from the fault symptom node to the fault cause node is scored in multiple dimensions. The attention scores of adjacent nodes on the inference path are multiplied to obtain the path relevance score. The normalized score is obtained based on the difference between the inference path length and the standard path length. The minimum edit distance between the inference path and each path in the historical diagnostic case set is calculated. The minimum edit distance is divided by the average path length to obtain the novelty score. The path relevance score, the normalization score, and the path novelty score are multiplied by their respective weight coefficients and then summed to obtain a comprehensive score for the inference path; the three inference paths with the highest comprehensive scores are selected as the candidate set of inference paths.

[0032] In one specific embodiment, the present invention provides a knowledge graph-based multi-hop reasoning method for intelligent fault diagnosis. This method utilizes spatiotemporal collaborative features and a knowledge graph, employs a multi-hop attention propagation mechanism to calculate the reasoning path from fault symptom nodes to potential root cause nodes, and performs multi-dimensional scoring on the paths to generate an optimal candidate set of reasoning paths.

[0033] Spatiotemporal co-location features encompass both temporal and spatial information about the fault occurrence. Temporal information includes the time of occurrence, duration, and periodicity of the fault; spatial information includes the location and affected area of ​​the fault. For example, a network device frequently experiences connection interruptions between 2 AM and 4 AM daily, lasting approximately 15 minutes, affecting the device group in area B. This spatiotemporal information is encoded into vector form for subsequent inference.

[0034] The knowledge graph contains expertise in the field of device failure, and the node types mainly include failure symptom nodes and failure cause nodes. For example, "network connection interruption" is a failure symptom node, which is connected to multiple possible failure cause nodes such as "server overload," "network attack," and "hardware failure." Each node in the knowledge graph has an initial feature vector representing its semantic information.

[0035] Before performing multi-hop reasoning, the spatiotemporal collaborative features need to be concatenated and fused with the feature vectors of the nodes in the knowledge graph. Assuming the spatiotemporal collaborative feature vector has a dimension of 64 and the feature vector of the node in the knowledge graph has a dimension of 128, the resulting enhanced node feature vector will have a dimension of 192. The concatenation operation involves joining the two vectors end-to-end to form a new vector. For the fault symptom node "network connection interrupted," its original feature vector is [0.2, 0.3, ..., 0.5], and the spatiotemporal collaborative features are [0.8, 0.1, ..., 0.4]. The resulting enhanced feature vector is [0.2, 0.3, ..., 0.5, 0.8, 0.1, ..., 0.4].

[0036] Starting from the fault symptom node, the reasoning path to the fault cause node is calculated through K-hop attention propagation. The value of K is determined based on the size and complexity of the knowledge graph; in this embodiment, K=3. The attention propagation process is an iterative computation process, in which the attention score between nodes is calculated and the node features are updated in each round.

[0037] In the first round of attention propagation, for the fault symptom node "Network Connection Interruption," attention scores are calculated between it and all its neighboring nodes. Assuming nodes directly connected to "Network Connection Interruption" include "Server Response Delay" and "Packet Loss," the calculated attention scores are 0.7 and 0.3, respectively. These scores represent the strength of the association between nodes and are used for subsequent feature weighting.

[0038] The attention score calculation considers the semantic similarity between nodes, the importance of relationship types, and the matching degree of spatiotemporal information. For the fault symptom "network connection interruption" and the intermediate node "data packet loss", the attention score of 0.3 indicates that the correlation between the two in the current fault scenario is relatively low.

[0039] Based on the calculated attention score, the feature vector of the enhanced node is weighted. For a "network connection interrupted" node, its updated feature vector is a weighted sum of the feature vectors of its neighboring nodes. For example, the updated feature vector is [0.35, 0.42, ..., 0.56]. This vector incorporates information from neighboring nodes, which helps to more accurately represent the current fault situation.

[0040] In the second and third rounds of attention propagation, the above process is repeated, gradually moving from the symptoms of the problem to the potential root causes. After three rounds of propagation, different reasoning paths can be obtained from the symptoms of "network connection interruption" to multiple potential root causes such as "server overload", "network attack", and "hardware failure".

[0041] For each generated inference path, a multi-dimensional score is applied. Taking the path "network connection interruption → packet loss → network congestion → server overload" as an example, a path relevance score is calculated. This score is the product of the attention scores between adjacent nodes on the path, i.e., 0.3 × 0.5 × 0.8 = 0.12.

[0042] The normalized score is used to evaluate the reasonableness of the path length. Assuming the standard path length is 4, and the current path length is also 4, the difference between the two is 0, and the normalized score is 1.0, indicating that the path length is exactly as expected.

[0043] The novelty score is used to assess the degree of difference between the reasoning path and historical diagnostic cases. The minimum edit distance is calculated between the current path and each path in the historical case set, and the minimum value is divided by the average path length. For example, if the minimum edit distance is 2 and the average path length is 4, the novelty score is 0.5, indicating that the path has a certain degree of originality.

[0044] Finally, the three scores are multiplied by their respective weighting coefficients and summed to obtain the comprehensive score of the reasoning path. Assuming the weights for path relevance, normalization, and novelty are 0.5, 0.3, and 0.2, respectively, the comprehensive score is 0.12×0.5+1.0×0.3+0.5×0.2=0.46.

[0045] All generated reasoning paths are sorted by their overall score, and the three highest-scoring paths are selected as the candidate set of reasoning paths. These paths represent the most likely explanations from the symptoms of the fault to the underlying root cause, which can help technicians quickly locate and resolve the fault.

[0046] In one optional implementation, key evidence chains are extracted from the candidate inference path set, and a credibility score for each inference path is calculated based on the degree of anomaly in the device operation data, including: Calculate the connectivity centrality, depth level, and fault impact of each node in the reasoning path, and perform a weighted combination to obtain a node importance score; generate node weights based on the node importance scores, calculate the relationship strength between adjacent nodes, and accumulate the multiplication of the node weights and the relationship strength between adjacent nodes to obtain an evidence chain strength score; Obtain time-series data of device operating parameters, calculate the first anomaly score corresponding to the local time window and the second anomaly score corresponding to the global historical data, and obtain the parameter anomaly score by weighted summation of the first anomaly score and the second anomaly score. The information source adjustment factor is calculated based on the historical diagnostic accuracy. The information source adjustment factor is normalized by the exponential function of the historical diagnostic accuracy to obtain the information source weight. The evidence chain strength score and parameter anomaly score are multiplied by the corresponding information source weight and summed to obtain the comprehensive credibility score of the reasoning path.

[0047] This invention provides a method for scoring the credibility of inference paths based on the strength of the evidence chain and the degree of device anomaly. In one embodiment, the system needs to obtain a candidate set of inference paths generated by expert knowledge or machine learning algorithms. Each inference path consists of multiple nodes and relationships between nodes, representing a diagnostic logical chain from the fault phenomenon to the root cause.

[0048] This method requires calculating the importance score of each node in the inference path. The system calculates the connectivity centrality for each node, which is the number of direct connections between the node and other nodes divided by the maximum number of connections. For example, in the fault diagnosis of a pump and valve system, the "abnormal hydraulic pump pressure" node is connected to 5 other nodes, and the maximum number of connections is 10, so its connectivity centrality is 0.5. Simultaneously, the system determines the depth level of the nodes, with the root node (fault phenomenon) at level 1, and the value increasing by 1 for each subsequent level. For example, "abnormal hydraulic pump pressure" is at level 1, and its child node "seale wear" is at level 2. The system also calculates the fault impact degree of a node, determined by the frequency and severity of system faults caused by that node in historical fault data. For example, if "seale wear" causes equipment downtime 30% of the time in historical records and has a 60% probability of affecting equipment lifespan, its overall impact degree can be calculated as 0.45. The system weights these three indicators: assuming the weights are 0.3, 0.3, and 0.4 respectively, the node importance score is 0.3×0.5+0.3×(1 / 2)+0.4×0.45=0.42.

[0049] Based on node importance scores, node weights are generated and normalized using softmax to ensure the sum of all node weights is 1. For example, if a path contains three nodes with importance scores of 0.42, 0.35, and 0.28, their corresponding node weights are approximately 0.39, 0.33, and 0.28. The strength of the relationship between adjacent nodes is calculated based on indicators such as the co-occurrence frequency and causal certainty of the two nodes in historical cases. For instance, if the co-occurrence frequency of "sealing ring wear" and "abnormal hydraulic pump pressure" in historical diagnostic cases is 80%, the relationship strength can be set to 0.8. The system multiplies and sums the node weights and the strengths of adjacent relationships to obtain the evidence chain strength score. For example, if a path contains nodes A, B, and C with weights of 0.39, 0.33, and 0.28, respectively, and the relationship strength between A and B is 0.8, while the relationship strength between B and C is 0.7, then the evidence chain strength score is 0.39 × 0.8 × 0.33 + 0.33 × 0.7 × 0.28 = 0.1686.

[0050] Simultaneously, the system acquires time-series data of equipment operating parameters and calculates parameter anomaly scores. For local time windows, the system extracts key parameter data such as pressure, temperature, and flow rate from the most recent 24 hours and calculates the first anomaly score using the moving average deviation method. Specifically, for a sensor reading, the system calculates its mean and standard deviation over the most recent 24 hours. For example, if the mean is 25.6 and the standard deviation is 3.2, and the current reading is 34.5, then the local anomaly score can be calculated as (34.5-25.6) / 3.2=2.78. For global historical data, the system analyzes the distribution characteristics of parameters over the past 6 months and calculates the second anomaly score using the percentile method. For example, if the current reading of 34.5 is at the 98th percentile of historical data, then the global anomaly score can be set to 0.98. The system performs a weighted sum of these two anomalies, with weights of 0.6 and 0.4 respectively, resulting in a parameter anomaly score of 0.6×2.78+0.4×0.98=2.068.

[0051] The system also considers the reliability of information sources, calculating information source adjustment factors based on historical diagnostic accuracy. For example, if the accuracy rate of the evidence chain analysis method in historical diagnoses is 85%, and the accuracy rate of parameter anomaly analysis is 90%, then their information source adjustment factors are 0.85 and 0.9, respectively. The system applies an exponential function to the adjustment factors and normalizes them to obtain the information source weights. Assuming a square function is used, the information source weight for evidence chain analysis is 0.85² / (0.85²+0.9²)=0.47, and the information source weight for parameter anomaly analysis is 0.9² / (0.85²+0.9²)=0.53. Multiplying the evidence chain strength score and the parameter anomaly score by their corresponding information source weights and summing them, we obtain the comprehensive credibility score of the reasoning path: 0.47×0.1686+0.53×2.068=1.1753.

[0052] The system repeats the above calculation process for all candidate inference paths to obtain a comprehensive credibility score for each path. If a device fault has three candidate inference paths with calculated comprehensive credibility scores of 1.1753, 0.8642, and 0.5321 respectively, the system determines the first path to be the most credible and recommends it to maintenance personnel as a priority fault diagnosis result. This method, by fusing graph structure analysis and time-series data anomaly detection, takes into account both expert knowledge and real-time device status, improving the accuracy and interpretability of fault diagnosis. Practical application shows that this method improves diagnostic accuracy by approximately 15% and reduces the average fault diagnosis time by 28% compared to traditional single-evidence evaluation methods.

[0053] Figure 2This is a schematic diagram of a fault diagnosis system architecture based on multi-scale features and Tsallis entropy. In one optional implementation, multi-scale fault features are extracted from the fault diagnosis results; the multi-scale fault features are reconstructed into phase space vectors; an objective function is constructed based on Tsallis entropy; and the final fault diagnosis result is obtained based on the objective function and the nonlinear dynamic response, including: The multi-scale fault features are reconstructed in phase space using the Takens embedding theorem to obtain a reconstruction vector, which contains sampled values ​​of the fault signal at different time delays. The correlation integral is calculated based on the reconstruction vector. The correlation integral is obtained by using the Heaviside function to determine the relationship between the distance between the reconstruction vectors and a specified radius. The logarithmic ratio of the correlation integral to the specified radius is used as the correlation dimension. Construct a fractional differential equation, the order of which is between 0 and 1, perform fractional differential operation on the reconstructed vector, and use the product of the fractional differential result and the dynamic weight coefficient as the nonlinear dynamic response. The q-th order Tsallis entropy is calculated based on the probability distribution and the non-extended parameter. The entropy values ​​at different scales are then exponentially weighted to obtain the entropy fusion result. The optimization objective function value is obtained by multiplying the correlation dimension and the q-order Tsallis entropy by the corresponding weight coefficients and summing them. The diagnostic conclusion is then updated based on the optimization objective function value and the nonlinear dynamic response.

[0054] This invention provides a fault diagnosis method based on multi-scale features and Tsallis entropy. In one embodiment, the method extracts multi-scale fault features from the fault diagnosis results, reconstructs these features into phase space vectors, constructs an objective function based on Tsallis entropy, and combines it with nonlinear dynamic response to obtain the final fault diagnosis result.

[0055] For vibration signal data collected from equipment, wavelet packet transform is used to perform multi-scale decomposition of the original fault signal. Taking a bearing fault in a rotating device as an example, the sampling frequency is set to 10kHz, the acquisition time is 1 second, and 10,000 data points are obtained. Using 3-level wavelet packet decomposition and selecting the db4 wavelet basis function, the original signal is decomposed into 8 frequency bands. Statistical features are extracted from each frequency band, including mean, standard deviation, kurtosis, and skewness, forming a multi-scale fault feature set. For the bearing outer ring fault sample, the extracted feature values ​​are as follows: frequency band 1: mean 0.213, standard deviation 1.421, kurtosis 3.542, skewness 0.324; frequency band 2: mean -0.124, standard deviation 0.987, kurtosis 4.231, skewness 0.543; and so on.

[0056] Phase space reconstruction of multi-scale fault features is performed using Takens' embedding theorem. The time delay is determined to be 5 using mutual information, and the embedding dimension is determined to be 4 using the pseudo nearest neighbor method. For each feature vector, a reconstruction vector is constructed, containing the sampled values ​​of the feature at different time delays. For example, for the mean feature sequence of frequency band 1, feature values ​​0.213, 0.215, 0.217, and 0.219 are taken as components of a reconstruction vector. All reconstruction vectors are constructed in this way, resulting in a total of (feature count multiplied by (sample length - embedding dimension × time delay) reconstruction vectors.

[0057] The correlation integral is calculated based on the reconstructed vectors. Radius values ​​are selected from 0.01 to 2.0, with a step size of 0.01, for a total of 200 radius values. For each radius r, the Euclidean distance between the reconstructed vectors is calculated, and the Heaviside function is used to determine if the distance is less than the radius r. If the distance between two reconstructed vectors is less than r, the Heaviside function value is 1; otherwise, it is 0. The number of all reconstructed vector pairs that meet the condition is counted, and divided by the total number of vector pairs to obtain the correlation integral. For example, when r=0.5, the correlation integral value is 0.127; when r=1.0, the correlation integral value is 0.356. The correlation integral is plotted as a curve after taking the logarithm of the radius r, and the slope is fitted using the least squares method in the linear region, yielding a correlation dimension D=1.56.

[0058] A fractional differential equation is constructed to process the reconstructed vector. The fractional differential order α = 0.75 is chosen, and the fractional differential equation is discretized using the finite difference method. The first five terms of the fractional differential coefficient matrix are: 1.000, -0.750, 0.219, -0.094, and 0.048. The fractional differential is applied to each reconstructed vector. For example, the fractional differential is applied to the reconstructed vector [0.213, 0.215, 0.217, 0.219] of the 1-band mean feature, yielding a result of 0.00621. The dynamic weight coefficient is set to 0.85, and the fractional differential result is multiplied by the dynamic weight to obtain the nonlinear dynamic response 0.00528.

[0059] Calculate the q-th order Tsallis entropy, choosing a non-extended parameter q=1.5. Divide the reconstructed vector space into 32 equal-sized boxes, count the number of reconstructed vectors in each box, and divide by the total number of vectors to obtain the probability distribution. For example, the first 5 terms of the probability distribution for the 32 boxes are: 0.032, 0.045, 0.067, 0.075, and 0.058. Calculate the q-th order Tsallis entropy based on the probability distribution and the non-extended parameter, obtaining an entropy value of 2.873. For features at different scales, calculate the entropy value separately; for example, the entropy value for band 1 is 2.873, for band 2 it is 2.765, and for band 3 it is 2.921. Assign exponential weights to each scale with values ​​of 0.35, 0.25, and 0.20, respectively, with the remaining band weights decreasing sequentially. Perform an exponentially weighted product of the entropy values ​​at different scales to obtain the entropy fusion result of 2.842.

[0060] An optimization objective function is constructed. Weight coefficients are assigned to the correlation dimension and the q-order Tsallis entropy, respectively, with values ​​of 0.4 and 0.6. Multiplying the correlation dimension (1.56) by the weight 0.4 yields 0.624, and multiplying the q-order Tsallis entropy (2.842) by the weight 0.6 yields 1.705. The sum of these two values ​​gives the optimization objective function value of 2.329. This objective function value is compared with the aforementioned nonlinear dynamic response (0.00528). The difference between the two exceeds a preset threshold of 2.32, thus determining the current fault type as a bearing outer race fault.

[0061] This method was validated on a bearing fault dataset, which included four states: normal, inner race fault, outer race fault, and rolling element fault, with 100 samples collected for each state. The diagnostic accuracy of this method was 96.5%, an improvement of 5.3 percentage points compared to traditional methods. Particularly for samples with unclear early fault characteristics, the diagnostic accuracy improved significantly, from 85.7% to 94.2%, validating the effectiveness and robustness of this method under complex operating conditions.

[0062] In one optional implementation, the q-th order Tsallis entropy is calculated based on a probability distribution and an unextended parameter. The entropy values ​​at different scales are then exponentially weighted to obtain an entropy fusion result, including: Obtain the time-domain waveform of multi-scale fault characteristics, segment the time-domain waveform, construct a probability histogram, and calculate the probability distribution of each segment interval based on the probability histogram; Select the range of values ​​for the non-extended parameter q, substitute the probability distribution into the Tsallis entropy calculation formula to obtain the initial entropy value; perform multi-scale decomposition on the time-domain waveform to obtain time series at different scales; perform probability distribution calculation and Tsallis entropy calculation on the time series at each scale to obtain entropy values ​​at multiple scales. Calculate the degree of fluctuation of entropy value at each scale, construct a weight allocation function based on the degree of fluctuation, and generate the exponential weight coefficient corresponding to each scale; perform exponential weighted product operation on the entropy value at different scales and the corresponding exponential weight coefficient to obtain the entropy fusion result.

[0063] The q-order Tsallis entropy calculation technique aims to achieve more accurate fault diagnosis by analyzing fault characteristics at multiple scales. This technique acquires the time-domain waveforms of fault characteristics, segments and processes these waveforms, calculates probability distributions and performs multi-scale decomposition, and finally obtains a comprehensive entropy value representing the fault characteristics through exponential weighted fusion.

[0064] In practical implementation, the system collects fault characteristic signals from the equipment and obtains their time-domain waveform data. For example, by collecting the vibration signal of a rotating machine, with a sampling frequency set to 10kHz and a sampling duration of 1 second, time-domain waveform data of 10,000 sampling points {x(1), x(2), ..., x(10000)} is obtained. This time-domain waveform data can be segmented into 100 segments, each containing 100 data points.

[0065] For each segmented interval, a probability histogram is constructed based on the frequency of occurrence of statistical data. Assuming the data amplitude range [-5, 5] is divided into 20 equal intervals, each data point is assigned to its corresponding interval. The number of data points in each interval is calculated, and this number is divided by the total number of data points to obtain the probability distribution {p(1), p(2), ..., p(20)} for each interval. For example, the probability p(1) = 0.02 for the first interval [-5, -4.5], the probability p(2) = 0.03 for the second interval [-4.5, -4], and so on.

[0066] The non-extended parameter q is selected to have a range of [0.1, 3.0], with a step size of 0.1, for a total of 30 values. For each q value, the probability distribution is substituted into the Tsallis entropy calculation formula to calculate the initial entropy value. In the Tsallis entropy calculation formula, the probability distribution p(i) is raised to the power of q, and the sum is substituted into a specific functional relationship to obtain the initial entropy value. For example, when q=0.5, the calculated initial entropy value is 0.85; when q=1.2, the initial entropy value is 0.76.

[0067] Multi-scale decomposition of time-domain waveforms can be performed using empirical mode decomposition or wavelet decomposition methods. Taking wavelet decomposition as an example, the db4 wavelet basis function is selected, and a 3-level decomposition is performed to obtain four time series sub-signals at different scales: {s1, s2, s3, s4}. Here, s1 represents the high-frequency component, and s4 represents the low-frequency component. For each scale of the time series, the probability distribution is calculated using the aforementioned method, and the Tsallis entropy value is calculated for different q values.

[0068] Specifically, for scale s1, the entropy sequence H1={h1(0.1), h1(0.2),..., h1(3.0)} is calculated for different q values; for scale s2, the entropy sequence H2={h2(0.1), h2(0.2),..., h2(3.0)} is obtained; and so on, to obtain H3 and H4. Example values ​​are as follows: when q=0.5, the entropy values ​​for the four scales are h1(0.5)=0.90, h2(0.5)=0.85, h3(0.5)=0.78, and h4(0.5)=0.72, respectively; when q=1.2, the entropy values ​​for the four scales are h1(1.2)=0.82, h2(1.2)=0.77, h3(1.2)=0.70, and h4(1.2)=0.65, respectively.

[0069] The degree of fluctuation in entropy values ​​at each scale can be measured by calculating the standard deviation of the entropy value sequence. Calculating the standard deviations of H1, H2, H3, and H4 respectively yields σ1=0.15, σ2=0.12, σ3=0.09, and σ4=0.06. A larger standard deviation indicates a higher sensitivity of the entropy value to the q parameter at that scale, and thus contains more information.

[0070] A weighting function is constructed based on the degree of entropy fluctuation, generating exponential weight coefficients for each scale. Weighting can be normalized so that the sum of the weights for each scale is 1. The calculation method is as follows: divide the standard deviation of each scale by the sum of its standard deviations to obtain the normalized weights. For example, the weight of scale s1 is w1 = σ1 / (σ1+σ2+σ3+σ4) = 0.15 / (0.15+0.12+0.09+0.06) = 0.36. Similarly, w2 = 0.29, w3 = 0.21, and w4 = 0.14 are calculated.

[0071] To enhance the impact of weights on the results, the normalized weights are converted into exponential weight coefficients. This can be achieved using the power of α, where α is an adjustment parameter with a value of 2. The calculated values ​​are: w1^α = 0.36^2 = 0.13, w2^α = 0.29^2 = 0.08, w3^α = 0.21^2 = 0.04, and w4^α = 0.14^2 = 0.02. Further normalization yields the final exponential weight coefficients: e1 = 0.48, e2 = 0.30, e3 = 0.15, and e4 = 0.07.

[0072] The entropy fusion result is obtained by performing an exponentially weighted product operation on the entropy values ​​at different scales and their corresponding exponential weight coefficients. For each q value, the fusion entropy value F(q) is calculated as F(q) = [h1(q)]^e1 × [h2(q)]^e2 × [h3(q)]^e3 × [h4(q)]^e4. For example, when q = 0.5, the fusion entropy value F(0.5) = 0.90^0.48 × 0.85^0.30 × 0.78^0.15 × 0.72^0.07 = 0.85; when q = 1.2, the fusion entropy value F(1.2) = 0.82^0.48 × 0.77^0.30 × 0.70^0.15 × 0.65^0.07 = 0.77.

[0073] Using the above method, a series of fusion entropy values ​​{F(0.1), F(0.2), ..., F(3.0)} corresponding to different q values ​​can be obtained, constituting the q-order Tsallis entropy spectrum of the fault features. This entropy spectrum contains the complexity information of the signal at multiple scales, which can effectively characterize fault features and realize the identification and classification of fault types.

[0074] In practical applications, the q-order Tsallis entropy spectra of different fault types can be used as feature vectors and input into classifiers such as support vector machines and random forests for fault diagnosis. Tests show that compared with entropy values ​​of a single scale or a single q-value, this method integrates multi-scale and multi-q-value information, improving fault identification accuracy by about 15%, especially showing significant advantages in the diagnosis of weak and complex faults.

[0075] The wind power booster station equipment fault prediction and diagnosis system of this invention includes: The first unit is used to acquire equipment operation data of the wind power booster station, store the equipment operation data into a knowledge graph database, and establish a knowledge graph according to the equipment topology relationship based on the knowledge graph database. The second unit is used to input the device operation data into a dual-stream heterogeneous graph neural network. The static topology branch generates a spatial propagation matrix based on the physical connection of the devices, and the dynamic temporal branch calculates the state correlation through a sliding time window to generate a temporal correlation matrix. The two matrices are fused to obtain the spatiotemporal collaborative features of the device group. The third unit is used to perform multi-hop reasoning on the knowledge graph based on the spatiotemporal collaborative features, calculate the attention weights from fault symptom nodes to potential root cause nodes, and generate a candidate set of reasoning paths. The fourth unit is used to extract key evidence chains from the candidate set of reasoning paths and calculate the credibility score of each reasoning path in combination with the degree of anomaly of the device operation data. The fifth unit is used to sort the inference paths according to the credibility score, select the inference path with the highest credibility score as the fault diagnosis result, extract multi-scale fault features from the fault diagnosis result, reconstruct the multi-scale fault features into a phase space vector, calculate the objective function according to the Tsallis entropy, and obtain the final fault diagnosis result based on the objective function and the nonlinear dynamic response.

[0076] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0077] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0078] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting and diagnosing equipment faults in wind power booster stations, characterized in that, include: Obtain equipment operation data from the wind power booster station and store the equipment operation data in a knowledge graph database; Based on the knowledge graph database, a knowledge graph is established according to the device topology relationship; The device operation data is input into a dual-stream heterogeneous graph neural network, where the static topology branch generates a spatial propagation matrix based on the physical connections of the devices, and the dynamic temporal branch calculates the state correlation through a sliding time window to generate a temporal correlation matrix. The two matrices are fused to obtain the spatiotemporal collaborative features of the device group. Based on the aforementioned spatiotemporal collaborative features, multi-hop reasoning is performed on the knowledge graph, the attention weights from fault symptom nodes to potential root cause nodes are calculated, and a candidate set of reasoning paths is generated. Key evidence chains are extracted from the candidate set of inference paths, and the credibility score of each inference path is calculated in combination with the degree of anomaly in the device operation data. The reasoning paths are sorted according to the credibility score, and the reasoning path with the highest credibility score is selected as the fault diagnosis result. Multi-scale fault features are extracted from the fault diagnosis result. The multi-scale fault features are reconstructed into phase space vectors, and the objective function is calculated based on the Tsallis entropy. The final fault diagnosis result is obtained based on the objective function and the nonlinear dynamic response.

2. The method according to claim 1, characterized in that, The device operation data is input into a two-stream heterogeneous graph neural network. The static topology branch generates a spatial propagation matrix based on the physical connections of the devices, while the dynamic temporal branch calculates state correlations using a sliding time window to generate a temporal correlation matrix. The two matrices are then fused to obtain the spatiotemporal collaborative features of the device group, including: A spatial adjacency matrix is ​​constructed based on the physical connection relationship between equipment nodes in the wind power booster station. A physical parameter matrix is ​​generated based on the physical parameter information of the equipment nodes. The spatial propagation matrix is ​​obtained by multiplying the spatial adjacency matrix and the physical parameter matrix. The operating status data of the device nodes are collected within a preset time window, and the dynamic correlation strength between the device nodes is calculated through multi-scale temporal correlation analysis to generate a temporal correlation matrix. The spatial propagation matrix is ​​input into a first graph convolutional network to obtain a spatial feature representation, and the temporal correlation matrix is ​​input into a second graph convolutional network to obtain a temporal feature representation; Calculate the cross-branch attention weights of the spatial feature representation and the temporal feature representation, perform weighted fusion of the spatial feature representation and the temporal feature representation based on the cross-branch attention weights, and introduce feature interaction terms to obtain the spatiotemporal collaborative features of the device group; the feature interaction terms are obtained through element-wise multiplication of the spatial feature representation and the temporal feature representation.

3. The method according to claim 1, characterized in that, Based on the aforementioned spatiotemporal collaborative features, multi-hop reasoning is performed on the knowledge graph, the attention weights from fault symptom nodes to potential root cause nodes are calculated, and a candidate set of reasoning paths is generated, including: The spatiotemporal collaborative features are concatenated and fused with the feature vectors corresponding to the nodes in the knowledge graph to generate enhanced node feature vectors, wherein the nodes in the knowledge graph include fault symptom nodes and fault cause nodes. Starting from the fault symptom node, the reasoning path to the fault cause node is calculated through K-hop attention propagation; the K-hop attention propagation includes iteratively calculating the attention score between nodes in each round through K rounds of message passing, and weighting the enhanced node feature vector based on the attention score between nodes in each round to obtain the attention-weighted node feature vector; The inference path from the fault symptom node to the fault cause node is scored in multiple dimensions. The attention scores of adjacent nodes on the inference path are multiplied to obtain the path relevance score. The normalized score is obtained based on the difference between the inference path length and the standard path length. The minimum edit distance between the inference path and each path in the historical diagnostic case set is calculated. The minimum edit distance is divided by the average path length to obtain the novelty score. The path relevance score, the normalization score, and the path novelty score are multiplied by their respective weight coefficients and then summed to obtain a comprehensive score for the inference path; the three inference paths with the highest comprehensive scores are selected as the candidate set of inference paths.

4. The method according to claim 1, characterized in that, Key evidence chains are extracted from the candidate inference paths, and a credibility score for each inference path is calculated based on the degree of anomaly in the device operation data, including: Calculate the connectivity centrality, depth level, and fault impact of each node in the reasoning path, and perform a weighted combination to obtain a node importance score; generate node weights based on the node importance scores, calculate the relationship strength between adjacent nodes, and accumulate the multiplication of the node weights and the relationship strength between adjacent nodes to obtain an evidence chain strength score; Obtain time-series data of device operating parameters, calculate the first anomaly score corresponding to the local time window and the second anomaly score corresponding to the global historical data, and obtain the parameter anomaly score by weighted summation of the first anomaly score and the second anomaly score. The information source adjustment factor is calculated based on the historical diagnostic accuracy. The information source adjustment factor is normalized by the exponential function of the historical diagnostic accuracy to obtain the information source weight. The evidence chain strength score and parameter anomaly score are multiplied by the corresponding information source weight and summed to obtain the comprehensive credibility score of the reasoning path.

5. The method according to claim 1, characterized in that, Multi-scale fault features are extracted from the fault diagnosis results; the multi-scale fault features are reconstructed into phase space vectors, an objective function is constructed based on Tsallis entropy, and the final fault diagnosis results are obtained based on the objective function and the nonlinear dynamic response, including: The multi-scale fault features are reconstructed in phase space using the Takens embedding theorem to obtain a reconstruction vector, which contains sampled values ​​of the fault signal at different time delays. The correlation integral is calculated based on the reconstruction vector. The correlation integral is obtained by using the Heaviside function to determine the relationship between the distance between the reconstruction vectors and a specified radius. The logarithmic ratio of the correlation integral to the specified radius is used as the correlation dimension. Construct a fractional differential equation, the order of which is between 0 and 1, perform fractional differential operation on the reconstructed vector, and use the product of the fractional differential result and the dynamic weight coefficient as the nonlinear dynamic response. Calculate the q-th order Tsallis entropy, which is obtained based on the probability distribution and the non-extended parameter. Then, perform an exponential weighted product of the entropy values ​​at different scales to obtain the entropy fusion result. The optimization objective function value is obtained by multiplying the correlation dimension and the q-order Tsallis entropy by the corresponding weight coefficients and summing them. The diagnostic conclusion is then updated based on the optimization objective function value and the nonlinear dynamic response.

6. The method according to claim 5, characterized in that, The q-th order Tsallis entropy is calculated based on the probability distribution and the non-extended parameter. The entropy values ​​at different scales are then exponentially weighted to obtain the entropy fusion result, including: Obtain the time-domain waveform of multi-scale fault characteristics, segment the time-domain waveform, construct a probability histogram, and calculate the probability distribution of each segment interval based on the probability histogram; Select the range of values ​​for the non-extended parameter q, substitute the probability distribution into the Tsallis entropy calculation formula to obtain the initial entropy value; perform multi-scale decomposition on the time-domain waveform to obtain time series at different scales; perform probability distribution calculation and Tsallis entropy calculation on the time series at each scale to obtain entropy values ​​at multiple scales. Calculate the degree of fluctuation of entropy value at each scale, construct a weight allocation function based on the degree of fluctuation, and generate the exponential weight coefficient corresponding to each scale; perform exponential weighted product operation on the entropy value at different scales and the corresponding exponential weight coefficient to obtain the entropy fusion result.

7. A wind power substation equipment fault prediction and diagnosis system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to acquire equipment operation data of the wind power booster station and store the equipment operation data into a knowledge graph database. Based on the knowledge graph database, a knowledge graph is established according to the device topology relationship; The second unit is used to input the device operation data into a dual-stream heterogeneous graph neural network. The static topology branch generates a spatial propagation matrix based on the physical connection of the devices, and the dynamic temporal branch calculates the state correlation through a sliding time window to generate a temporal correlation matrix. The two matrices are fused to obtain the spatiotemporal collaborative features of the device group. The third unit is used to perform multi-hop reasoning on the knowledge graph based on the spatiotemporal collaborative features, calculate the attention weights from fault symptom nodes to potential root cause nodes, and generate a candidate set of reasoning paths. The fourth unit is used to extract key evidence chains from the candidate set of reasoning paths and calculate the credibility score of each reasoning path in combination with the degree of anomaly of the device operation data. The fifth unit is used to sort the reasoning paths according to the credibility score, select the reasoning path with the highest credibility score as the fault diagnosis result, and extract multi-scale fault features from the fault diagnosis result. The multi-scale fault features are reconstructed into phase space vectors, and the objective function is calculated based on the Tsallis entropy. The final fault diagnosis result is obtained based on the objective function and the nonlinear dynamic response.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images