Wind turbine generator performance anomaly detection method and system

Abstract

The invention provides a wind turbine generator performance anomaly detection method and system, and the method comprises the steps: collecting an operation state data set of a wind turbine generator in a continuous operation period, constructing a dynamic energy topological graph which reflects the energy conversion and transmission process of the wind turbine generator based on the operation state data set, and mapping the operation state data set to the dynamic energy topological graph, analyzing and extracting an energy flow feature vector representing the transfer characteristic of energy between the components through a topological structure; calling a pre-constructed historical health state energy flow model to perform dynamic matching processing on the energy flow feature vector, and calculating an energy flow deviation degree between a current energy flow mode and a historical health mode; and according to the energy flow deviation degree and the wind speed environment parameters collected in real time, optimizing an anomaly judgment threshold value, and outputting a wind turbine generator performance anomaly detection result based on the optimized anomaly judgment threshold value. According to the method, the performance anomaly detection of the wind turbine generator can be more accurately and reliably realized.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to a method and system for detecting abnormal performance of wind turbine generators. Background Technology

[0002] As a core component of new energy power generation, the operating performance of wind turbines directly impacts power generation efficiency and equipment safety. Anomaly detection is crucial for ensuring the stable operation of wind turbines. Real-time monitoring and analysis of operating status, timely identification and early warning of potential anomalies, are of great significance for reducing operation and maintenance costs and improving power generation reliability. A wind turbine is essentially a complex energy conversion and transfer system. Energy capture from wind to electrical output requires the coordinated action of multiple components. Traditional methods focus only on isolated component parameters or static structural relationships, making it difficult to capture the dynamic characteristics of energy transfer between components. This results in insufficient sensitivity in identifying system-level performance anomalies. Furthermore, fixed thresholds or simple statistical models do not consider the differences in energy flow patterns under different operating conditions, making them prone to misjudgment or missed detection when environmental parameters such as wind speed change, affecting the accuracy and reliability of anomaly detection. Summary of the Invention

[0003] In view of this, the present invention provides a method and system for detecting abnormal performance of wind turbine units.

[0004] According to one aspect of the present invention, a method for detecting performance anomalies in wind turbine generators is provided. The method includes: collecting an operational status dataset of the wind turbine generators during a continuous operating cycle, wherein the operational status dataset is synchronously collected by a sensor array and an environmental monitoring device distributed on key components of the wind turbine generators; constructing a dynamic energy topology map reflecting the energy conversion and transfer process of the wind turbine generators based on the operational status dataset, wherein the nodes of the dynamic energy topology map correspond to the energy conversion components of the wind turbine generators, and the edges correspond to the energy transfer paths between components; mapping the operational status dataset to the dynamic energy topology map, and extracting energy flow feature vectors characterizing the energy transfer properties between components through topological structure analysis; calling a pre-constructed historical healthy state energy flow model to dynamically match the energy flow feature vectors, and calculating the energy flow deviation between the current energy flow mode and the historical healthy mode; optimizing an anomaly judgment threshold based on the energy flow deviation and real-time collected wind speed environmental parameters, and outputting the wind turbine generator performance anomaly detection result based on the optimized anomaly judgment threshold.

[0005] According to another aspect of the present invention, a computer system is provided, comprising: a processor; and a memory, wherein the memory stores computer-readable code that, when executed by the processor, causes the processor to perform the method as described above.

[0006] The wind turbine performance anomaly detection method provided by this invention collects operational status datasets from continuous operating cycles of the wind turbine, constructs a dynamic energy topology map reflecting the energy conversion and transfer process, and abstracts the wind turbine into a dynamic system model with energy conversion components as nodes and energy transfer paths as edges. This method can capture the dynamic characteristics of energy transfer between components from a system-wide perspective, overcoming the limitations of traditional isolated component parameter monitoring. By mapping the operational status dataset to the dynamic energy topology map and extracting energy flow feature vectors through topology analysis, the extracted features integrate the component energy conversion characteristics and path transfer characteristics, providing a more comprehensive characterization of the system. The system operates under a single operating state; it calls upon historical health state energy flow models to dynamically match energy flow feature vectors and calculate energy flow deviation. By dynamically matching with historical health models, it quantifies the difference between the current energy flow pattern and the health state, rather than relying on a fixed threshold, thus improving the accuracy of anomaly identification. It combines energy flow deviation and real-time wind speed environmental parameters to optimize the anomaly judgment threshold, making the anomaly judgment threshold adaptable to the energy flow characteristics under different wind speed conditions. This avoids misjudgment under complex conditions by using fixed thresholds and improves the robustness of detection. Through the synergistic effect of the above technical means, it is possible to more accurately and reliably detect anomalies in wind turbine performance. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the application environment architecture provided by the present invention; Figure 2 This is a flowchart illustrating a method for detecting abnormal performance of wind turbine generators provided by the present invention. Figure 3 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be explained below with reference to the accompanying drawings. To facilitate a clearer understanding of the present invention, the application environment in which the present invention is implemented will be introduced first, such as... Figure 1 As shown, in the application environment of this invention, a computer system 10, a sensor array, and an environmental monitoring device are provided. The sensor array includes multiple sensors; the number of sensors is not limited here. Figure 1 As shown, the sensor array may specifically include sensor 1, sensor 2, ..., sensor n; it is understood that sensor 1, sensor 2, sensor n, ..., and environmental monitoring device can all be network connected to computer system 10 so that each sensor and environmental monitoring device can interact with computer system 10 through network connection.

[0009] Understandably, computer system 10 can refer to equipment that performs anomaly detection and analysis, which may be a computer or server located in the monitoring backend. Sensors may specifically refer to temperature sensors, vibration sensors, etc., but are not limited to these. Environmental monitoring devices may include anemometers or wind vanes, etc., but are not limited to these. The number of sensors and environmental monitoring devices is not limited. Each sensor and environmental monitoring device can be directly or indirectly connected to computer system 10 via wired or wireless communication.

[0010] Further, please see Figure 2 This is a flowchart illustrating a method for detecting performance anomalies in wind turbine units provided in an embodiment of the present invention. Figure 2 As shown, this method can be derived from... Figure 1 The method for detecting abnormal performance of a wind turbine generator is executed by a computer system 10, and may include the following steps: Step S100: Collect the operating status dataset of the wind turbine during a continuous operating cycle. The operating status dataset is collected synchronously by sensor arrays and environmental monitoring devices distributed on key components of the wind turbine.

[0011] The operational status dataset is a collection of data reflecting various aspects of the wind turbine's operation. The sensor array is a system composed of multiple sensors of different types, distributed across key components of the wind turbine, used to monitor the operational status of these components in real time. For example, temperature sensors monitor temperature changes in components, and vibration sensors detect vibrations. Environmental monitoring devices collect environmental information surrounding the wind turbine, such as wind speed, wind direction, and ambient temperature. Synchronous acquisition means that the sensor array and environmental monitoring devices collect data at the same point in time to ensure that the collected data accurately reflects the wind turbine's operational status and environmental conditions at the set time.

[0012] In this embodiment of the invention, the sensor array may include a temperature sensor and a vibration sensor mounted on the gearbox, used to monitor changes in gearbox oil temperature and vibration frequency, respectively; and a current sensor and a voltage sensor mounted on the generator, used to collect generator current and voltage data. The environmental monitoring device may employ an anemometer and wind vane to measure wind speed and direction.

[0013] Step S200: Construct a dynamic energy topology map based on the operating status dataset to reflect the energy conversion and transfer process of the wind turbine. The nodes of the dynamic energy topology map correspond to the energy conversion components of the wind turbine, and the edges correspond to the energy transfer paths between the components.

[0014] Dynamic energy topology diagrams are graphical representations used to describe the energy conversion and transfer processes of wind turbines. Nodes represent various energy conversion components in the wind turbine, such as blades, gearboxes, and generators, each with corresponding physical properties and energy conversion functions. Edges represent energy transfer paths between these components, with the edge direction indicating the direction of energy transfer and the edge weight indicating the intensity of energy transfer. By constructing dynamic energy topology diagrams, the energy flow of wind turbines at different points in time can be visually displayed, which helps in analyzing energy conversion efficiency and identifying potential anomalies.

[0015] In an optional implementation, step S200 can be specifically implemented as the following steps S210~S240: Step S210: Extract component attributes from the operating status dataset, identify the energy conversion components and their physical attribute parameters in the wind turbine, and establish a set of topological nodes containing unique component identifiers, type classification labels, and rated energy conversion efficiency parameters.

[0016] Component attribute extraction involves filtering information related to energy conversion components from the operational status dataset and extracting their physical attribute parameters. Energy conversion components are key parts of wind turbines that convert energy into mechanical energy, such as blades that convert wind energy into mechanical energy and generators that convert mechanical energy into electrical energy. Physical attribute parameters include component temperature, vibration frequency, and power, which reflect the component's operational status and performance. A unique component identifier is assigned to each energy conversion component to accurately distinguish different components in the topology map. Type classification labels are used to categorize components; for example, blades are classified as wind energy capture components, and generators are classified as electrical output components. The rated energy conversion efficiency parameter refers to the component's energy conversion efficiency under ideal conditions and is an important reference for evaluating the component's actual operating efficiency.

[0017] In this embodiment of the invention, energy conversion components in a wind turbine can be identified by analyzing equipment identification information and physical parameter data in the operating status dataset. For example, components such as blades, gearboxes, and generators are determined based on the equipment model and installation location. Then, physical attribute parameters such as temperature and power of these components are extracted. A unique identifier is assigned to each component, and corresponding classification tags are added according to the component's function and type. Simultaneously, the rated energy conversion efficiency parameters of each component are obtained from the equipment's technical documentation, and a set of topology nodes containing this information is established.

[0018] Step S220: Based on the energy conversion type and physical connection relationship of each component in the topology node set, extract the energy transfer power data between components from the operating status dataset, construct a set of directed weighted edges describing the direction and intensity of energy transfer, and calculate the weight value of the directed weighted edge by the ratio of real-time transfer power to rated transfer power.

[0019] Energy conversion type refers to the way a component converts one form of energy into another; for example, blades convert wind energy into mechanical energy, and generators convert mechanical energy into electrical energy. Physical connection relationship refers to the actual physical connection between components, such as connections via shafts, gears, etc. Energy transfer power data refers to the magnitude of energy transferred between components, reflecting the intensity of energy transfer. A directed weighted edge set is a set of directed edges, each representing an energy transfer path. The direction of the edge indicates the direction of energy transfer, and the weight of the edge indicates the intensity of energy transfer. Real-time transfer power refers to the actual power transferred between components at a given moment, while rated transfer power refers to the maximum transfer power specified in the component's design. By calculating the ratio of real-time transfer power to rated transfer power, the weight value of the directed weighted edge can be obtained, reflecting the efficiency and stability of energy transfer.

[0020] In an optional implementation, step S220 can be performed as the following steps S221~S225: Step S221: Perform hierarchical division of the energy conversion components in the topology node set. According to the energy transfer path of the wind turbine, from the wind energy capture component to the power output component, they are sequentially labeled as the primary conversion layer, the secondary conversion layer, and the terminal output layer, generating a node hierarchical topology structure containing hierarchical attributes.

[0021] Hierarchical partitioning divides the energy conversion components of a wind turbine into different levels based on the energy transfer path. The primary conversion layer mainly includes wind energy capture components, such as blades, which directly convert wind energy into mechanical energy. The secondary conversion layer includes components that further process and convert mechanical energy, such as gearboxes, which can increase the rotational speed and torque of the mechanical energy. The final output layer consists of components that ultimately convert mechanical energy into electrical energy, such as generators. This hierarchical partitioning clearly illustrates the energy transfer process within the wind turbine, facilitating subsequent analysis and processing.

[0022] In this embodiment of the invention, the blades are marked as components of the primary conversion layer because they are the first components in the wind turbine to capture wind energy and perform energy conversion. The gearbox is marked as a component of the secondary conversion layer, receiving and converting the mechanical energy transmitted by the blades. The generator is marked as a component of the terminal output layer, converting the mechanical energy converted by the gearbox into electrical energy. Through this hierarchical division, a node-level topology containing hierarchical attributes is generated.

[0023] Step S222: Based on the physical connection relationships of each component in the node hierarchy topology, construct an initial undirected connection graph. The edges in the initial undirected connection graph represent direct physical connections between components.

[0024] Physical connections refer to the actual connections between components via mechanical structures or electrical circuits. An initial undirected connection graph is a simple graphical representation where edges only indicate direct physical connections between components, without considering the direction of energy transfer. By constructing an initial undirected connection graph, the physical connections between components can be visually displayed, providing a foundation for subsequently determining the direction of energy transfer and constructing directed graphs.

[0025] In this embodiment of the invention, the physical connection relationships between components are determined based on the installation position and connection method of each component in the node-level topology. For example, a blade is connected to a gearbox via a shaft, and the gearbox is connected to a generator via a shaft. Based on these connection relationships, an initial undirected connection graph is constructed, wherein there is an edge connecting the blade node and the gearbox node, and there is also an edge connecting the gearbox node and the generator node.

[0026] Step S223: Extract the input power data and output power data of each component from the operating status dataset, calculate the power transfer difference between adjacent level components, and determine that there is an effective energy transfer relationship when the proportion of the power transfer difference to the output power of the upstream component exceeds a preset threshold.

[0027] Input power data refers to the energy power received by a component, while output power data refers to the energy power output by the component. Power transfer difference refers to the difference between the input and output power between adjacent components, reflecting power loss during energy transfer. The preset threshold is a pre-defined percentage value used to determine whether the power transfer difference is within a normal range. When the proportion of the power transfer difference to the output power of the upstream component exceeds the preset threshold, it indicates significant power loss along the energy transfer path, suggesting a possible effective energy transfer relationship.

[0028] In this embodiment of the invention, the output power data of the blades and the input power data of the gearbox, as well as the output power data of the gearbox and the input power data of the generator, are extracted from the operating status dataset. The power transfer difference between the blades and the gearbox, and the power transfer difference between the gearbox and the generator are calculated.

[0029] Step S224: Based on the effective energy transfer relationship, the initial undirected connection graph is converted into a directed graph. The direction of the directed edge is from the energy output component to the energy input component, and the power transfer difference is used as the initial weight value of the directed edge.

[0030] Effective energy transfer relationships determine the actual direction of energy transfer between components. Converting an initial undirected graph into a directed graph involves adding directions to the edges based on these effective energy transfer relationships. The direction of a directed edge points from an energy output component to an energy input component, visually representing the direction of energy transfer. The power transfer difference serves as the initial weight value for each directed edge, reflecting the power loss along that energy transfer path.

[0031] In this embodiment of the invention, based on the effective energy transfer relationship determined in step S223, the edges in the initial undirected connected graph are converted into directed edges. For example, since the blade is an energy output component and the gearbox is an energy input component, a direction is added to the edge from the blade node to the gearbox node, making it a directed edge. Simultaneously, the power transfer difference between the blade and the gearbox is used as the initial weight value of this directed edge.

[0032] Step S225: Normalize the initial weight values, calculate the standardized weight values ​​by the ratio of the current transmission power to the design maximum transmission power, and generate a set of directed weighted edges containing the standardized weight values.

[0033] Normalization transforms initial weight values ​​from different ranges into a unified range to facilitate comparison and analysis. Current transmitted power refers to the actual power transmitted between components at a given moment, while the design maximum transmitted power refers to the maximum transmitted power specified in the design. By calculating the ratio of current transmitted power to the design maximum transmitted power, a standardized weight value can be obtained, which more accurately reflects the intensity and efficiency of energy transfer.

[0034] In this embodiment of the invention, the initial weight values ​​of the directed edges obtained in step S224 are normalized. First, the designed maximum transmission power of each component is obtained. Then, the ratio of the current transmission power to the designed maximum transmission power is calculated to obtain the normalized weight value. For example, for the directed edge between the blade and the gearbox, its initial weight value is divided by the designed maximum transmission power of the blade to obtain the normalized weight value. The normalized weight values ​​of all directed edges are integrated to generate a set of directed weighted edges containing the normalized weight values.

[0035] Step S230: Perform sliding window segmentation on the continuous running cycle according to the preset time window division rules to generate multiple static topology snapshots of time windows. Each static topology snapshot contains the set of topology nodes and the set of directed weighted edges within the corresponding time window.

[0036] The preset time window division rules refer to the pre-defined rules used to divide time windows, such as the length of the time window and the sliding step size. Sliding window segmentation divides the continuous operating cycle according to the time window division rules, resulting in multiple overlapping or non-overlapping time windows. Each time window corresponds to a static topology snapshot, which contains the set of topological nodes and the set of directed weighted edges within that time window. The static topology snapshot reflects the energy conversion and transfer of the wind turbine within the time window and is the basis for constructing the dynamic energy topology map.

[0037] Step S240: Perform topological structure difference analysis on static topological snapshots of adjacent time windows, calculate the change rate of node attributes and the gradient of edge weight change, and construct a dynamic energy topology map containing time dimension evolution characteristics. The time resolution of the dynamic energy topology map is consistent with the sampling frequency of the sensor array.

[0038] Topology difference analysis involves comparing static topology snapshots from adjacent time windows to analyze the differences in their topology. Node attribute change rate refers to the rate of change of a node's physical attribute parameters within adjacent time windows, such as temperature change rate and vibration frequency change rate. Edge weight change gradient refers to the gradient of the weight values ​​of directed weighted edges within adjacent time windows, reflecting changes in energy transfer intensity. Temporal evolution characteristics refer to the temporal changes in the energy conversion and transfer process of the wind turbine, reflected by the node attribute change rate and edge weight change gradient. The temporal resolution of the dynamic energy topology map is consistent with the sampling frequency of the sensor array, meaning that the dynamic energy topology map can accurately reflect the energy conversion and transfer of the wind turbine at the sensor sampling time.

[0039] In an optional implementation, step S240 can be specifically implemented as the following steps S241~S245: Step S241: Perform node matching processing on static topology snapshots of two consecutive adjacent time windows, establish node correspondence across time windows through the unique identifier of the component, and identify newly added nodes, removed nodes, and nodes with changed attributes.

[0040] Node matching involves matching nodes in static topology snapshots from adjacent time windows to find their correspondences. A component unique identifier is used to distinguish different components, allowing for accurate location of the corresponding node in topology snapshots across different time windows. New nodes are those that appear in later time windows but were not present in earlier ones; removed nodes are those that existed in earlier time windows but disappeared in later ones. Nodes with changed attributes are those whose physical attribute parameters change within adjacent time windows.

[0041] In this embodiment of the invention, node matching is performed on static topology snapshots from two consecutive adjacent time windows. For example, for the static topology snapshots of the first and second time windows, the corresponding blade nodes in the two snapshots are found using the unique component identifier "WT-01-Blade". During the matching process, it is found that a new node, "WT-01-Sensor-New", has been added to the snapshot of the second time window; this is a newly added node. Simultaneously, the "WT-01-Old-Sensor" node from the first time window disappears in the second time window; this is a removed node. Furthermore, it is also found that the temperature attribute of the blade node changed within the two time windows; this node is an attribute change node.

[0042] Step S242: Perform time series difference processing on the physical attribute parameters of the attribute change nodes, calculate the parameter change per unit time, and generate node attribute change rate index. The node attribute change rate index includes temperature change rate, vibration frequency change rate, and efficiency change rate.

[0043] Time series differencing involves calculating the difference in physical attribute parameters of nodes over time, specifically calculating the parameter difference between adjacent time points. The parameter change per unit time refers to the magnitude of the parameter change within that unit of time, reflecting the rate of change. Node attribute change rate indices are used to measure the changes in node physical attribute parameters, including temperature change rate, vibration frequency change rate, and efficiency change rate. Temperature change rate refers to the percentage change in node temperature per unit time; vibration frequency change rate refers to the percentage change in node vibration frequency per unit time; and efficiency change rate refers to the percentage change in node energy conversion efficiency per unit time.

[0044] In this embodiment of the invention, for nodes with changing attributes (such as blade nodes), their temperature attributes change within adjacent time windows. Time series differencing is performed on the temperature parameters of the blade nodes to calculate the temperature change per unit time. Assuming the temperature of the first time window is T1, the temperature of the second time window is T2, and the length of the time window is Δt, the rate of temperature change is (T2-T1) / Δt. Similarly, time series differencing is performed on the vibration frequency and energy conversion efficiency parameters of the blade nodes to calculate the rate of change of vibration frequency and efficiency, generating a node attribute change rate index.

[0045] Step S243: Calculate the weight difference for each directed edge in the set of directed weighted edges to obtain the change in edge weight between adjacent time windows, and calculate the gradient of edge weight change in combination with the time window interval. The sign of the gradient of edge weight change indicates the direction of increase or decrease of the weight value.

[0046] Weight difference calculation refers to calculating the difference in weight values ​​of directed weighted edges between adjacent time windows, thus obtaining the change in edge weights. The time window interval refers to the time interval between adjacent time windows. The edge weight change gradient is the rate of change of edge weight values ​​per unit time, obtained by dividing the change in edge weights by the time window interval. The sign of the edge weight change gradient indicates the direction of weight increase or decrease; a positive sign indicates an increase in weight value, and a negative sign indicates a decrease in weight value.

[0047] In this embodiment of the invention, for each directed edge in the set of directed weighted edges, such as the directed edge from the blade node to the gearbox node, the difference in its weight value between adjacent time windows is calculated. Assuming the edge weight value is W1 in the first time window, W2 in the second time window, and the time window interval is Δt, then the change in edge weight is W2 - W1, and the gradient of the edge weight change is (W2 - W1) / Δt. If the gradient of the edge weight change is positive, it indicates that the weight value of the directed edge is increasing, i.e., the energy transfer intensity is increasing; if the gradient of the edge weight change is negative, it indicates that the weight value is decreasing, and the energy transfer intensity is weakening.

[0048] Step S244: Construct a topology evolution feature set containing a matrix of node attribute change rates and a gradient tensor of edge weight change. The dimension of the topology evolution feature set should match the number of nodes and edges in the static topology snapshot.

[0049] The node attribute change rate matrix is ​​a matrix composed of the node attribute change rate indices of all nodes with changing attributes arranged according to certain rules. Rows in the matrix correspond to different nodes, and columns correspond to different attribute change rate indices. The edge weight change gradient tensor is a tensor composed of the edge weight change gradients of all directed edges arranged according to certain rules. The dimension of the tensor is related to the number of edges and the associated attributes. The topology evolution feature set is a set containing the node attribute change rate matrix and the edge weight change gradient tensor. Its dimension matches the number of nodes and edges in the static topology snapshot to accurately reflect the temporal evolution characteristics of the wind turbine's energy conversion and transfer process.

[0050] In this embodiment of the invention, assuming there are n attribute change nodes, and each node has m attribute change rate indices, the node attribute change rate matrix is ​​an n×m matrix. Assuming there are k directed edges, the dimension of the edge weight change gradient tensor is determined based on the relevant attributes of the edges. The node attribute change rate matrix and the edge weight change gradient tensor are combined to construct a topological evolution feature set, ensuring that its dimension is consistent with the number of nodes and edges in the static topological snapshot.

[0051] Step S245: Perform spatiotemporal correlation modeling on the set of topological nodes, the set of directed weighted edges, and the set of topological evolution features to generate a dynamic energy topology graph containing static node attributes, static edge weights, and temporal evolution features. The dynamic energy topology graph is stored using a spatiotemporal tensor data structure.

[0052] Spatiotemporal correlation modeling involves comprehensively processing the set of topological nodes, the set of directed weighted edges, and the set of topological evolution features to establish spatiotemporal correlations among them. Node static attributes refer to the fixed values ​​of a node's physical property parameters at a given moment, such as temperature and power. Edge static weights refer to the fixed values ​​of the weights of directed weighted edges at a given moment. Temporal evolution features refer to the temporal changes in the energy conversion and transfer processes of wind turbines, as reflected by the set of topological evolution features. The dynamic energy topology map is stored using a spatiotemporal tensor data structure, effectively representing the energy conversion and transfer of wind turbines in both time and space.

[0053] In this embodiment of the invention, spatiotemporal correlation modeling is performed on the static attributes of nodes in the topological node set, the static weights of edges in the directed weighted edge set, and the temporal evolution features in the topological evolution feature set. For example, the temperature attribute of a node is correlated with the temperature change rate in the temporal evolution features, and the weight value of a directed edge is correlated with the gradient of the edge weight change. Through this correlation modeling, a dynamic energy topology graph containing node static attributes, edge static weights, and temporal evolution features is generated and stored using a spatiotemporal tensor data structure.

[0054] Step S300: Map the running status dataset to a dynamic energy topology map, and extract energy flow feature vectors that characterize the energy transfer properties between components through topology analysis.

[0055] Mapping the operational status dataset to a dynamic energy topology graph involves associating each data point in the dataset with a corresponding node and edge in the graph, allowing for a visual representation of the data. Topology analysis examines the structure and connectivity of the dynamic energy topology graph, extracting key information characterizing energy transfer between components. Energy flow feature vectors are a set of feature vectors describing energy transfer characteristics between components, including energy transfer efficiency, energy loss, and energy transfer paths.

[0056] In an optional implementation, step S300 can be performed as the following steps S310~S350: Step S310: Perform timestamp alignment processing on the operating status dataset to ensure that the data collected by the sensor array is consistent with the time resolution of the dynamic energy topology map, and generate a time-synchronized operating status data sequence.

[0057] Timestamp alignment refers to adding accurate timestamp information to each data point in the operational status dataset and matching these timestamps with the temporal resolution of the dynamic energy topology map. Data collected by sensor arrays may have different sampling frequencies and time intervals; timestamp alignment ensures that this data is temporally consistent with the dynamic energy topology map, guaranteeing data accuracy and validity. A time-synchronized operational status data sequence refers to the operational status data sequence arranged in chronological order after timestamp alignment.

[0058] Step S320: Distribute the monitoring data of each component in the time-synchronized operating status data sequence to the topology nodes corresponding to the dynamic energy topology map, and generate an enhanced node set with real-time monitoring values. The enhanced node set includes the static physical attributes and dynamic monitoring data of the nodes.

[0059] The monitoring data for each component refers to the monitoring data related to each energy conversion component in the time-synchronized operating status data sequence, such as temperature, power, and vibration frequency. Assigning this monitoring data to the corresponding topology nodes in the dynamic energy topology map means associating each monitoring data point with a corresponding node in the topology map, giving each node real-time monitoring values. The enhanced node set is an additional set of nodes with real-time monitoring values ​​added to the original topology node set. This set includes the nodes' static physical properties (such as rated power and rated temperature) and dynamic monitoring data (such as real-time temperature and real-time power).

[0060] In this embodiment of the invention, blade temperature monitoring data in a time-synchronized operating status data sequence is assigned to the corresponding blade nodes in the dynamic energy topology graph. Assuming the static physical properties of the blade nodes include rated temperature and rated power, real-time temperature monitoring values ​​are added to the blade nodes, generating an enhanced node set with real-time monitoring values. In this way, the topology nodes not only possess static properties but also reflect real-time operating status.

[0061] Step S330: Perform topology traversal processing on the dynamic energy topology map, starting from the wind energy capture component and sequentially extracting the weight values ​​of the directed weighted edges and the dynamic monitoring data of the corresponding nodes along the energy transfer direction to construct the sequence data on the energy transfer path.

[0062] Topology traversal processing refers to traversing the nodes and edges of a dynamic energy topology graph according to certain rules to obtain relevant information. Traversal begins with the wind energy capture unit (TEU) along the energy transfer direction because the TEU is the starting point for energy conversion in wind turbines; following the energy transfer direction, each energy conversion unit and energy transfer path can be accessed sequentially. The weights of directed weighted edges reflect the intensity of energy transfer, and the dynamic monitoring data of the corresponding nodes reflect the real-time operating status of the units. By extracting this information and constructing sequential data along the energy transfer path, the energy transfer process between units can be clearly displayed.

[0063] In an optional implementation, step S330 can be performed as the following steps S331 to S335: Step S331: Using the wind energy capture component node in the dynamic energy topology graph as the starting node, the depth-first search algorithm is used to perform path traversal processing on the directed weighted edges, and all complete energy transfer paths from the starting node to the terminal output node are recorded.

[0064] Depth-first search (DFS) is an algorithm used to traverse graphs. Starting from the initial node, it visits nodes as deep as possible along a path until it cannot continue, then backtracks to the previous node and continues visiting other paths. In this embodiment of the invention, wind energy capture component nodes (such as blade nodes) in the dynamic energy topology graph are used as the starting node, and the DFS algorithm is employed to traverse the directed weighted edges. During the traversal, all complete energy transfer paths from the starting node to the terminal output node (such as generator nodes) are recorded. These paths reflect the complete process of energy capture from wind energy to electrical output in the wind turbine.

[0065] In this embodiment of the invention, it is assumed that there are multiple nodes and directed edges in the dynamic energy topology graph. Starting from the blade node, a depth-first search algorithm is used to traverse the path. Starting from the blade node, the gearbox node, generator node, etc., are visited sequentially along the directed edges, and all complete paths from the blade node to the generator node are recorded. For example, a complete energy transfer path is: blade node to gearbox node to generator node.

[0066] Step S332: Calculate the path length for each complete energy transfer path. The path length is defined as the number of directed weighted edges contained in the path. Main energy transfer paths with path lengths within a preset range are retained, and abnormal paths containing anomalous jumps are filtered out.

[0067] Path length calculation refers to calculating the number of directed weighted edges contained in each complete energy transfer path. The preset range is a pre-defined range of path lengths used to filter out the main energy transfer paths. Abnormal jumps refer to edges or nodes in the path that do not conform to the laws of energy transfer, such as a sudden change in energy transfer direction or an abnormal increase or decrease in energy transfer intensity. Filtering out abnormal paths containing abnormal jumps ensures that the main energy transfer paths retained are valid paths that conform to actual energy transfer conditions.

[0068] In this embodiment of the invention, it is assumed that the preset path length range is 2-4 directed weighted edges. The path length is calculated for all complete energy transfer paths recorded in step S331. For example, a path containing 3 directed weighted edges has a path length of 3, which is within the preset range; this path is retained as the primary energy transfer path. Another path containing 5 directed weighted edges exceeds the preset range and is filtered out. Simultaneously, the path is checked for abnormal transitions. If an abnormal transition is found where the energy transfer direction suddenly changes from the generator node to the blade node, this is considered an abnormal transition, and the path is filtered out.

[0069] Step S333: Perform path standardization on the main energy transfer paths that are retained, renumber the nodes on the paths according to the energy transfer order, and generate a standard path structure that includes the node sequence and the corresponding directed weighted edge sequence.

[0070] Path standardization refers to the unified processing of the retained main energy transfer paths to give them a consistent representation. Renumbering the nodes along the path according to the energy transfer order facilitates subsequent analysis and processing. The standard path structure consists of a sequence of nodes and corresponding directed weighted edge sequences. The node sequence is arranged in the order of energy transfer, and the directed weighted edge sequences correspond to the node sequences, reflecting the energy transfer relationships between the nodes.

[0071] In this embodiment of the invention, for the main energy transfer paths that are retained, such as "blade node to gearbox node to generator node", the nodes are renumbered according to the energy transfer order, for example, the blade node is numbered 1, the gearbox node is numbered 2, and the generator node is numbered 3. At the same time, the directed weighted edge sequence between the corresponding nodes is recorded to generate a standard path structure containing the node sequence (1, 2, 3) and the directed weighted edge sequence.

[0072] Step S334: Extract dynamic monitoring data of each node in the standard path structure from the enhanced node set. The dynamic monitoring data includes real-time power value, temperature monitoring value and vibration monitoring value.

[0073] The enhanced node set is a set of nodes containing real-time monitoring values, including dynamic monitoring data for each node. Extracting dynamic monitoring data for each node in the standard path structure from the enhanced node set refers to finding the corresponding node in the enhanced node set based on its node number in the standard path structure and extracting its dynamic monitoring data. Dynamic monitoring data includes real-time power values, temperature monitoring values, and vibration monitoring values, which reflect the real-time operating status of the nodes. In this embodiment of the invention, for the standard path structure (1, 2, 3), node 1 corresponds to the blade node, node 2 corresponds to the gearbox node, and node 3 corresponds to the generator node. The real-time power values, temperature monitoring values, and vibration monitoring values ​​of the blade node, the gearbox node, and the generator node are extracted from the enhanced node set.

[0074] Step S335: Arrange the directed weighted edge weights and corresponding dynamic monitoring data of the nodes in the standard path structure in order to construct a three-dimensional sequence data array containing path identifiers, node data sequences, and edge weight sequences. The first dimension of the three-dimensional sequence data array is the path number, the second dimension is the position index on the path, and the third dimension is the data type index.

[0075] A path identifier is an identifier used to uniquely identify each major energy transfer path. Arranging the directed weighted edge weights and corresponding node dynamic monitoring data in the standard path structure in sequence means arranging the weights of the directed weighted edges and the dynamic monitoring data of the nodes according to their position on the path. The three-dimensional sequence data array is a data structure with three dimensions: the first dimension represents the path number, used to distinguish different major energy transfer paths; the second dimension represents the position index on the path, used to determine the position of each node and edge within the path; and the third dimension represents the data type index, used to distinguish different types of data, such as power values, temperature values, and edge weight values.

[0076] In this embodiment of the invention, it is assumed that there are p main energy transfer paths, each path has q nodes and q-1 directed weighted edges, and each node has r types of dynamic monitoring data. The first dimension of the three-dimensional sequence data array is p, the second dimension is q+(q-1), and the third dimension is r+1 (where 1 represents the edge weight value). For each path, the directed weighted edge weight values ​​and the corresponding node's dynamic monitoring data are arranged sequentially and placed into the three-dimensional sequence data array. For example, for the first path, the path is identified as 1, and the node data sequence and edge weight sequence on the path are arranged in the first row of the three-dimensional sequence data array according to their position index and data type index.

[0077] Step S340: Perform spatiotemporal feature extraction processing on the sequence data, calculate the correlation of energy conversion efficiency of each node on the path in the spatial dimension, calculate the sequence similarity of adjacent time windows in the time dimension, and generate a spatiotemporal joint feature matrix.

[0078] Spatiotemporal feature extraction involves comprehensively analyzing a three-dimensional sequence data array to extract its spatiotemporal features. In the spatial dimension, the correlation of energy conversion efficiency among nodes along the path is calculated, analyzing the relationship between the energy conversion efficiencies of adjacent nodes—for example, whether a change in the energy conversion efficiency of one node affects the energy conversion efficiencies of adjacent nodes. In the temporal dimension, the sequence similarity of adjacent time windows is calculated, comparing the similarity of the three-dimensional sequence data arrays of adjacent time windows to reflect the temporal stability of the wind turbine's energy conversion and transfer process. The spatiotemporal joint feature matrix is ​​a matrix that integrates the spatial dimension's energy conversion efficiency correlation and the temporal dimension's sequence similarity information, comprehensively reflecting the spatiotemporal characteristics of the wind turbine's energy conversion and transfer process.

[0079] In an optional implementation, step S340 can be performed as the following steps S341~S345: Step S341: Extract spatial features from each path data in the three-dimensional sequence data array, and calculate the energy conversion efficiency ratio of adjacent nodes on the path. The energy conversion efficiency is calculated by the ratio of node output power to input power.

[0080] Spatial feature extraction involves extracting spatially relevant feature information from a three-dimensional sequence data array. The energy conversion efficiency ratio of adjacent nodes on a path refers to the ratio of the energy conversion efficiencies of two adjacent nodes, reflecting the relative relationship between their energy conversion efficiencies. Energy conversion efficiency is calculated as the ratio of a node's output power to its input power and is an important indicator for measuring a node's energy conversion capability.

[0081] Step S342: Construct a path energy transfer efficiency matrix based on the energy conversion efficiency ratio, calculate the spatial autocorrelation coefficient of the elements in the matrix, and identify node pairs with abnormal fluctuations in energy transfer efficiency.

[0082] The path energy transfer efficiency matrix is ​​a matrix that arranges the energy conversion efficiency ratios along a path according to certain rules. The elements of the matrix represent the energy conversion efficiency ratios of adjacent nodes. The spatial autocorrelation coefficient is an indicator used to measure the spatial correlation between elements in the matrix, reflecting the relationship between the energy conversion efficiency ratios of adjacent nodes. Identifying node pairs with abnormal fluctuations in energy transfer efficiency refers to analyzing the spatial autocorrelation coefficients to find adjacent node pairs with abnormally changing energy transfer efficiency ratios; these node pairs may indicate anomalies in energy transfer.

[0083] In this embodiment of the invention, assuming there are n nodes on the path, the path energy transfer efficiency matrix is ​​an (n-1)×(n-1) matrix, where the elements of the matrix are the energy conversion efficiency ratios of adjacent nodes. The spatial autocorrelation coefficients of the elements in the matrix are calculated, for example using Moran's I index. By analyzing the spatial autocorrelation coefficients, it is found that the coefficient values ​​of some elements are abnormally large or small, indicating that the energy transfer efficiency ratios of the adjacent node pairs corresponding to these elements fluctuate abnormally, thus identifying these node pairs.

[0084] Step S343: Spatial location marking is performed on the identified abnormal fluctuation node pairs to generate spatial abnormal feature vectors. The dimension of the spatial abnormal feature vectors is consistent with the number of nodes on the path. The abnormal node pairs are marked as 1 at their corresponding locations and as 0 at other locations.

[0085] Spatial location marking of identified anomalous fluctuation node pairs refers to marking the locations of these pairs along the path. The spatial anomaly feature vector is a vector equal to the number of nodes on the path, used to represent the anomalous situation of each node. Anomalous node pairs are marked with a 1 at their corresponding locations, and other locations with a 0. This method provides a visual representation of the locations of anomalous nodes along the path.

[0086] Step S344: Perform time series comparison on the three-dimensional sequence data arrays of adjacent time windows, calculate the dynamic time curvature distance of the corresponding path, and use it as a sequence similarity index. The smaller the dynamic time curvature distance, the higher the sequence similarity.

[0087] Time series alignment involves comparing three-dimensional time series data arrays from adjacent time windows to analyze their similarity. Dynamic time warping distance (VTW) is a metric used to measure the similarity between two time series. It is calculated by finding the optimal matching path between the two series and then summing the distances between matching points along that path. A smaller VTW indicates a higher similarity between the two time series.

[0088] In this embodiment of the invention, for a three-dimensional sequence data array of adjacent time windows, it is assumed that there are p paths in each array. For each path, the dynamic time curvature distance of the corresponding path in adjacent time windows is calculated. For example, for the first path, the dynamic time curvature distance between the path data in the first time window and the path data in the second time window is calculated. The dynamic time curvature distance is used as a sequence similarity index to reflect the degree of similarity between the path data in adjacent time windows.

[0089] Step S345: Perform matrix fusion processing on the spatial anomaly feature vector and the sequence similarity index to construct a spatiotemporal joint feature matrix with the spatial dimension being the number of nodes and the temporal dimension being the number of windows. The matrix element values ​​represent the energy transfer characteristic index of the corresponding spatiotemporal location.

[0090] Matrix fusion processing integrates spatial anomaly feature vectors and sequence similarity indices into a single matrix. The spatial dimension of the spatiotemporal joint feature matrix represents the number of nodes, reflecting the information of each node along the path; the temporal dimension represents the number of windows, reflecting the information of different time windows. Matrix element values ​​represent energy transfer characteristics at the corresponding spatiotemporal location, such as node anomalies and sequence similarity.

[0091] In this embodiment of the invention, it is assumed that there are m time windows and n nodes on the path. The spatial anomaly feature vector and sequence similarity index of each time window are fused into an n×m spatiotemporal joint feature matrix. Each row of the matrix corresponds to a node, each column corresponds to a time window, and the matrix element value represents the energy transfer characteristic index of that node in that time window.

[0092] Step S350: Perform dimensional reduction on the spatiotemporal joint feature matrix, use principal component analysis to extract principal component components whose cumulative contribution rate exceeds a preset threshold, and construct a fixed-dimensional energy flow feature vector. The dimension of the energy flow feature vector is positively correlated with the number of components of the wind turbine.

[0093] Dimensionality reduction refers to reducing the dimensionality of a high-dimensional spatiotemporal joint feature matrix while retaining its main information. Principal component analysis (PCA) is a commonly used dimensionality reduction method. By identifying the principal components of the data, the original data is projected into the principal component space, thus achieving dimensionality reduction. The cumulative contribution rate refers to the cumulative proportion of each principal component's contribution to the variance of the original data. A preset threshold is a pre-defined cumulative contribution rate value used to filter out the main principal component components. A fixed-dimensional energy flow feature vector is a vector with fixed dimensions obtained after dimensionality reduction. Its dimension is positively correlated with the number of components in a wind turbine, effectively representing the energy flow characteristics of the wind turbine.

[0094] In this embodiment of the invention, assuming the spatiotemporal joint feature matrix is ​​an n×m matrix, principal component analysis (PCA) is used to reduce its dimensionality. The contribution rate and cumulative contribution rate of each principal component are calculated, with a preset threshold of 90%. Principal component components with a cumulative contribution rate exceeding 90% are extracted, and the spatiotemporal joint feature matrix is ​​projected onto these principal component components to obtain the dimensionality-reduced matrix. The dimension of the energy flow feature vector is determined based on the number of components in the wind turbine; for example, the more components, the higher the dimension of the energy flow feature vector. The dimensionality-reduced matrix is ​​then converted into a fixed-dimensional energy flow feature vector.

[0095] Step S400: Call the pre-built historical health state energy flow model to perform dynamic matching processing on the energy flow feature vector, and calculate the energy flow deviation between the current energy flow mode and the historical health mode.

[0096] The pre-built historical health state energy flow model refers to a pre-trained model used to describe the energy flow characteristics of wind turbines under historical health conditions. Dynamic matching involves comparing and matching the current energy flow feature vector with the historical health state energy flow model to find the most similar historical health pattern. Energy flow deviation refers to the degree of difference between the current energy flow pattern and the historical health pattern, reflecting any abnormalities in the current energy flow state of the wind turbine.

[0097] In an optional implementation, step S400 can be performed as the following steps S410~S450: Step S410: Construct a physical field linkage sensing network. The physical field linkage sensing network includes a mechanical vibration field sensing layer, an electromagnetic coupling field sensing layer, and a thermodynamic field sensing layer. Each sensing layer corresponds to the feature extraction channel of different physical fields of the wind turbine.

[0098] The physical field linkage sensing network is a network structure used to comprehensively sense the characteristics of different physical fields in a wind turbine. The mechanical vibration field sensing layer extracts features related to mechanical vibration in the wind turbine, such as bearing vibration spectrum characteristics and gearbox meshing frequency characteristics. The electromagnetic coupling field sensing layer extracts features related to electromagnetic coupling in the wind turbine, such as generator stator current characteristics and rotor magnetic field strength characteristics. The thermodynamic field sensing layer extracts features related to thermodynamics in the wind turbine, such as gearbox oil temperature characteristics and generator winding temperature characteristics. Each sensing layer corresponds to a feature extraction channel for different physical fields of the wind turbine, and can independently extract feature information for each physical field.

[0099] In this embodiment of the invention, the mechanical vibration field sensing layer of the physical field linkage sensing network can employ a convolutional neural network (CNN) to extract mechanical vibration features. The input layer of the CNN receives data collected by the mechanical vibration sensor, and after processing by convolutional layers, pooling layers, and fully connected layers, outputs bearing vibration spectrum features, gearbox meshing frequency features, etc. The electromagnetic coupling field sensing layer can employ a recurrent neural network (RNN) to extract electromagnetic coupling features. RNNs can process time-series data and output generator stator current features, rotor magnetic field strength features, etc. The thermodynamic field sensing layer can employ a multilayer perceptron (MLP) to extract thermodynamic features. The MLP, through processing by multiple layers of neurons, outputs gearbox oil temperature features, generator winding temperature features, etc.

[0100] Step S420: Decompose the energy flow feature vector into mechanical vibration feature sub-vectors, electromagnetic coupling feature sub-vectors, and thermodynamic feature sub-vectors, and input them into the corresponding sensing layer of the physical field linkage sensing network for feature enhancement processing.

[0101] Decomposing the energy flow feature vector into mechanical vibration feature sub-vectors, electromagnetic coupling feature sub-vectors, and thermodynamic feature sub-vectors means splitting the energy flow feature vector into three sub-vectors based on the feature information related to different physical fields. Inputting each sub-vector into the corresponding perceptual layer of the physical field linkage sensing network for feature enhancement means inputting each sub-vector into its corresponding perceptual layer. Through processing by the perceptual layer, the expressive power and discriminative power of the features are enhanced.

[0102] In this embodiment of the invention, assuming the dimension of the energy flow feature vector is d, it is decomposed into three sub-vectors based on the feature information related to mechanical vibration, electromagnetic coupling, and thermodynamics: a mechanical vibration feature vector (dimension d1), an electromagnetic coupling feature vector (dimension d2), and a thermodynamic feature vector (dimension d3), where d = d1 + d2 + d3. The mechanical vibration feature vector is input into the mechanical vibration field sensing layer of the physical field linkage sensing network, the electromagnetic coupling feature vector is input into the electromagnetic coupling field sensing layer, and the thermodynamic feature vector is input into the thermodynamic field sensing layer. Through processing by each sensing layer, the features are enhanced, for example, by increasing the feature dimension or improving the feature stability.

[0103] Step S430: Model the inter-field correlation of the enhanced features output by each sensing layer through the cross-field correlation layer of the physical field linkage sensing network, calculate the mutual information entropy between the mechanical vibration field and the electromagnetic coupling field, and the covariance matrix between the electromagnetic coupling field and the thermodynamic field, and generate the inter-field correlation feature tensor.

[0104] The cross-field correlation layer is a layer in the physical field linkage sensing network used to establish correlations between different physical fields. Modeling the inter-field correlations of the enhancement features output by each sensing layer involves analyzing the interrelationships between different physical fields and identifying their intrinsic connections. The mutual information entropy between the mechanical vibration field and the electromagnetic coupling field is an indicator used to measure the degree of information sharing between these two fields; a larger mutual information entropy value indicates a stronger correlation between the two fields. The covariance matrix between the electromagnetic coupling field and the thermodynamic field is a matrix used to measure the linear relationship between these two fields; the elements of the matrix represent the covariance between different feature dimensions of the two fields. The inter-field correlation feature tensor is a tensor that integrates the correlation information between the mechanical vibration field, the electromagnetic coupling field, and the thermodynamic field, and can comprehensively reflect the interrelationships between different physical fields.

[0105] In an optional implementation, step S430 can be performed as the following steps S431 to S436: Step S431: Extract the enhanced mechanical vibration feature sub-vectors from the mechanical vibration field sensing layer. The mechanical vibration feature sub-vectors include bearing vibration spectrum features, gearbox meshing frequency features, and blade flapping vibration features.

[0106] Extracting enhanced mechanical vibration feature sub-vectors from the mechanical vibration field sensing layer refers to obtaining enhanced mechanical vibration feature information from the output of the mechanical vibration field sensing layer. Bearing vibration spectrum features refer to the spectral distribution characteristics of the bearing during vibration, reflecting the bearing's operating state. Gearbox meshing frequency features refer to the frequency characteristics of the gearbox during meshing, capable of detecting gearbox faults. Blade flapping vibration features refer to the vibration characteristics of the blades during flapping, significantly impacting blade stability and safety.

[0107] In this embodiment of the invention, the mechanical vibration field sensing layer employs a convolutional neural network (CNN) for feature extraction and enhancement. Enhanced mechanical vibration feature sub-vectors are extracted from the output layer of the CNN. These sub-vectors include bearing vibration spectrum features, gearbox meshing frequency features, and blade flapping vibration features. For example, by processing data collected by vibration sensors, the CNN can extract the frequency components in the bearing vibration spectrum as bearing vibration spectrum features.

[0108] Step S432: Extract the enhanced electromagnetic coupling feature sub-vector from the electromagnetic coupling field sensing layer. The electromagnetic coupling feature sub-vector includes generator stator current characteristics, rotor magnetic field strength characteristics, and converter output harmonic characteristics.

[0109] Extracting enhanced electromagnetic coupling feature sub-vectors from the electromagnetic coupling field sensing layer refers to obtaining enhanced electromagnetic coupling feature information from the output of the electromagnetic coupling field sensing layer. Generator stator current characteristics refer to the waveform, frequency, and other characteristics of the generator stator current, reflecting the generator's power generation performance. Rotor magnetic field strength characteristics refer to the intensity distribution characteristics of the rotor magnetic field, which have a significant impact on the generator's efficiency and stability. Converter output harmonic characteristics refer to the harmonic components in the converter output current, enabling the detection of the converter's operating status.

[0110] In this embodiment of the invention, the electromagnetic coupling field sensing layer employs a recurrent neural network (RNN) for feature extraction and enhancement. Enhanced electromagnetic coupling feature sub-vectors are extracted from the RNN's output layer. These sub-vectors contain generator stator current features, rotor magnetic field strength features, and converter output harmonic features. For example, by processing the time-series data of the generator stator current, the waveform and frequency features of the current are extracted as the generator stator current features.

[0111] Step S433: Calculate the joint probability distribution and marginal probability distribution of the mechanical vibration feature vector and the electromagnetic coupling feature vector. Calculate the inter-field mutual information entropy based on the probability distribution. The larger the mutual information entropy value, the stronger the correlation between the two field features.

[0112] The joint probability distribution refers to the probability distribution of the simultaneous occurrence of mechanical vibration feature vectors and electromagnetic coupling feature vectors, while the marginal probability distribution refers to the probability distribution of the occurrence of each feature vector individually. Calculating the inter-field mutual information entropy based on the probability distribution involves calculating the mutual information entropy between the mechanical vibration field and the electromagnetic coupling field according to the joint and marginal probability distributions. A larger mutual information entropy value indicates a higher degree of information sharing and stronger correlation between the two field features.

[0113] Step S434: Extract the enhanced thermodynamic feature sub-vectors from the thermodynamic field sensing layer. The thermodynamic feature sub-vectors include gearbox oil temperature features, generator winding temperature features, and ambient temperature gradient features.

[0114] Extracting enhanced thermodynamic feature sub-vectors from the thermodynamic field sensing layer refers to acquiring enhanced thermodynamic feature information from the output of the thermodynamic field sensing layer. Gearbox oil temperature characteristics refer to the variation characteristics of oil temperature within the gearbox, reflecting the gearbox's lubrication and heat dissipation. Generator winding temperature characteristics refer to the temperature variation characteristics of the generator windings, which have a significant impact on the generator's insulation performance and lifespan. Ambient temperature gradient characteristics refer to the gradient variation characteristics of the ambient temperature surrounding the wind turbine, which can affect the wind turbine's operating performance.

[0115] In this embodiment of the invention, the thermodynamic field sensing layer employs a multilayer perceptron (MLP) for feature extraction and enhancement. Enhanced thermodynamic feature sub-vectors are extracted from the output layer of the MLP. These sub-vectors include gearbox oil temperature features, generator winding temperature features, and ambient temperature gradient features. For example, the MLP processes data collected by temperature sensors to extract the changing trend of the gearbox oil temperature, which is then used as the gearbox oil temperature feature.

[0116] Step S435: Calculate the covariance matrix of the electromagnetic coupling field and the thermodynamic field. The diagonal elements of the covariance matrix represent the variance of each characteristic dimension, and the off-diagonal elements represent the covariance between different field characteristic dimensions.

[0117] Calculating the covariance matrix of the electromagnetic coupling field and the thermodynamic field involves determining the covariance matrix between them based on the values ​​of the electromagnetic coupling eigenvectors and the thermodynamic eigenvectors. The diagonal elements of the covariance matrix represent the variance of each eigendimensional feature, reflecting the degree of dispersion of that feature dimension. The off-diagonal elements represent the covariance between different field eigendimensional features, reflecting the linear relationship between the two fields.

[0118] Step S436: Concatenate the mutual information entropy, covariance matrix and feature vectors of each field according to the channel dimension to generate a four-dimensional inter-field correlation feature tensor containing intra-field features and inter-field correlation information. The four dimensions correspond to the mechanical field dimension, electromagnetic field dimension, thermodynamic field dimension and spatiotemporal dimension, respectively.

[0119] Concatenating the mutual information entropy, covariance matrix, and field feature vectors along the channel dimension involves combining the mutual information entropy between the mechanical vibration field and the electromagnetic coupling field, the covariance matrix between the electromagnetic coupling field and the thermodynamic field, and the mechanical vibration feature vectors, electromagnetic coupling feature vectors, and thermodynamic feature vectors according to the channel dimension. This generates a four-dimensional inter-field correlation feature tensor containing intra-field features and inter-field correlation information. The four dimensions of this tensor correspond to the mechanical field dimension, electromagnetic field dimension, thermodynamic field dimension, and spatiotemporal dimension, respectively, and can comprehensively represent the correlation information and intra-field feature information between different physical fields of the wind turbine.

[0120] Step S440: Input the inter-field correlation feature tensor into the dynamic matching layer of the historical health state energy flow model, and perform dynamic time warping matching with the multi-physics health benchmark feature library stored in the model to generate an inter-field feature similarity matrix.

[0121] The dynamic matching layer is used for feature matching in the historical health state energy flow model. The inter-field correlation feature tensor is input into the dynamic matching layer and dynamically time-warped (DTW) matched with samples in the health benchmark feature library stored in the model. This means that the current inter-field correlation feature tensor is matched with samples in the health benchmark feature library using the DTW algorithm to find the most similar samples. The inter-field feature similarity matrix is ​​a matrix used to represent the degree of similarity between the current inter-field correlation feature tensor and each sample in the health benchmark feature library.

[0122] In an optional implementation, step S440 can be specifically implemented as the following steps S441 to S447: Step S441: Load the multiphysics health benchmark feature library from the historical health state energy flow model. The feature library contains a sample set of health state field correlation feature tensors divided into different wind speed ranges. Each sample set corresponds to the stable operating state within the corresponding wind speed range.

[0123] Loading a multiphysics health benchmark feature library from a historical health state energy flow model refers to reading pre-stored multiphysics health benchmark feature information from the model. The feature library contains sample sets of health state inter-field correlation feature tensors divided into different wind speed ranges. This means that the inter-field correlation feature tensors of the health state are classified according to wind speed ranges, with each sample set corresponding to a wind speed range, reflecting the stable operating status of the wind turbine within that wind speed range.

[0124] In this embodiment of the invention, it is assumed that the historical health state energy flow model stores a health state field correlation feature tensor sample set for multiple wind speed ranges. Each sample set contains multiple field correlation feature tensor samples, which are extracted from the historical health operation data of wind turbines in the corresponding wind speed range.

[0125] Step S442: Extract the real-time collected wind speed environmental parameters, and based on the wind speed environmental parameters, select a target healthy sample subset that matches the current wind speed range from the multi-physics health benchmark feature library. The target healthy sample subset contains health status data for at least three consecutive operating cycles.

[0126] Extracting real-time acquired wind speed environmental parameters refers to obtaining the current wind speed data from environmental monitoring devices. Selecting a target healthy sample subset matching the current wind speed range from a multi-physics health benchmark feature library based on wind speed environmental parameters means identifying a sample set corresponding to the current wind speed range from the health benchmark feature library as the target healthy sample subset based on the current wind speed data. The target healthy sample subset contains health status data from at least three consecutive operating cycles to ensure the stability and reliability of the samples.

[0127] Step S443: Perform time axis scaling and alignment processing on the current inter-field correlation feature tensor and each sample tensor in the target healthy sample subset. The time series of different lengths are mapped to the same dimension space by the dynamic time warping algorithm to eliminate the time scale difference.

[0128] Performing time-axis scaling and alignment on the current inter-field correlation feature tensor and each sample tensor in the target healthy sample subset refers to adjusting the time axis length of the current inter-field correlation feature tensor and the sample tensor so that they have the same length in the time dimension. Mapping time series of different lengths to the same dimensional space using the Dynamic Time Warping (DTW) algorithm means using the DTW algorithm to find the optimal matching path between two time series, aligning their time scales, and eliminating time scale differences.

[0129] In this embodiment of the invention, it is assumed that the time series length of the current inter-field correlation feature tensor is T1, and the time series length of a sample tensor in the target healthy sample subset is T2, and T1 and T2 may be different. A dynamic time warping algorithm is employed. This algorithm searches for an optimal matching path between the time series of the current inter-field correlation feature tensor and the sample tensor, minimizing the cumulative distance between the two time series along that path. In this way, time series of different lengths are mapped to the same dimensional space, much like adjusting rulers of different lengths to the same scale range, thereby eliminating time scale differences and ensuring the accuracy of subsequent matching. For example, if the current inter-field correlation feature tensor shows dense data changes over a certain period, while the sample tensor shows relatively sparse data in the corresponding time period, the dynamic time warping algorithm will reasonably stretch or compress the time axis so that the two can accurately correspond in the time dimension.

[0130] Step S444: In the dynamic time warping matching process, a physical field weighting factor is introduced, assigning a first weight value to the mechanical vibration field characteristic dimension, a second weight value to the electromagnetic coupling field characteristic dimension, and a third weight value to the thermodynamic field dimension, wherein the first weight value is greater than the second and third weight values.

[0131] In dynamic time warping matching, the characteristics of different physical fields may have varying degrees of importance in determining the health status of wind turbines. Therefore, a physical field weighting factor is introduced to assign different weights to the characteristic dimensions of different physical fields. The operating state of the mechanical vibration field often has a critical impact on the overall stability of the wind turbine; for example, abnormal bearing vibration may indicate mechanical failure. Therefore, the mechanical vibration field characteristic dimension is assigned the first weight value. The characteristics of the electromagnetic coupling field reflect the operating conditions of electrical components such as generators, and its importance is secondary, so it is assigned the second weight value. The characteristics of the thermodynamic field mainly reflect the temperature changes of the equipment, and its importance is slightly lower than the first two, so it is assigned the third weight value, with the first weight value being greater than both the second and third weight values. This allows for a more accurate highlighting of the importance of different physical field characteristics during dynamic time warping matching, improving the accuracy of the matching.

[0132] Step S445: Calculate the cumulative distance between the current inter-field correlation feature tensor and each healthy sample based on the weighted dynamic time warping path. The cumulative distance is obtained by weighted summation of the Euclidean distances of each matching point on the path.

[0133] Based on the dynamically time-warped path weighted by the physical field weighting factor, the cumulative distance between the current inter-field correlation feature tensor and each healthy sample is calculated. For each matching point on the dynamically time-warped path, the Euclidean distance between its corresponding feature vectors is calculated. Then, these Euclidean distances are weighted according to the physical field weighting factor, and finally, the weighted Euclidean distances of all matching points are summed to obtain the cumulative distance. This cumulative distance reflects the degree of difference between the current inter-field correlation feature tensor and the healthy samples; the smaller the cumulative distance, the more similar they are.

[0134] Step S446: Normalize the cumulative distance. Construct a normal distribution model using the mean and standard deviation of the cumulative distances of all samples in the target healthy sample subset, and calculate the similarity probability value corresponding to the current cumulative distance.

[0135] Normalizing the calculated cumulative distances ensures comparability between different healthy samples. A normal distribution model is constructed using the mean and standard deviation of the cumulative distances of all samples in the target healthy sample subset, assuming that the cumulative distances follow a normal distribution. Based on the properties of the normal distribution, the probability value corresponding to the current cumulative distance within that distribution can be calculated; this probability value is the similarity probability value. A higher similarity probability value indicates that the current inter-field correlation feature tensor is more similar to that of healthy samples, and the current operating state of the wind turbine is closer to a healthy state.

[0136] Step S447: Decompose the similarity probability value according to the physical field feature dimension to generate an inter-field similarity matrix containing mechanical vibration field similarity vector, electromagnetic coupling field similarity vector and thermodynamic field similarity vector. The row dimension of the matrix corresponds to the health sample number, the column dimension corresponds to the physical field feature dimension, and the matrix element value is the normalized similarity probability value.

[0137] The obtained similarity probability values ​​are decomposed according to the physical field feature dimensions to obtain similarity vectors for the mechanical vibration field, electromagnetic coupling field, and thermodynamic field. The mechanical vibration field similarity vector reflects the degree of similarity between the mechanical vibration field features in the current inter-field correlation feature tensor and the mechanical vibration field features in the healthy samples. The electromagnetic coupling field similarity vector and the thermodynamic field similarity vector are obtained similarly. These vectors are combined into an inter-field similarity matrix, where the row dimensions correspond to different healthy sample numbers, the column dimensions correspond to different physical field feature dimensions, and the matrix elements are normalized similarity probability values. Through the inter-field similarity matrix, the similarity between the current inter-field correlation feature tensor and each healthy sample in different physical field dimensions can be visually observed.

[0138] Step S450: Calculate the contribution weight coefficient of each physical field based on the inter-field feature similarity matrix, obtain the comprehensive similarity index by weighted summation, and define the difference obtained by subtracting the comprehensive similarity index from 1 as the energy flow deviation.

[0139] The inter-field feature similarity matrix reflects the degree of similarity between the current inter-field correlation feature tensor and healthy samples across different physical field dimensions. To obtain a comprehensive index to measure the difference between the current energy flow state and the healthy state of the wind turbine, the contribution weight coefficients of each physical field need to be calculated. These contribution weight coefficients can be determined based on the degree of influence of different physical fields on the overall performance of the wind turbine. For example, mechanical vibration fields have a significant impact on the stability of the wind turbine, and their contribution weight coefficient may be relatively high. By weighting and summing the elements in the inter-field feature similarity matrix according to the contribution weight coefficients of each physical field, a comprehensive similarity index is obtained. The higher the comprehensive similarity index, the closer the current energy flow state of the wind turbine is to a healthy state. The difference between 1 and the comprehensive similarity index is the energy flow deviation, reflecting the degree of deviation between the current energy flow pattern of the wind turbine and the historical healthy pattern.

[0140] Step S500: Optimize the anomaly judgment threshold based on the energy flow deviation and the real-time collected wind speed environmental parameters, and output the wind turbine performance anomaly detection result based on the optimized anomaly judgment threshold.

[0141] Energy flow deviation reflects the degree of difference between the current energy flow state of a wind turbine and its historical healthy pattern, while wind speed environmental parameters have a significant impact on the operating status of wind turbines. Under different wind speed conditions, the normal operating range and energy flow characteristics of wind turbines may vary. Therefore, combining energy flow deviation with real-time collected wind speed environmental parameters to optimize the anomaly detection threshold can make anomaly detection more accurate and reasonable. Based on the optimized anomaly detection threshold, the current operating status of the wind turbine is compared with it, thereby outputting the wind turbine performance anomaly detection result and determining whether the wind turbine has an anomaly and its severity.

[0142] In an optional implementation, step S500 can be performed as the following steps S510~S560: Step S510: Construct a multi-physics field coupling threshold optimization model, decompose the energy flow deviation into mechanical vibration field deviation component, electromagnetic coupling field deviation component and thermodynamic field deviation component, and establish the nonlinear mapping relationship between each physical field deviation component and wind speed environmental parameters.

[0143] The multiphysics coupling threshold optimization model is used to comprehensively consider the interactions between different physical fields of a wind turbine and the influence of wind speed environmental parameters. The energy flow deviation is decomposed into mechanical vibration field deviation components, electromagnetic coupling field deviation components, and thermodynamic field deviation components because the operating states of different physical fields may contribute differently to the energy flow deviation. For example, anomalies in the mechanical vibration field may increase energy loss during mechanical transmission, thus affecting energy flow; problems in the electromagnetic coupling field may affect the generator's power generation efficiency, thereby altering the energy flow pattern; anomalies in the thermodynamic field may affect the equipment's performance and stability, also impacting energy flow. A nonlinear mapping relationship is established between each physical field deviation component and wind speed environmental parameters because the relationship between wind speed environmental parameters and each physical field deviation component is not a simple linear relationship, but a complex nonlinear one. By constructing such a nonlinear mapping relationship, the influence of wind speed environmental parameters on each physical field deviation component can be described more accurately, providing a more precise basis for optimizing the anomaly detection threshold.

[0144] In an optional implementation, step S510 can be specifically implemented as the following steps S511~S515: Step S511: Perform physical field source decomposition on the energy flow deviation, and separate the mechanical vibration field feature vector, electromagnetic coupling field feature vector and thermodynamic field feature vector through independent component analysis algorithm. Each feature vector contains the time domain evolution characteristics of the corresponding physical field.

[0145] Physical field source decomposition refers to tracing the deviation of energy flow back to the influencing factors of different physical fields. Independent component analysis (ICA) algorithms, used to separate independent components in mixed signals, are used here to decompose the energy flow deviation into mechanical vibration field eigenvectors, electromagnetic coupling field eigenvectors, and thermodynamic field eigenvectors. Each eigenvector contains the time-domain evolution characteristics of the corresponding physical field, reflecting the changes of each physical field over time. For example, the mechanical vibration field eigenvector might contain the variation of bearing vibration frequency over time, the electromagnetic coupling field eigenvector might contain the fluctuation of generator stator current over time, and the thermodynamic field eigenvector might contain the trend of gearbox oil temperature over time. Through this decomposition, the specific contribution and changes of each physical field to the energy flow deviation can be clearly understood.

[0146] Step S512: Collect wind speed environmental parameters and physical field feature vector data under historical healthy conditions, and construct a three-dimensional training sample set. The sample dimensions include wind speed, time, and physical field features.

[0147] Collecting historical wind speed environmental parameters and physical field feature vector data under healthy conditions is to obtain relevant information about wind turbines under normal operating conditions. Wind speed environmental parameters reflect the external environmental conditions at that time, while physical field feature vector data reflects the operating status of each physical field. This data is organized according to three dimensions: wind speed, time, and physical field features, to construct a three-dimensional training sample set. This training sample set contains physical field feature information of the wind turbine under different wind speed conditions and at different time points, providing a rich data foundation for subsequently constructing nonlinear mapping functions. For example, at different wind speeds, feature vector data of each physical field is collected at regular intervals, and this data is integrated into the three-dimensional training sample set.

[0148] Step S513: Use the gradient boosting tree algorithm to construct a nonlinear mapping function between the deviation components of each physical field and the wind speed environmental parameters. The input of the mapping function is wind speed, wind direction angle and wind speed change rate, and the output is the predicted value of the deviation component of the corresponding physical field.

[0149] Gradient boosting tree algorithm is an ensemble learning algorithm that builds a powerful prediction model by iteratively training multiple decision trees. In this embodiment of the invention, the gradient boosting tree algorithm is used to construct a nonlinear mapping function between the deviation components of each physical field and wind speed environmental parameters. The inputs to this mapping function include wind speed, wind direction angle, and wind speed change rate, which comprehensively describe the characteristics of the wind speed environment. The output is the predicted value of the deviation component of the corresponding physical field, that is, the deviation component of each physical field is predicted based on the input wind speed environmental parameters. Through this nonlinear mapping function, the complex relationship between wind speed environmental parameters and the deviation components of each physical field can be accurately reflected, providing accurate prediction results for subsequent threshold optimization.

[0150] Step S514: Calculate the sum of squared residuals of the mapping functions of each physical field, identify the cross-influence terms between physical fields through residual analysis, and generate mechanical-electromagnetic coupling terms, electromagnetic-thermal coupling terms, and mechanical-thermal coupling terms. The coupling terms are represented by the product terms of the dual-field characteristics.

[0151] The sum of squared residuals of each physical field mapping function refers to the sum of squared differences between the actual deviation components of the physical field and the deviation components predicted by the mapping function. By calculating the sum of squared residuals, the fitting effect of the mapping function can be evaluated. Residual analysis is an in-depth study of residuals to discover hidden information that may exist in the data. In this embodiment of the invention, residual analysis is used to identify cross-influence terms between physical fields. Different physical fields may have mutual influences. For example, the vibration of a mechanical vibration field may affect the electromagnetic induction of an electromagnetic coupling field, and changes in the electromagnetic coupling field may affect the temperature distribution of a thermodynamic field. By analyzing the residuals, these cross-influence terms are discovered, generating mechanical-electromagnetic coupling terms, electromagnetic-thermal coupling terms, and mechanical-thermal coupling terms. These coupling terms are represented by product terms of the characteristics of the two fields. For example, the mechanical-electromagnetic coupling term can be represented as the product of the characteristics of the mechanical vibration field and the characteristics of the electromagnetic coupling field, reflecting the interaction between the two physical fields.

[0152] Step S515: Superimpose the main effect mapping function with the cross-influence term to construct a multi-physics coupling threshold optimization model. The model output includes the total deviation prediction value of each physical field deviation component and the inter-field coupling deviation component. The determination coefficient of the deviation from the actual energy flow is kept above the preset threshold.

[0153] The main effect mapping function is a nonlinear mapping function between the deviation components of each physical field and the wind speed environmental parameters, constructed using the gradient boosting tree algorithm. It describes the direct impact of wind speed environmental parameters on the deviation components of each physical field. The cross-influence term reflects the interaction between different physical fields. The main effect mapping function and the cross-influence term are superimposed to construct a multi-physics coupling threshold optimization model. The output of this model includes the predicted total deviation of each physical field deviation component and the inter-field coupling deviation component. This predicted value comprehensively considers the deviation of each physical field itself and the mutual influence between fields. To ensure the accuracy of the model, the coefficient of determination between the predicted total deviation and the actual energy flow deviation must remain above a preset threshold. The coefficient of determination is an indicator of the model's goodness of fit; a higher coefficient of determination indicates a better fit to the actual data and more accurate prediction results.

[0154] Step S520: Extract the distribution characteristics of the physical field deviation components under different wind speed ranges from historical health data, and generate the mechanical vibration field reference threshold curve, the electromagnetic coupling field reference threshold curve, and the thermodynamic field reference threshold curve. Each curve shows the differential fluctuation characteristics of wind speed correlation.

[0155] Historical health data contains various information about wind turbines under normal operating conditions. The distribution characteristics of physical field deviation components under different wind speed ranges are extracted from this data; that is, the distribution of deviation components of the mechanical vibration field, electromagnetic coupling field, and thermodynamic field under different wind speed conditions is analyzed. Based on these distribution characteristics, reference threshold curves for the mechanical vibration field, electromagnetic coupling field, and thermodynamic field are generated. These curves reflect the normal range of each physical field deviation component under different wind speeds. Because different physical fields respond differently to wind speed, each curve exhibits differentiated fluctuation characteristics related to wind speed. For example, the reference threshold curve for the mechanical vibration field may fluctuate less at low wind speeds and more at high wind speeds; the reference threshold curve for the electromagnetic coupling field may show a gradually increasing trend with increasing wind speed. These reference threshold curves can provide a reference for anomaly detection under different wind speed conditions.

[0156] Step S530: Calculate the reference thresholds of each physical field corresponding to the real-time wind speed environmental parameters, and perform weighted fusion of the multiple physical field reference thresholds through the inter-field correlation weight matrix to generate the initial comprehensive anomaly judgment threshold. The inter-field correlation weight matrix is ​​dynamically updated through the physical field importance evaluation algorithm.

[0157] Based on real-time collected wind speed environmental parameters, the corresponding physical field benchmark thresholds are retrieved from the mechanical vibration field benchmark threshold curve, electromagnetic coupling field benchmark threshold curve, and thermodynamic field benchmark threshold curve. An inter-field correlation weight matrix is ​​used to represent the degree of correlation and importance between different physical fields. The inter-field correlation weight matrix is ​​dynamically updated using a physical field importance assessment algorithm. This algorithm adjusts the weights of each physical field in real time based on the actual operating conditions of the wind turbine and the impact of different physical fields on overall performance. The inter-field correlation weight matrix is ​​used to perform weighted fusion of the multi-physical field benchmark thresholds, that is, the benchmark thresholds of each physical field are weighted and summed according to their corresponding weights to generate an initial comprehensive anomaly judgment threshold. This initial comprehensive anomaly judgment threshold comprehensively considers the influence of different physical fields and the inter-field correlation, providing a preliminary standard for subsequent anomaly judgment.

[0158] Step S540: Introduce the time series trend characteristics of energy flow deviation. When the deviation of three consecutive time windows shows a monotonically increasing trend, activate the threshold elastic contraction mechanism to reduce the initial comprehensive anomaly judgment threshold according to a preset ratio.

[0159] The time-series trend of energy flow deviation reflects how the deviation changes over time. When the deviation shows a monotonically increasing trend for three consecutive time windows, it indicates that the energy flow state of the wind turbine is gradually deteriorating, potentially indicating an anomaly risk. At this point, a threshold elastic contraction mechanism is activated, lowering the initial comprehensive anomaly judgment threshold by a preset ratio. Lowering the threshold makes anomaly detection more sensitive, enabling timely detection of abnormal changes in the wind turbine and preventing further deterioration of the situation. For example, if the initial comprehensive anomaly judgment threshold is high, when the energy flow deviation monotonically increases for three consecutive time windows, the threshold is lowered by a certain ratio, allowing anomalies to be judged even at lower deviation levels.

[0160] Step S550: Adjust the threshold elastic contraction range by combining the wind speed fluctuation coefficient. The wind speed fluctuation coefficient is calculated by the wind speed standard deviation within the sliding window. The larger the fluctuation coefficient, the larger the contraction range. Generate an optimized anomaly judgment threshold that integrates multi-dimensional constraints.

[0161] The wind speed fluctuation coefficient reflects the fluctuation of wind speed within a certain time range and is calculated using the standard deviation of wind speed within a sliding window. A larger standard deviation indicates more severe wind speed fluctuations. Adjusting the threshold elastic contraction range in conjunction with the wind speed fluctuation coefficient is because severe wind speed fluctuations can significantly impact the operating status of wind turbines, increasing the likelihood of anomalies. When the wind speed fluctuation coefficient is large, increasing the threshold elastic contraction range further lowers the anomaly detection threshold, making anomaly detection more stringent. In this way, an optimized anomaly detection threshold integrating multi-dimensional constraints is generated, comprehensively considering the time-series trend of energy flow deviation and the impact of wind speed fluctuations, thus improving the accuracy and reliability of anomaly detection.

[0162] Step S560: Compare the current energy flow deviation with the optimization anomaly judgment threshold. When the deviation exceeds the threshold, trigger the anomaly classification process, determine the dominant factor of the anomaly based on the proportion of deviation components of each physical field, and generate a performance anomaly detection result containing the anomaly dominant field identifier.

[0163] The calculated energy flow deviation is compared with the optimized anomaly detection threshold. If the energy flow deviation exceeds the threshold, it indicates a potential anomaly in the wind turbine's operating status. This triggers an anomaly grading process to assess the severity of the anomaly. The dominant anomaly factor is determined based on the proportion of deviation components in each physical field. This involves analyzing the percentage of deviation components from the mechanical vibration field, electromagnetic coupling field, and thermodynamic field within the total deviation; the physical field with the largest proportion is identified as the dominant anomaly field. Performance anomaly detection results containing the dominant anomaly field identifier are generated. These results clearly indicate the main cause and severity of the wind turbine anomaly, providing crucial information for subsequent maintenance and handling.

[0164] In an optional implementation, step S560 can be performed as the following steps S561~S565: Step S561: When the energy flow deviation exceeds the optimization anomaly judgment threshold, calculate the percentage of the mechanical vibration field deviation component, electromagnetic coupling field deviation component, and thermodynamic field deviation component in the total deviation, and determine the physical field with the highest percentage as the dominant anomaly field.

[0165] When the energy flow deviation exceeds the optimization anomaly detection threshold, further analysis is performed on the energy flow deviation. The percentages of the mechanical vibration field deviation component, the electromagnetic coupling field deviation component, and the thermodynamic field deviation component in the total deviation are calculated. By comparing these percentages, the physical field with the highest percentage is identified as the dominant anomaly field. For example, if the mechanical vibration field deviation component accounts for the largest proportion of the total deviation, then the mechanical vibration field is the dominant anomaly field, indicating that the anomaly of the wind turbine is likely mainly caused by mechanical vibration issues.

[0166] Step S562: Extract the abnormal sensitive feature dimension from the inter-field correlation feature tensor based on the abnormal dominant field. The abnormal sensitive feature dimension is obtained by filtering the feature variation coefficient. Dimensions with a variation coefficient greater than a preset value are marked as key abnormal features.

[0167] The inter-field correlation feature tensor contains correlation information and features between different physical fields of the wind turbine. Extracting anomaly-sensitive feature dimensions from the inter-field correlation feature tensor based on the dominant anomaly field aims to identify feature dimensions that are highly sensitive to anomaly changes and are related to the dominant anomaly field. These anomaly-sensitive feature dimensions are obtained through feature variation coefficient (CV), which measures the dispersion of feature data. When the CV of a feature dimension is greater than a preset value, it indicates that the feature dimension changes significantly under abnormal conditions, and it is marked as a key anomaly feature. For example, if the dominant anomaly field is a mechanical vibration field, feature dimensions related to the mechanical vibration field are extracted from the inter-field correlation feature tensor, the CV of these dimensions is calculated, and dimensions with CVs greater than a preset value are marked as key anomaly features.

[0168] Step S563: Map the key anomaly feature dimensions to the nodes and edges of the dynamic energy topology graph, identify the abnormal energy conversion components and abnormal energy transmission paths, and construct an anomaly propagation path model, which includes the anomaly starting node, propagation direction, and path attenuation coefficient.

[0169] Mapping key anomaly feature dimensions to nodes and edges in a dynamic energy topology graph allows for a visual representation of the anomaly's location and propagation. Nodes in the dynamic energy topology graph represent the energy conversion components of a wind turbine, while edges represent energy transfer paths between these components. By associating key anomaly feature dimensions with the nodes and edges of the topology graph, anomalous energy conversion components and anomalous energy transfer paths are identified. For example, if a key anomaly feature dimension is associated with a generator node, then the generator node is likely an anomalous energy conversion component; if this feature dimension is associated with an edge from the generator to the transformer, then the energy transfer path corresponding to this edge is likely an anomalous energy transfer path. An anomaly propagation path model is constructed, which includes the anomaly's origin node, propagation direction, and path attenuation coefficient. The anomaly origin node is the source of the anomaly, the propagation direction indicates the direction of the anomaly's propagation along the energy transfer path, and the path attenuation coefficient indicates the degree of attenuation of the anomaly during propagation. This model provides a clear understanding of the anomaly's propagation mechanism and scope of impact.

[0170] Step S564: Calculate the energy transfer efficiency attenuation rate of each node on the anomaly propagation path, and combine it with the deviation component value of the anomaly dominant field to assess the severity level of the anomaly. The level classification is based on the betweenness centrality index of the anomaly component in the energy topology diagram.

[0171] The energy transfer efficiency decay rate at each node along the anomaly propagation path reflects the degree of impact of the anomaly on the node's energy conversion efficiency. Calculating the energy transfer efficiency decay rate for each node involves comparing the changes in energy conversion efficiency before and after the anomaly occurs. The severity of the anomaly is then comprehensively assessed by combining this with the deviation component values ​​of the dominant anomaly field. A larger deviation component value indicates a more severe impact. The severity level is determined by the betweenness centrality index of the anomaly component in the energy topology graph. The betweenness centrality index represents a node's ability to connect other nodes as an intermediate node in the energy topology graph. A higher betweenness centrality index indicates a more important node in the energy transfer process, and the anomaly's impact on it may lead to more serious consequences. Based on these factors, the severity of the anomaly is classified into different levels, such as mild, moderate, and severe anomalies.

[0172] Step S565: Generate wind turbine performance anomaly detection results including anomaly dominant field identifier, unique code of anomaly component, anomaly propagation path diagram, severity level, and suggested intervention time window. The suggested intervention time window is calculated by the anomaly development rate prediction model.

[0173] The generated wind turbine performance anomaly detection results contain several key pieces of information: the dominant field identifier clearly identifies the main source of the anomaly; the unique code of the anomalous component accurately locates it; the anomaly propagation path diagram visually illustrates the propagation of the anomaly; the severity level helps assess the degree of harm caused by the anomaly; and the recommended intervention time window provides maintenance personnel with a time reference for taking action. The recommended intervention time window is calculated using an anomaly development rate prediction model. This model can be trained based on historical anomaly data and current anomaly conditions. By analyzing the development trend and speed of the anomaly, it predicts the time required for the anomaly to reach its severity level, thereby determining the recommended intervention time window.

[0174] The various algorithms described above in the embodiments of this invention can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on further in the embodiments of this invention. Furthermore, those skilled in the art can supplement the details when implementing the solutions of this invention based on common knowledge in the field. For example, normalization or standardization can be used to eliminate dimensional conflicts before feature fusion, based on common knowledge in the field. For example, when "introducing a physical field weighting factor in the dynamic time warping matching process, assigning a first weight value to the mechanical vibration field feature dimension, a second weight value to the electromagnetic coupling field feature dimension, and a third weight value to the thermodynamic field feature dimension", the inter-field correlation feature tensor can be standardized to make each feature dimension have zero mean and unit variance; the physical field weighting factor is introduced to assign a first weight value to the mechanical vibration field feature dimension, a second weight value to the electromagnetic coupling field feature dimension, and a third weight value to the thermodynamic field feature dimension.

[0175] Furthermore, interpolation can be used to eliminate dimensional differences, thresholds can be reasonably set based on historical data, experience, or business scenario requirements, models can be trained using general model training methods, the number of layers in the model structure can be set according to actual needs, and activation functions can be selected, etc. In addition, to save computational overhead and training costs, those skilled in the art can select general models to implement the functions of the models listed in this invention, provided that existing general models can be used.

[0176] This invention will no longer provide redundant details about the implementation process.

[0177] Please see Figure 3 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. Figure 3As shown, the computer system 1000 described above may include: a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the computer system 1000 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The user interface 1003 may include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. The memory 1005 may also optionally be at least one storage device located remotely from the processor 1001. Figure 3 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application program. Figure 3 In the computer system 1000 shown, the network interface 1004 provides network communication functions; the user interface 1003 is mainly used to provide an input interface; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the methods provided in the above embodiments.

Claims

1. A method for detecting abnormal performance of wind turbine generators, characterized in that, The method includes: The operating status dataset of the wind turbine is collected during a continuous operating cycle. The operating status dataset is collected synchronously by sensor arrays and environmental monitoring devices distributed on key components of the wind turbine. Based on the aforementioned operating status dataset, a dynamic energy topology map reflecting the energy conversion and transfer process of the wind turbine is constructed. The nodes of the dynamic energy topology map correspond to the energy conversion components of the wind turbine, and the edges correspond to the energy transfer paths between the components. The operational status dataset is mapped to the dynamic energy topology map, and energy flow feature vectors characterizing the energy transfer properties between components are extracted through topology analysis. The energy flow feature vector is dynamically matched by calling a pre-built historical health state energy flow model, and the energy flow deviation between the current energy flow mode and the historical health mode is calculated. The anomaly detection threshold is optimized based on the energy flow deviation and the real-time collected wind speed environmental parameters, and the wind turbine performance anomaly detection result is output based on the optimized anomaly detection threshold.

2. The method according to claim 1, characterized in that, The construction of a dynamic energy topology map reflecting the energy conversion and transfer process of wind turbine units based on the operating status dataset includes: The operating status dataset is processed to extract component attributes, identify energy conversion components and their physical attribute parameters in the wind turbine, and establish a set of topological nodes containing unique component identifiers, type classification labels and rated energy conversion efficiency parameters. Based on the energy conversion type and physical connection relationship of each component in the topological node set, the energy transfer power data between components is extracted from the operating status dataset, and a set of directed weighted edges describing the energy transfer direction and intensity is constructed. The weight value of the directed weighted edge is calculated by the ratio of real-time transfer power to rated transfer power. According to the preset time window division rules, the continuous running cycle is divided into sliding windows to generate multiple static topology snapshots of time windows. Each static topology snapshot contains a set of topology nodes and a set of directed weighted edges within the corresponding time window. A topological structure difference analysis is performed on static topological snapshots of adjacent time windows, the change rate of node attributes and the gradient of edge weight change are calculated, and a dynamic energy topology graph containing time dimension evolution characteristics is constructed. The time resolution of the dynamic energy topology graph is consistent with the sampling frequency of the sensor array.

3. The method according to claim 2, characterized in that, Based on the energy conversion type and physical connection relationship of each component in the topological node set, the energy transfer power data between components is extracted from the operating status dataset, and a directed weighted edge set describing the energy transfer direction and intensity is constructed, including: The energy conversion components in the topological node set are hierarchically divided. According to the energy transfer path of the wind turbine, from the wind energy capture component to the power output component, they are sequentially labeled as the primary conversion layer, the secondary conversion layer, and the sensor output layer, generating a node hierarchical topology structure containing hierarchical attributes. Based on the physical connection relationships of each component in the node hierarchy topology, an initial undirected connection graph is constructed, where the edges in the initial undirected connection graph represent direct physical connections between components. The input power data and output power data of each component are extracted from the operating status dataset. The power transfer difference between adjacent level components is calculated. When the proportion of the power transfer difference to the output power of the upstream component exceeds a preset threshold, an effective energy transfer relationship is determined. Based on the effective energy transfer relationship, the initial undirected connection graph is converted into a directed graph, with the direction of the directed edge pointing from the energy output component to the energy input component, and the power transfer difference is used as the initial weight value of the directed edge. The initial weight values ​​are normalized, and the standardized weight values ​​are calculated by the ratio of the current transmission power to the designed maximum transmission power, generating a set of directed weighted edges containing the standardized weight values.

4. The method according to claim 2, characterized in that, The process of performing topological structure difference analysis on static topological snapshots of adjacent time windows, calculating the rate of change of node attributes and the gradient of edge weight changes, and constructing a dynamic energy topology graph containing time-dimensional evolution features includes: Node matching is performed on static topology snapshots of two consecutive adjacent time windows. The node correspondence across time windows is established through the unique identifier of the component, and newly added nodes, removed nodes, and nodes with changed attributes are identified. The physical attribute parameters of the attribute change nodes are subjected to time series difference processing to calculate the parameter change per unit time and generate node attribute change rate index, which includes temperature change rate, vibration frequency change rate and efficiency change rate. For each directed edge in the set of directed weighted edges, the weight difference is calculated to obtain the change in edge weight between adjacent time windows, and the gradient of edge weight change is calculated in combination with the time window interval. The sign of the gradient of edge weight change indicates the direction of increase or decrease of the weight value. Construct a topological evolution feature set that includes a node attribute change rate matrix and an edge weight change gradient tensor, wherein the dimension of the topological evolution feature set matches the number of nodes and edges in the static topological snapshot; The set of topological nodes, the set of directed weighted edges, and the set of topological evolution features are subjected to spatiotemporal correlation modeling to generate a dynamic energy topology graph that includes static node attributes, static edge weights, and temporal evolution features. The dynamic energy topology graph is stored using a spatiotemporal tensor data structure.

5. The method according to claim 1, characterized in that, The step of mapping the operating status dataset to the dynamic energy topology map and extracting energy flow feature vectors characterizing the energy transfer properties between components through topology analysis includes: The operating status dataset is timestamped to ensure that the data collected by the sensor array is consistent with the time resolution of the dynamic energy topology map, thereby generating a time-synchronized operating status data sequence. The monitoring data of each component in the time-synchronized operating status data sequence are allocated to the topology nodes corresponding to the dynamic energy topology map to generate an enhanced node set with real-time monitoring values. The enhanced node set includes the static physical attributes and dynamic monitoring data of the nodes. The dynamic energy topology map is subjected to topology traversal processing. Starting from the wind energy capture component, the weight values ​​of the directed weighted edges and the dynamic monitoring data of the corresponding nodes are extracted sequentially along the energy transfer direction to construct the sequence data on the energy transfer path. The sequence data is subjected to spatiotemporal feature extraction processing. The correlation of energy conversion efficiency of each node on the path is calculated in the spatial dimension, and the sequence similarity of adjacent time windows is calculated in the temporal dimension to generate a spatiotemporal joint feature matrix. The spatiotemporal joint feature matrix is ​​subjected to dimensional reduction processing, and principal component components with cumulative contribution rates exceeding a preset threshold are extracted to construct a fixed-dimensional energy flow feature vector. The dimension of the energy flow feature vector is positively correlated with the number of components of the wind turbine.

6. The method according to claim 5, characterized in that, The process of performing a topological traversal on the dynamic energy topology map involves extracting the weight values ​​of directed weighted edges and the dynamic monitoring data of corresponding nodes sequentially along the energy transfer direction, starting from the wind energy capture component, to construct sequence data along the energy transfer path, including: Using the wind energy capture component node in the dynamic energy topology graph as the starting node, the directed weighted edges are traversed to record all complete energy transfer paths from the starting node to the sensor output node. The path length of each complete energy transfer path is calculated. The path length is defined as the number of directed weighted edges contained in the path. The main energy transfer paths with path lengths within a preset range are retained, and abnormal paths containing anomalous jumps are filtered out. The main energy transfer paths that are retained are standardized, and the nodes on the paths are renumbered according to the energy transfer order to generate a standard path structure containing a node sequence and a corresponding directed weighted edge sequence. Dynamic monitoring data of each node in the standard path structure is extracted from the enhanced node set. The dynamic monitoring data includes real-time power value, temperature monitoring value and vibration monitoring value. The directed weighted edge weights and corresponding dynamic monitoring data of the nodes in the standard path structure are arranged in order to construct a three-dimensional sequence data array containing path identifiers, node data sequences, and edge weight sequences. The first dimension of the three-dimensional sequence data array is the path number, the second dimension is the position index on the path, and the third dimension is the data type index.

7. The method according to claim 5, characterized in that, The process of extracting spatiotemporal features from the sequence data includes calculating the correlation of energy conversion efficiency among nodes along the path in the spatial dimension, calculating the sequence similarity of adjacent time windows in the temporal dimension, and generating a spatiotemporal joint feature matrix, comprising: Spatial features are extracted from each path data in the three-dimensional sequence data array, and the energy conversion efficiency ratio of adjacent nodes on the path is calculated. The energy conversion efficiency is obtained by the ratio of node output power to input power. Based on the energy conversion efficiency ratio, a path energy transfer efficiency matrix is ​​constructed, the spatial autocorrelation coefficient of the elements in the matrix is ​​calculated, and node pairs with abnormal fluctuations in energy transfer efficiency are identified. The identified abnormal fluctuation node pairs are spatially marked to generate spatial abnormal feature vectors. The dimension of the spatial abnormal feature vectors is consistent with the number of nodes on the path. The abnormal node pairs are marked as 1 at their corresponding positions and as 0 at other positions. Time series comparison is performed on three-dimensional sequence data arrays of adjacent time windows, and the dynamic time curvature distance of the corresponding path is calculated as a sequence similarity index. The smaller the dynamic time curvature distance, the higher the sequence similarity. The spatial anomaly feature vector and the sequence similarity index are fused into a matrix to construct a spatiotemporal joint feature matrix with the spatial dimension being the number of nodes and the temporal dimension being the number of windows. The matrix element values ​​represent the energy transfer characteristics index of the corresponding spatiotemporal location.

8. The method according to claim 1, characterized in that, The process of dynamically matching the energy flow feature vector by invoking a pre-built historical health state energy flow model, and calculating the energy flow deviation between the current energy flow pattern and the historical health pattern, includes: A physical field linkage sensing network is constructed, which includes a mechanical vibration field sensing layer, an electromagnetic coupling field sensing layer, and a thermodynamic field sensing layer. Each sensing layer corresponds to the feature extraction channel of different physical fields of the wind turbine. The energy flow feature vector is decomposed into mechanical vibration feature sub-vectors, electromagnetic coupling feature sub-vectors, and thermodynamic feature sub-vectors, and then input into the sensing layer of the physical field linkage sensing network for feature enhancement processing. By modeling the inter-field correlation of the enhanced features output by each sensing layer through the cross-field correlation layer of the physical field linkage sensing network, the mutual information entropy of the mechanical vibration field and the electromagnetic coupling field, and the covariance matrix of the electromagnetic coupling field and the thermodynamic field are calculated to generate the inter-field correlation feature tensor. The inter-field correlation feature tensor is input into the dynamic matching layer of the historical health state energy flow model, and is dynamically time-warped and matched with the multi-physics field health benchmark feature library stored in the model to generate an inter-field feature similarity matrix. The contribution weight coefficients of each physical field are calculated based on the inter-field feature similarity matrix. The comprehensive similarity index is obtained by weighted summation. The difference obtained by subtracting the comprehensive similarity index from 1 is defined as the energy flow deviation.

9. The method according to claim 1, characterized in that, The process of optimizing the anomaly detection threshold based on the energy flow deviation and real-time collected wind speed environmental parameters, and outputting the wind turbine performance anomaly detection result based on the optimized anomaly detection threshold, includes: A multi-physics field coupling threshold optimization model is constructed, which decomposes the energy flow deviation into mechanical vibration field deviation components, electromagnetic coupling field deviation components, and thermodynamic field deviation components, and establishes a nonlinear mapping relationship between each physical field deviation component and wind speed environmental parameters. The distribution characteristics of the physical field deviation components under different wind speed ranges are extracted from historical health data to generate mechanical vibration field reference threshold curves, electromagnetic coupling field reference threshold curves and thermodynamic field reference threshold curves. Each curve shows the differential fluctuation characteristics of wind speed correlation. Calculate the reference thresholds of each physical field corresponding to the real-time wind speed environmental parameters, and perform weighted fusion of the multiple physical field reference thresholds through the inter-field correlation weight matrix to generate an initial comprehensive anomaly judgment threshold. The inter-field correlation weight matrix is ​​dynamically updated through the physical field importance evaluation algorithm. Introducing the time series trend characteristics of energy flow deviation, when the deviation shows a monotonically increasing trend for three consecutive time windows, the threshold elastic contraction mechanism is activated to reduce the initial comprehensive anomaly judgment threshold according to a preset ratio. The threshold elastic contraction amplitude is adjusted by combining the wind speed fluctuation coefficient. The wind speed fluctuation coefficient is calculated by the wind speed standard deviation within the sliding window. The larger the fluctuation coefficient, the larger the contraction amplitude, thus generating an optimized anomaly judgment threshold that integrates multi-dimensional constraints. The current energy flow deviation is compared with the optimization anomaly judgment threshold. When the deviation exceeds the threshold, the anomaly classification process is triggered. The dominant factor of the anomaly is determined according to the proportion of deviation components of each physical field, and a performance anomaly detection result containing the anomaly dominant field identifier is generated.

10. A computer system, characterized in that, include: processor; And a memory, wherein the memory stores computer-readable code that, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 9.

Images