Artificial intelligence-based wind turbine fault classification mining method and system

Abstract

The application provides an artificial intelligence-based wind turbine fault classification mining method and system, and relates to the technical field of data mining.The industrial cloud platform is used to obtain a time sequence alarm sequence of a wind turbine in a safety chain triggering process and a working condition data set within a preset time length before the safety chain triggering process; a state deviation matrix is obtained based on the logic state of different position points in the time sequence alarm sequence; the alarm nodes and the corresponding time information are respectively associated with the state deviation matrix and spliced to obtain a joint encoding matrix; the alarm nodes in the joint encoding matrix are taken as nodes, and the causal logic path between the alarm nodes is taken as a directed edge to construct a structure feature vector; and the target fault category corresponding to the structure feature vector is determined to minimize the logarithmic loss between the structure feature vector and a preset discriminant function set, thereby improving the fault classification precision and diagnosis reliability of the wind turbine under complex multi-source data.

Description

Technical Field

[0001] This application relates to the field of data mining technology, and in particular to a method and system for classifying and mining faults in wind turbine generators based on artificial intelligence. Background Technology

[0002] With the continuous expansion of global wind power generation, improving the reliability of wind turbine operation and reducing operation and maintenance costs have become core demands of the industry. Utilizing advanced monitoring data through industrial cloud platforms for intelligent fault classification and analysis is a key technological means to achieve proactive predictive maintenance of turbines and ensure long-term stable returns for wind farms. Industrial cloud platforms break down the data silos of traditional wind farms, providing fundamental support for the centralized management of massive operational data and the training of high-precision artificial intelligence models.

[0003] However, existing solutions are typically limited to using operating parameters or status codes collected locally by the SCADA system, and constructing a mapping relationship between data and fault types through deep neural networks or traditional classification models. Some solutions further attempt to integrate multi-dimensional operating indicators and alarm status bits, aiming to improve the accuracy of identifying typical unit fault modes through feature stacking of multi-source information.

[0004] However, existing solutions focus on single-dimensional numerical trend analysis, making it difficult to capture the profound causal evolution relationships between alarm signals from different subsystems during the safety chain triggering process, resulting in insufficient ability to represent the dynamic propagation process of faults. This neglect of the logical topological characteristics of fault triggering makes it difficult for the model to accurately uncover the core causes behind the alarm sequence when facing highly coupled alarm scenarios. Therefore, existing technologies suffer from low classification accuracy and insufficient diagnostic reliability in complex fault triggering scenarios. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for classifying and mining wind turbine faults based on artificial intelligence, so as to solve the technical problems of low classification accuracy and insufficient diagnostic reliability in complex fault triggering scenarios in the prior art.

[0006] Firstly, this application provides an artificial intelligence-based method for classifying and mining faults in wind turbine generators, including:

[0007] The timing alarm sequence of the wind turbine during the safety chain triggering process and the operating condition dataset within a preset time before the safety chain triggering process are obtained. The timing alarm sequence includes the logical state combination of different subsystems in the wind turbine at multiple time points.

[0008] Based on the logical state of different locations in the time-series alarm sequence, the alarm node sequence and the time information corresponding to each alarm node are determined. The limit parameters of the wind turbine are subjected to boundary processing to obtain the limit constraint set. The Euclidean distance between each parameter in the working condition dataset and the corresponding limit threshold in the limit constraint set is calculated to obtain the state deviation matrix.

[0009] The alarm node and its corresponding time information are spatiotemporally correlated with the state deviation matrix and then concatenated to obtain the joint coding matrix.

[0010] Using alarm nodes in the joint coding matrix as nodes and causal logical paths between alarm nodes as directed edges, a directed acyclic graph is constructed. By calculating the global importance and centrality of each node in the directed acyclic graph, a structural feature vector is constructed.

[0011] With the goal of minimizing the log loss between the structural feature vector and the preset discriminant function set, the regression coefficient matrix in the discriminant function set is iteratively optimized to determine the target fault category corresponding to the structural feature vector.

[0012] Optionally, based on the logical states of different locations in the time-series alarm sequence, the alarm node sequence and the time information corresponding to each alarm node are determined. Boundary processing is then performed on the limit parameters of the wind turbine to obtain a set of limit constraints. The Euclidean distance between each parameter in the operating condition dataset and the corresponding limit threshold in the set of limit constraints is calculated to obtain a state deviation matrix, including:

[0013] The location where the logical state changes in the timing alarm sequence is taken as the alarm node. According to the order of the alarm nodes in the timing alarm sequence, all alarm nodes are arranged into an alarm node sequence, and the time when the logical state changes in each alarm node in the alarm node sequence is taken as the corresponding time information.

[0014] The operating conditions of the wind turbine are divided into intervals to obtain multiple operating condition intervals. In each operating condition interval, a limit threshold corresponding to each limit parameter of the wind turbine is set. All limit thresholds are combined to obtain a set of limit constraints.

[0015] The limit constraint set and the working condition dataset are mapped to a preset multidimensional feature space to calculate the Euclidean distance between each parameter in the working condition dataset and the corresponding limit threshold. According to the sampling time of each parameter in the working condition dataset, all Euclidean distances are arranged into a matrix to obtain the state deviation matrix.

[0016] Optionally, the alarm node and its corresponding time information are spatiotemporally correlated with the state deviation matrix and then concatenated to obtain a joint coding matrix, including:

[0017] The state deviation matrix is ​​retrieved to determine the Euclidean distance between the time information corresponding to each alarm node and the time information at the same moment, thus obtaining the instantaneous offset corresponding to each alarm node.

[0018] Each alarm node, its corresponding time information, and instantaneous offset are arranged horizontally to obtain a comprehensive attribute vector for each alarm node.

[0019] According to the chronological order of the time information corresponding to each alarm node, all comprehensive attribute vectors are arranged vertically to obtain the joint coding matrix.

[0020] Optionally, a directed acyclic graph (DAG) is constructed using alarm nodes in the joint encoding matrix as nodes and causal logical paths between alarm nodes as directed edges. A structural feature vector is then constructed by calculating the global importance and centrality of each node in the DAG, including:

[0021] Based on the trigger sequence of the comprehensive attribute vectors in the joint coding matrix and the preset logical topology between each subsystem in the wind turbine, the causal logical path between alarm nodes is determined.

[0022] Construct a directed acyclic graph using alarm nodes as nodes and causal logical paths as directed edges;

[0023] The global importance of each node is obtained by summing the number of out-degrees of each node in the directed acyclic graph, and the centrality of each node is obtained by averaging the topological distances between each node and the remaining nodes in the directed acyclic graph.

[0024] Following the order of alarm nodes in the alarm node sequence, the global importance and centrality of all nodes are concatenated, and the concatenation result is truncated or padded to a fixed length to obtain the structural feature vector.

[0025] Optionally, the discriminant function set includes sub-discriminant functions corresponding to physical component over-limit faults, external load impact faults, sensor calibration errors, control software deadlocks, and communication protocol anomalies. The goal is to minimize the logarithmic loss between the structural feature vector and the preset discriminant function set. The regression coefficient matrix in the discriminant function set is iteratively optimized to determine the target fault category corresponding to the structural feature vector, including:

[0026] Probability mapping is performed between the structural feature vector and the regression coefficient matrix in the discriminant function set to obtain the category probability of each fault category. The fault categories include physical component over-limit faults, external load impact faults, sensor calibration errors, control software deadlocks, and communication protocol anomalies.

[0027] The distribution difference between the category probability and the preset label value is compared to obtain the log loss. With the goal of reducing the log loss value, each weight parameter of the regression coefficient matrix is ​​iteratively updated to obtain the optimized regression coefficient matrix.

[0028] The target prediction sequence is obtained by performing a dot product operation on the structural feature vector and the optimized regression coefficient matrix. The target fault category is determined based on the classification label corresponding to the largest component in the target prediction sequence.

[0029] Optionally, a probability mapping is performed between the structural feature vector and the regression coefficient matrix in the discriminant function set to obtain the class probability of each fault category, including:

[0030] The initial score for each fault category is obtained by performing a dot product operation between the structural feature vector and the regression coefficient matrix.

[0031] The initial score is raised to an exponent to obtain the mapped score. The ratio of each mapped score to the sum of all mapped scores is then calculated to obtain the category probability corresponding to each fault category.

[0032] Optionally, the distribution difference between the class probabilities and the preset label values ​​is compared to obtain the log loss. With the goal of reducing the log loss, each weight parameter of the regression coefficient matrix is ​​iteratively updated to obtain the optimized regression coefficient matrix, including:

[0033] The log loss is obtained by calculating the distribution difference between the category probability and the preset label value;

[0034] Calculate the partial derivative of each weight parameter in the regression coefficient matrix to obtain the rate of change of the log loss with respect to each weight parameter, and determine the rate of change as the gradient component;

[0035] Adjust each weight parameter in the opposite direction of the gradient components, and then perform a dot product operation between the structural feature vector and the updated regression coefficient matrix until the regression coefficient matrix converges or reaches the preset number of iterations, thus obtaining the optimized regression coefficient matrix.

[0036] Secondly, this application provides an artificial intelligence-based wind turbine fault classification and mining system, including:

[0037] The acquisition module is used to acquire the timing alarm sequence of the wind turbine during the safety chain triggering process and the operating condition dataset within a preset time before the safety chain triggering process. The timing alarm sequence includes the logical state combination of different subsystems in the wind turbine at multiple time points.

[0038] The determination module is used to determine the alarm node sequence and the time information corresponding to each alarm node based on the logical state of different locations in the time-series alarm sequence, perform boundary processing on the limit parameters of the wind turbine to obtain the limit constraint set, and calculate the Euclidean distance between each parameter in the operating condition dataset and the corresponding limit threshold in the limit constraint set to obtain the state deviation matrix.

[0039] The association module is used to spatiotemporally associate alarm nodes and their corresponding time information with the state deviation matrix and then concatenate them to obtain a joint coding matrix.

[0040] The module is used to construct a directed acyclic graph (DAG) with alarm nodes in the joint encoding matrix as nodes and causal logical paths between alarm nodes as directed edges. It also constructs a structural feature vector by calculating the global importance and centrality of each node in the DAG.

[0041] The iterative module is used to iteratively optimize the regression coefficient matrix in the discriminant function set with the goal of minimizing the log loss between the structural feature vector and the preset discriminant function set in order to determine the target fault category corresponding to the structural feature vector.

[0042] Thirdly, this application provides an electronic device, comprising:

[0043] Memory, used to store computer programs;

[0044] A processor is used to execute computer programs to implement the steps of the artificial intelligence-based wind turbine fault classification and mining method described in the first aspect above.

[0045] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the steps of the artificial intelligence-based wind turbine fault classification and mining method described in the first aspect above.

[0046] The AI-based wind turbine fault classification and mining method provided in this application leverages the massive data integration and real-time interaction capabilities of an industrial cloud platform to simultaneously acquire discrete state switching signals at the logical level and continuous operating parameters at the physical level. This overcomes the bottlenecks of limited computing power and storage in traditional local data centers, ensuring the simultaneous capture of abnormal turbine performance from both logical evolution and physical state deviation dimensions. It solves the problem of difficulty in quantifying the severity of anomalies through single numerical trend analysis, transforming raw operating data into a standard matrix that reflects the degree to which the system deviates from safety boundaries. It overcomes the limitation of existing technologies in capturing the causal evolution relationship between alarm signals. This improves classification accuracy and diagnostic reliability in highly coupled alarm scenarios, providing a reliable solution for large-scale clustered intelligent monitoring of wind turbines based on industrial cloud platforms.

[0047] Furthermore, this application extracts the trigger sequence relationship of the comprehensive attribute vector in the joint encoding matrix and combines it with the preset logical topological relationship between each subsystem of the wind turbine to jointly determine the causal logical path between alarm nodes. Then, a directed acyclic graph is constructed with alarm nodes as vertices and causal logical paths as directed edges. The global importance is obtained by accumulating the out-degree of each node in the graph, and the centrality is obtained by calculating the mean topological distance between nodes. Finally, the global importance and centrality of all nodes are concatenated according to the order of the alarm node sequence, and after fixed-length truncation or padding, a standardized structural feature vector is constructed. This solves the problem of insufficient performance caused by existing solutions ignoring the topological characteristics of alarm logic. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating the artificial intelligence-based wind turbine fault classification and mining method provided in this application embodiment;

[0050] Figure 2 A flowchart illustrating the method for obtaining the state deviation matrix provided in an embodiment of this application;

[0051] Figure 3 A flowchart illustrating the method for constructing structural feature vectors provided in this application embodiment;

[0052] Figure 4 A schematic diagram of the structure of the artificial intelligence-based wind turbine fault classification and mining system provided in the embodiments of this application;

[0053] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0054] To address the technical bottleneck of existing solutions that over-rely on single-dimensional numerical trends and struggle to capture the causal evolution relationship between alarm signals, leading to inaccurate fault cause identification, this application introduces Euclidean distance to measure the deviation of operating parameters from limit boundaries. This transforms the ambiguous operating state into a quantified state deviation matrix, solving the problem of insufficient characterization of anomaly severity in existing solutions. Subsequently, by spatiotemporally associating time-series alarm sequences with state deviations and constructing a directed acyclic graph, the analysis dimension is elevated from simple numerical fluctuations to the fault causal propagation topology. Finally, the core structural characteristics of fault evolution are extracted using global importance and centrality indices in graph theory, forming a highly recognizable structural feature vector. This fundamentally overcomes the diagnostic reliability defects caused by existing solutions ignoring logical evolution relationships, achieving accurate determination and classification of complex fault triggering causes.

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] The core of this application is to provide an artificial intelligence-based method for classifying and identifying faults in wind turbines. A flowchart illustrating one specific implementation is shown below. Figure 1 As shown, the method includes:

[0057] Step 101: Obtain the timing alarm sequence of the wind turbine during the safety chain triggering process and the operating condition dataset within a preset time period before the safety chain triggering process. The timing alarm sequence includes the logical state combination of different subsystems in the wind turbine at multiple time points.

[0058] In this step, a wind turbine refers to a mechatronic device that captures wind energy and converts it into electrical energy. The safety chain triggering process refers to the dynamic response process where the unit detects abnormal operating conditions that may threaten hardware safety or component lifespan, and the hardware control loop or monitoring logic executes a protective shutdown. The timing alarm sequence refers to a set of codes or signals arranged in chronological order of alarm events within the observation time window. The preset duration refers to the backtracking sampling time step set for analyzing fault precursors, which can be 600 seconds before the fault trigger moment. The operating condition dataset refers to a set of digitized information recording the physical state of the unit's operation, including wind speed environmental parameters, shaft physical parameters, and generator operating indicators. Subsystems refer to the various independent functional units that constitute the unit, including the pitch subsystem, yaw subsystem, and main shaft subsystem. The logical state combination refers to the Boolean arrangement of alarm signals corresponding to all subsystems in the spatial dimension at the same sampling moment, with 0 representing normal operation and 1 representing alarm triggering.

[0059] In this embodiment, various sensor signals and register bit signals are first read in real time by the industrial controller inside the unit, and the data is continuously aggregated and uploaded to the industrial cloud platform for centralized storage and management via an edge computing gateway or real-time communication network. When a safety chain action is detected to be activated, the action time is used as the basis for calculation. Based on this, the time range obtained covers from the backtracking starting point, i.e. Subtract the preset time from the complete cycle of fault diagnosis to ensure the working condition data set is accurate. Completely includes time-series alarm sequences in the time dimension. All alarm times. The acquired data includes two categories. The first category is the time-series alarm sequence, captured in real-time through the controller's log recording and received by the cloud node. This data structure records multiple sampling points. Combinations of logical states generated by different subsystems The second category is operating condition datasets retrieved from the distributed database of the industrial cloud platform. This dataset records continuous physical operating parameters synchronized with the alarm time, avoiding retrieval failures caused by non-overlapping data time periods.

[0060] For example, taking an overspeed shutdown fault of unit B in region A as an example, the preset duration is set to 600 seconds. The industrial cloud platform first calls its cloud storage resources to obtain the speed data within the 600 seconds prior to the fault. Torque and paddle angle The constructed working condition dataset Next, obtain the sequence of timing alarms generated during the safety chain triggering. The sequence is recorded at time [time]. The pitch module detected an anomaly, and the logic state combination was established. Represented as 1,0,0. In subsequent moments... The spindle module also generates an alarm, at which point the logic state combination... Updated to 1,1,0. The final result will be... and This data is integrated and used as the original data source for subsequently establishing the association coding matrix.

[0061] Step 102: Based on the logical state of different locations in the time-series alarm sequence, determine the alarm node sequence and the time information corresponding to each alarm node, perform boundary processing on the limit parameters of the wind turbine to obtain the limit constraint set, and calculate the Euclidean distance between each parameter in the operating condition dataset and the corresponding limit threshold in the limit constraint set to obtain the state deviation matrix.

[0062] In this step, a location point refers to the discrete sampled coordinates recording the logical state in the time-series alarm sequence. A logical state refers to the operational attribute exhibited by the subsystem at a specific moment, including normal operation state 0 and alarm trigger state 1. An alarm node refers to a critical location point where the logical state switches from 0 to 1. An alarm node sequence refers to an ordered set of multiple alarm nodes arranged in chronological order. Time information refers to the state transition time corresponding to the alarm node. Limit parameters refer to key physical variables used to monitor the safe operation of wind turbine units, which may include generator speed, main shaft torque, and pitch rate. The limit constraint set refers to a standard numerical space composed of limit thresholds corresponding to multiple limit parameters. Limit thresholds refer to the safe boundary values ​​allowed for operation by the limit parameters, including preset upper and lower safety limits for each limit parameter. Euclidean distance refers to the degree of geometric deviation between the operating parameters and the limit thresholds after mapping to a multi-dimensional feature space. The state deviation matrix is ​​a two-dimensional data structure formed by arranging the Euclidean distances corresponding to all operating condition data sampling points along a time axis.

[0063] like Figure 2 As shown, Figure 2 This is a flowchart illustrating the method for obtaining the state deviation matrix provided in an embodiment of this application.

[0064] Step 201: Take the location point where the logic state changes in the timing alarm sequence as the alarm node. According to the order of the alarm nodes in the timing alarm sequence, arrange all alarm nodes into an alarm node sequence, and take the time when the logic state changes in each alarm node in the alarm node sequence as the corresponding time information.

[0065] In this step, the sequential position refers to the order in which the alarm nodes are arranged relative to the starting sampling time within the observation time window.

[0066] In this embodiment of the application, the timing alarm sequence is first traversed. Capture the changes in logical states. Specifically, identify the transition points where the status bits of each subsystem flip from 0 to 1 and mark them as alarm nodes. Next, based on each alarm node... They are stacked in an orderly manner according to their order of appearance.

[0067] For example, taking a fault in unit B in region A as an example, if at time The pitch subsystem triggers an alarm at a specific time. If the yaw subsystem triggers an alarm, the generated alarm node sequence can be represented as follows: Finally, the time information corresponding to each node in the sequence is extracted. Specifically, the corresponding time information sequence can be represented as follows: .

[0068] Step 202: Divide the operating condition range of the wind turbine into intervals to obtain multiple operating condition intervals, and set a limit threshold corresponding to each limit parameter of the wind turbine in each operating condition interval. Combine all the limit thresholds to obtain a limit constraint set.

[0069] In this step, the operating condition range refers to the physical boundaries within which the wind turbine can maintain safe operation under different environmental constraints. Interval partitioning refers to the operation of subdividing the operating condition range based on specific control variables.

[0070] In this embodiment, the historical operating records of the wind turbine are first retrieved to determine the operating condition range. Then, the operating condition range is divided into intervals based on wind speed or power output status.

[0071] For example, the operating conditions are divided into low wind speed zones. and rated wind speed range Within each obtained operating condition range, corresponding limit thresholds are set for generator speed, main shaft torque, and blade angle. Specifically, within the rated wind speed range... Inside, set the limit threshold vector. ,in Indicates the speed threshold. Indicates the torque threshold. This indicates the paddle angle threshold.

[0072] For each limiting parameter, there is an upper threshold. and a lower threshold To eliminate the differences in dimensions and magnitudes between speed measured in revolutions per minute (rpm), torque measured in Newton-meters (Nm), and angle measured in degrees (°), this application first utilizes the mean-variance normalization method to normalize the operating condition dataset. The original parameters are dimensionless to transform them into standard normal distribution data.

[0073] When calculating the Euclidean distance, for each normalized parameter... The corresponding instantaneous offset calculation method is improved as follows: the smaller of the absolute value of the difference between this parameter and the upper threshold, and the absolute value of the difference between this parameter and the lower threshold, is selected as the one-dimensional deviation. The calculation formula is as follows: 1. This calculation method ensures that, regardless of whether parameters exceed or fall short of limits, the geometric degree of deviation of the physical state from the safety boundary can be accurately quantified, and all parameters have the same weight and influence in the multidimensional feature space. Finally, the limit threshold vectors set within all operating condition intervals are combined to obtain a set of limit constraints used to determine the degree of deviation of the physical state.

[0074] Step 203: Map the limit constraint set and the working condition dataset to a preset multidimensional feature space to calculate the Euclidean distance between each parameter in the working condition dataset and the corresponding limit threshold. Arrange all Euclidean distances into a matrix according to the sampling time of each parameter in the working condition dataset to obtain the state deviation matrix.

[0075] In this step, the pre-defined multidimensional feature space refers to a linear metric space constructed by using multiple limiting parameters as orthogonal bases.

[0076] In this embodiment of the application, relying on the distributed computing power of the industrial cloud platform, the working condition dataset is first... The real-time physical parameters and the limit thresholds in the limit constraint set are jointly mapped to a preset multi-dimensional feature space. For example, the time interval is... Operating condition parameter vector With the corresponding limit threshold vector They are placed within the same multidimensional feature space. Then, using spatial geometric calculation methods, the Euclidean distance between the operating parameters and the corresponding limit thresholds at each sampling time is calculated in the cloud. The calculation formula is shown in formula (1):

[0077] (1)

[0078] in, Represents Euclidean distance. Indicates the first Real-time observed values ​​of each limiting parameter Indicates the first The limit threshold corresponding to each limiting parameter This represents the total number of limiting parameters. Then, following the chronological order of sampling times, the one-dimensional deviation values ​​calculated for each parameter at each sampling time are horizontally concatenated and vertically stacked along the time axis to obtain the state deviation matrix. ,in This represents the total number of sampling points. Specifically, the state deviation matrix. The number of rows is the total number of sampling points. The number of columns is the total number of limiting parameters. Under this representation, Each row represents the system offset distribution of all physical parameters at a sampling time, rather than a single scalar value.

[0079] Step 103: After spatiotemporally associating the alarm node and its corresponding time information with the state deviation matrix, concatenate them to obtain the joint coding matrix.

[0080] In this step, the joint coding matrix refers to the feature representation obtained by multidimensionally concatenating discrete logical alarm features with continuous physical offset features.

[0081] Step 301: Search the state deviation matrix to determine the Euclidean distance at the same time as the time information corresponding to each alarm node, and obtain the instantaneous offset corresponding to each alarm node.

[0082] In this step, "same moment" refers to the point on the time axis where the logical state switch of the alarm node coincides with the data sampling time point recorded in the state deviation matrix. Instantaneous offset refers to the quantized geometric distance of the unit's operating parameters relative to the limit boundary at the instant the alarm node is triggered.

[0083] In this embodiment of the application, firstly, based on each alarm node in the alarm node sequence... and its corresponding time information In the state deviation matrix Data retrieval is performed within the system. Specifically, a timestamp indexing algorithm is used to locate time information. Data records at the same time.

[0084] For example, taking a fault in unit B in region A as an example, if the identified pitch subsystem alarm node The corresponding time information is By retrieving the state deviation matrix The Euclidean distance value corresponding to this moment is determined as follows: Then The instantaneous offset corresponding to the alarm node is determined. Repeating the above retrieval process, the resulting sequence of instantaneous offsets can be represented as follows: ,in This indicates the total number of alarm nodes.

[0085] Step 302: Arrange each alarm node, its corresponding time information, and instantaneous offset horizontally to obtain the comprehensive attribute vector of each alarm node.

[0086] In this step, the comprehensive attribute vector refers to a one-dimensional, multi-column data feature composed of the alarm node's identification information, trigger time, and physical offset degree.

[0087] In this embodiment, the identifier of each alarm node is first obtained. and the corresponding time information and instantaneous offset Next, the heterogeneous data is processed by feature concatenation. For example, for the first alarm node... Encode its identifier and the time of occurrence and offset values Arrange them horizontally. The resulting comprehensive attribute vector. .

[0088] In this way, the event information of the logic layer and the degree of deviation of the state of the physical layer are initially bound together by data dimensions, forming a feature unit that can fully describe the attributes of a single alarm event.

[0089] Step 303: Arrange all comprehensive attribute vectors vertically according to the chronological order of the time information corresponding to each alarm node to obtain the joint coding matrix.

[0090] In this embodiment, the comprehensive attribute vector of all alarm nodes generated during the current fault process is first determined. Next, according to the time information corresponding to each node in the alarm node sequence... The order in which all comprehensive attribute vectors are stacked vertically in an ordered manner.

[0091] For example, the first generated composite attribute vector As the first row of the matrix, the second generated comprehensive attribute vector As the second row of the matrix, and so on, the joint encoding matrix is ​​finally obtained through this ordered stacking. The generated joint encoding matrix It can be represented in the following matrix format:

[0092]

[0093] in, Represents the joint encoding matrix, Indicates the first One alarm node, Indicates the first This is a time-related information. Indicates the first Instantaneous offset. Joint encoding matrix. The number of rows is determined by the number of alarm nodes, and each row fully represents an alarm event and its associated physical offset characteristics.

[0094] Step 104: Using the alarm nodes in the joint encoding matrix as nodes and the causal logical paths between alarm nodes as directed edges, construct a directed acyclic graph, and construct a structural feature vector by calculating the global importance and centrality of each node in the directed acyclic graph.

[0095] In this step, a node refers to a vertex in a directed acyclic graph (DAG) that represents a specific alarm event. A causal logical path refers to the direction of propagation between alarm nodes based on trigger timing and physical logical relationships. A directed edge is an arrow connecting two nodes in the graph, representing a causal logical path. A DAG is a loop-free topology structure composed of alarm nodes and directed edges. Global importance is an indicator of a node's influence in the entire fault evolution network, which can include the number of subsequent alarms triggered by the node. Centrality refers to the structural coreness of a node in the network topology, which can be the average topological distance from the node to all other nodes. The structural feature vector is a mathematical vector obtained by fixing the graph's topological attributes to a fixed dimension.

[0096] like Figure 3 As shown, Figure 3 This is a flowchart illustrating the method for constructing structural feature vectors provided in an embodiment of this application.

[0097] Step 401: Determine the causal logical path between alarm nodes based on the trigger sequence of the comprehensive attribute vectors in the joint coding matrix and the preset logical topology relationship between each subsystem in the wind turbine.

[0098] In this step, the trigger sequence refers to the order in which alarm nodes are triggered on the time axis based on corresponding time information. The preset logical topology refers to the inherent signal transmission paths or control affiliation rules between the various subsystems of the wind turbine unit. It is abstracted from the control loop logic diagram of the unit. Specifically, the logical topology defines the subsystems... With subsystem Is there a physical link between them for issuing commands or providing feedback signals? For example, if an abnormal pitch rate in the pitch subsystem is transmitted through the load, it can cause torque fluctuations in the main shaft subsystem. In the topology diagram, this is defined as the pitch subsystem pointing to the main shaft subsystem.

[0099] A causal logical path refers to the path along which an alarm signal, determined by combining the trigger sequence with a pre-defined logical topology, is transmitted from the cause node to the result node. The establishment of a causal logical path follows the principles of topology priority and timing verification. First, the joint encoding matrix is ​​retrieved. The system checks whether the alarm nodes have a connection path within the preset logical topology; if so, it further verifies the trigger sequence. If the preset path direction is met and the trigger alarm time information is available... Information on the time of alarm before the result Then establish by point to Directed edges; if the actual triggering order violates the preset topological logic, i.e. Prior to If an alarm is triggered, a judgment will be made. Not by It is not induced, but rather caused by an independent precipitating factor resulting from a sudden external load. and No directed edges are established between them.

[0100] In this embodiment of the application, the joint coding matrix is ​​first extracted. Time information corresponding to each alarm node According to time information The numerical value determines the alarm node. The triggering sequence between them is determined. Then, the preset logical topology is retrieved.

[0101] For example, taking a fault in unit B in region A as an example, if an alarm node of the pitch subsystem is detected... At any moment Triggered, spindle subsystem alarm node At any moment The trigger, the corresponding time information sequence is as follows And satisfy Meanwhile, the preset logical topology defines the logical connection from the pitch subsystem to the spindle subsystem. By comparing the trigger sequence with the preset logical topology, the causal logical path between alarm nodes can be determined. Ultimately, a path from... point to The causal logic path.

[0102] Step 402: Construct a directed acyclic graph with alarm nodes as nodes and causal logical paths as directed edges.

[0103] In this embodiment of the application, the joint coding matrix is ​​first... alarm nodes in This is mapped to independent vertices in a graph structure. Then, based on the determined causal logical path, directed edges are established between the nodes that act as causes and the nodes that act as effects.

[0104] For example, using alarm nodes Starting from the alarm node Establish a directed edge for the endpoint. Because alarm events have irreversible temporal attributes and there is no cyclic triggering logic between subsystems, the established directed edges do not form cycles. Ultimately, by combining all identified alarm nodes... By topologically integrating the directed edges between them, a directed acyclic graph capable of describing the dynamic evolution of the fault is constructed. .

[0105] Step 403: Accumulate the number of out-degrees of each node in the directed acyclic graph to obtain the global importance of each node, and calculate the mean of the topological distances between each node and the remaining nodes in the directed acyclic graph to obtain the centrality of each node.

[0106] In this step, out-degree refers to the number of directed edges originating from a specific node in a directed acyclic graph (DAG) and pointing to other nodes. Global importance is a metric that measures the degree to which a node influences the ability to trigger subsequent anomalies during the failure evolution process. Topological distance is the minimum number of directed edges traversed between two nodes in a DAG. Centrality is a metric indicating how well a node occupies a core evolutionary position in the DAG.

[0107] In this embodiment of the application, the directed acyclic graph is first traversed. For each alarm node, count the number of directed edges originating from that node, and calculate the global importance of each node by summing the out-degree counts. .

[0108] For example, for nodes If it points to the directed edges respectively as well as Then its global importance The value is 2. Next, the shortest path algorithm is used to calculate the topological distance between the current node and all remaining nodes in the graph. Then, the mean of all calculated topological distances is calculated to obtain the centrality of each node. The calculation formula is shown in formula (2):

[0109] (2)

[0110] In the above formula, Indicates the first Centrality of each alarm node Represents a node To the node Topological distance, This represents the total number of alarm nodes in the graph. For example, the calculated centrality sequence for all nodes can be represented as... .

[0111] Step 404: Following the order of alarm nodes in the alarm node sequence, concatenate the global importance and centrality of all nodes, and then truncate or pad the concatenation result to a fixed length to obtain the structural feature vector.

[0112] In this step, fixed-length truncation or padding refers to the standardization process that forcibly adjusts the data length to ensure consistency in the feature dimensions input to the model. The structural feature vector refers to the final mathematical feature representation used to determine the target fault category.

[0113] In this embodiment of the application, the alarm nodes are first arranged according to the alarm node sequence. The initial order is used to extract the global importance of each node in turn. and centrality Next, the global importance of the corresponding alarm node will be determined. and centrality Arrange the beginning and end to form local features.

[0114] For example, for the first alarm node Concatenating the vectors yields a local vector. Then, the local vectors of all alarm nodes in the sequence are concatenated sequentially using long vector concatenation. For example, the concatenated vectors yield the original vector. Finally, the spliced ​​result is truncated or padded to a fixed length. If the preset fixed length is... When the length of the original vector exceeds Truncation is performed when the original vector length is insufficient. The end is padded with zeros. This ultimately yields the structural feature vector. The generated structural feature vector It can be represented as: Structural feature vectors The dimension and the preset fixed length Consistent.

[0115] Step 105: With the goal of minimizing the logarithmic loss between the structural feature vector and the preset discriminant function set, iteratively optimize the regression coefficient matrix in the discriminant function set to determine the target fault category corresponding to the structural feature vector.

[0116] In this step, the discriminant function set refers to a pre-defined set of mathematical mapping models used to perform the classification task. The logarithmic loss is a cost function used to measure the difference in distribution between the predicted class probabilities of the structural feature vectors and the true fault labels. The regression coefficient matrix refers to the set of adjustable weight parameters within the discriminant function set. The target fault category refers to the specific fault type ultimately identified, which may include physical component over-limit faults, external load impact faults, sensor calibration errors, control software deadlocks, and communication protocol anomalies.

[0117] Step 501: Perform probability mapping between the structural feature vector and the regression coefficient matrix in the discriminant function set to obtain the category probability of each fault category. The fault categories include physical component over-limit faults, external load impact faults, sensor calibration errors, control software deadlocks, and communication protocol anomalies.

[0118] In this step, the category probability refers to the numerical value of the likelihood of occurrence for each fault category.

[0119] Step 5011: Perform a dot product operation between the structural feature vector and the regression coefficient matrix to obtain the initial score for each fault category.

[0120] In this step, the initial score refers to the original mapping result obtained after linear weighted summation of the structural feature vector and the regression coefficient matrix.

[0121] In this embodiment of the application, structural feature vectors are first extracted. Next, With regression coefficient matrix Perform the dot product operation.

[0122] For example, taking a fault in unit B in region A as an example, the structural feature vector With regression coefficient matrix The inner product is calculated using the corresponding weight vectors in the matrix. The formula is: , in the formula Indicates the first The initial score for each fault category, Represents the first eigenvector in the structural feature vector One portion, Represents the correlation in the regression coefficient matrix. The weight parameters of each feature. The final initial score sequence. .

[0123] Step 5012: Perform an exponential operation on each initial score to obtain a mapped score, and calculate the ratio of each mapped score to the sum of all mapped scores to obtain the category probability corresponding to each fault category.

[0124] In this step, the mapped score refers to the intermediate calculated value in the non-negative range obtained after processing the initial score using the natural exponential function.

[0125] In this embodiment of the application, the initial score for each fault category is first obtained. Next, each initial score is raised to its natural power to obtain the corresponding mapped score. For example, the resulting sequence of mapped scores... Then, calculate the ratio of each mapping score to the sum of all mapping scores. The calculation formula is shown in formula (3):

[0126] (3)

[0127] in the formula Indicates the first The category probabilities corresponding to each fault category are obtained. This ultimately yields a category probability sequence. .

[0128] Step 502: Compare the distribution differences between the category probabilities and the preset label values ​​to obtain the log loss. With the goal of reducing the log loss value, iteratively update each weight parameter of the regression coefficient matrix to obtain the optimized regression coefficient matrix.

[0129] In this step, the preset label values ​​refer to the pre-determined standard vector used to label the true types of faults. The logarithmic loss is a cost function used to measure the difference in distribution between the class probability sequence and the preset label value sequence. The optimized regression coefficient matrix refers to the set of parameters after iterative optimization has been completed and the preset termination condition has been met.

[0130] Step 5021: Calculate the distribution difference between the category probability and the preset label value to obtain the log loss.

[0131] In this embodiment of the application, the category probability sequence is first obtained. and preset label value vector For example, if the current actual fault type of unit B in region A is a communication protocol anomaly, then the corresponding preset label value vector... Then, by comparing the distribution patterns of the two, log-likelihood estimation is performed to obtain the log loss. The calculation formula is shown in formula (4):

[0132] (4)

[0133] in the formula Represents logarithmic loss. This represents the components in the preset label value vector. This represents the components in the category probability sequence. The calculated numerical value... This reflects the level of discrimination accuracy under the current regression coefficient matrix.

[0134] Step 5022: Calculate the partial derivative of each weight parameter in the regression coefficient matrix to obtain the rate of change of the logarithmic loss with respect to each weight parameter, and determine the rate of change as the gradient component.

[0135] In this step, the rate of change refers to the first derivative of the logarithmic loss function with respect to specific weight parameters, reflecting the trend of the loss function changing with parameter variations. The gradient component is a vector element composed of multiple rates of change that indicates the direction of the fastest local growth of the loss function.

[0136] In this embodiment of the application, the differential chain rule is first used to analyze the regression coefficient matrix. Each weight parameter in Perform differentiation. Then calculate the logarithmic loss. The rate of change for each weight parameter is calculated using the following formula: , in the formula This represents the gradient components. Finally, all calculated gradient components are aggregated to form the gradient matrix used to guide parameter adjustments. .

[0137] For example, the resulting gradient matrix Represented as: ,in, This represents the total number of weight parameters in the regression coefficient matrix.

[0138] Step 5023: Adjust each weight parameter in the opposite direction of the gradient components, and perform a dot product operation between the structural feature vector and the updated regression coefficient matrix until the regression coefficient matrix converges or reaches the preset number of iterations to obtain the optimized regression coefficient matrix.

[0139] In this step, the opposite direction refers to the negative vector relative to the gradient growth direction, i.e., the search path where the loss function decreases fastest. The preset number of iterations refers to the maximum number of loop steps for optimizing the regression coefficient matrix.

[0140] In this embodiment, each weight parameter is first adjusted stepwise in the opposite direction of the gradient components. The calculation formula is: , in the formula This represents the adjusted weight parameters. This represents the current weight parameters. This indicates the preset learning rate, which is a proportional coefficient used to control the step size of a single parameter update. This represents the gradient components. Then, the structural feature vectors are re-evaluated. The updated regression coefficient matrix is ​​then multiplied by a dot product, and the probability mapping and loss calculation processes are repeated. This iterative process continues, monitoring the magnitude of changes in the logarithmic loss. The computation stops when the logarithmic loss reaches a preset stable state (meaning the change in logarithmic loss decreases to below a preset threshold), or when the preset number of iterations is reached. The optimized regression coefficient matrix is ​​then obtained. .

[0141] For example, the optimized regression coefficient matrix It is represented in the following format:

[0142]

[0143] in, This represents the optimized regression coefficient matrix. This represents a fixed length for the structural feature vector. This process enables continuous optimization of the fault classification model's performance.

[0144] Step 503: Perform a dot product operation on the structural feature vector and the optimized regression coefficient matrix to obtain the target prediction sequence, and determine the target fault category based on the classification label corresponding to the largest component in the target prediction sequence.

[0145] In this step, the target prediction sequence refers to the set of classification scores obtained by mapping the structural feature vectors using the optimized regression coefficient matrix. The classification identifier is an index symbol used to uniquely determine the fault type. The target fault category refers to the final fault type determined after logical discrimination.

[0146] In this embodiment of the application, the structural feature vector to be determined is first obtained. Next, the structural feature vectors With the optimized regression coefficient matrix Perform a dot product operation. This step maps the topological features to the fault determination space.

[0147] For example, the target prediction sequence obtained by the operation ,in, arrive This represents the predicted score for each fault category. Finally, in the target prediction sequence... The system retrieves the component with the largest value. Specifically, it extracts the classification identifier corresponding to the largest component by comparing the component sizes. For example, if... If the value is the maximum, the corresponding category identifier is retrieved as a physical component over-limit fault.

[0148] Graph structure features can effectively distinguish different types of faults: physical component over-limit faults typically exhibit long chain-like causal transmission characteristics in directed acyclic graphs, with high centrality of the inducing nodes; communication protocol anomalies typically manifest as a large number of alarm nodes occurring concurrently from different subsystems within a very short time, exhibiting an extremely uneven distribution of node degree and a flattened, explosive topology in the directed acyclic graph; external load impact faults manifest as deviations in operating parameters. At the moment of multiple alarm triggering, there were severe, physically correlated synchronization peaks. These qualitative physical logics were transformed into structural feature vectors. The quantitative values ​​improved the classification accuracy in complex multi-source alarm coupling scenarios, ultimately determining the target fault category for this fault.

[0149] This application's embodiments ensure that abnormal unit performance can be captured simultaneously from both logical evolution and physical state deviation dimensions by synchronously acquiring discrete state switching signals at the logical level and continuous operating parameters at the physical level. This solves the problem that single numerical trend analysis cannot quantify the severity of anomalies, transforming raw operating data into a standard matrix that reflects the degree to which the system deviates from safety boundaries, further improving the intelligent operation and maintenance level of the industrial cloud platform under complex multi-source data in the wind power field. It overcomes the limitation of existing technologies in capturing the causal evolution relationship between alarm signals. It also improves the classification accuracy and diagnostic reliability in highly coupled alarm scenarios.

[0150] Figure 4 This is a schematic diagram of a specific implementation of the artificial intelligence-based wind turbine fault classification and mining system provided in this application embodiment, with reference to... Figure 4 The system may include:

[0151] The acquisition module 21 is used to acquire the timing alarm sequence of the wind turbine during the safety chain triggering process and the operating condition dataset within a preset time before the safety chain triggering process. The timing alarm sequence includes the logical state combination of different subsystems in the wind turbine at multiple time points.

[0152] The determination module 22 is used to determine the alarm node sequence and the time information corresponding to each alarm node based on the logical state of different locations in the time-series alarm sequence, perform boundary processing on the limit parameters of the wind turbine to obtain the limit constraint set, and calculate the Euclidean distance between each parameter in the working condition dataset and the corresponding limit threshold in the limit constraint set to obtain the state deviation matrix.

[0153] The association module 23 is used to spatiotemporally associate the alarm node and the corresponding time information with the state deviation matrix and then concatenate them to obtain a joint encoding matrix;

[0154] Module 24 is used to construct a directed acyclic graph using alarm nodes in the joint coding matrix as nodes and causal logical paths between alarm nodes as directed edges, and to construct structural feature vectors by calculating the global importance and centrality of each node in the directed acyclic graph.

[0155] The iteration module 25 is used to iteratively optimize the regression coefficient matrix in the discriminant function set with the goal of minimizing the log loss between the structural feature vector and the preset discriminant function set in order to determine the target fault category corresponding to the structural feature vector.

[0156] The AI-based wind turbine fault classification and mining system of this application embodiment is used to implement the aforementioned AI-based wind turbine fault classification and mining method. Therefore, the specific implementation of the AI-based wind turbine fault classification and mining system can be found in the embodiment section of the AI-based wind turbine fault classification and mining method above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0157] Figure 5 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0158] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described artificial intelligence-based wind turbine fault classification and mining method.

[0159] The electronic device may include a processor 510 and a memory 520 storing computer program instructions.

[0160] Specifically, the processor 510 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0161] Memory 520 may include mass storage for data or instructions. For example, and not limitingly, memory 520 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 520 may include removable or non-removable (or fixed) media. Where appropriate, memory 520 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 520 is non-volatile solid-state memory.

[0162] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to the first aspect of this disclosure.

[0163] The processor 510 reads and executes computer program instructions stored in the memory 520 to implement any of the artificial intelligence-based wind turbine fault classification and mining methods in the above embodiments.

[0164] In one example, the electronic device may also include a communication interface 530 and a bus 540. Wherein, such as Figure 5 As shown, the processor 510, memory 520, and communication interface 530 are connected through bus 540 and complete communication with each other.

[0165] The communication interface 530 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0166] Bus 540 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 540 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0167] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described artificial intelligence-based wind turbine fault classification and mining methods.

[0168] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0169] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the artificial intelligence-based wind turbine fault classification and mining method described above.

[0170] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0171] The above provides a detailed description of the artificial intelligence-based wind turbine fault classification and mining method and system provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for classifying and mining faults in wind turbine units based on artificial intelligence, characterized in that, include: The timing alarm sequence of the wind turbine during the safety chain triggering process and the operating condition dataset within a preset time period before the safety chain triggering process are obtained. The timing alarm sequence includes the logical state combination of different subsystems in the wind turbine at multiple time points. Based on the logical states of different locations in the time-series alarm sequence, the alarm node sequence and the time information corresponding to each alarm node are determined. The limit parameters of the wind turbine are subjected to boundary processing to obtain a set of limit constraints. The Euclidean distance between each parameter in the operating condition dataset and the corresponding limit threshold in the set of limit constraints is calculated to obtain the state deviation matrix. The alarm node and its corresponding time information are spatiotemporally correlated with the state deviation matrix to obtain a joint coding matrix. Using the alarm nodes in the joint coding matrix as nodes and the causal logical paths between the alarm nodes as directed edges, a directed acyclic graph is constructed, and a structural feature vector is constructed by calculating the global importance and centrality of each node in the directed acyclic graph. With the goal of minimizing the logarithmic loss between the structural feature vector and the preset discriminant function set, the regression coefficient matrix in the discriminant function set is iteratively optimized to determine the target fault category corresponding to the structural feature vector.

2. The method according to claim 1, characterized in that, Based on the logical states of different locations in the time-series alarm sequence, the alarm node sequence and the time information corresponding to each alarm node are determined. Boundary processing is then performed on the limit parameters of the wind turbine to obtain a set of limit constraints. The Euclidean distance between each parameter in the operating condition dataset and the corresponding limit threshold in the set of limit constraints is calculated to obtain a state deviation matrix, including: The location point where the logical state changes in the time-series alarm sequence is taken as the alarm node. According to the order of the alarm nodes in the time-series alarm sequence, all alarm nodes are arranged into an alarm node sequence, and the time when the logical state changes in each alarm node in the alarm node sequence is taken as the corresponding time information. The operating condition range of the wind turbine is divided into intervals to obtain multiple operating condition intervals. In each operating condition interval, a limit threshold corresponding to each limit parameter of the wind turbine is set. All limit thresholds are combined to obtain a set of limit constraints. The limit constraint set and the working condition dataset are mapped to a preset multidimensional feature space to calculate the Euclidean distance between each parameter in the working condition dataset and the corresponding limit threshold. According to the sampling time of each parameter in the working condition dataset, all Euclidean distances are arranged in a matrix to obtain the state deviation matrix.

3. The method according to claim 1, characterized in that, The alarm node and its corresponding time information are spatiotemporally correlated with the state deviation matrix to obtain a joint coding matrix, including: The state deviation matrix is ​​retrieved to determine the Euclidean distance between the time information corresponding to each alarm node and the time information at the same moment, so as to obtain the instantaneous offset corresponding to each alarm node. Each alarm node, its corresponding time information, and instantaneous offset are arranged horizontally to obtain a comprehensive attribute vector for each alarm node. According to the chronological order of the time information corresponding to each alarm node, all comprehensive attribute vectors are arranged vertically to obtain the joint coding matrix.

4. The method according to claim 3, characterized in that, Using the alarm nodes in the joint encoding matrix as nodes and the causal logical paths between the alarm nodes as directed edges, a directed acyclic graph is constructed. A structural feature vector is then constructed by calculating the global importance and centrality of each node in the directed acyclic graph, including: Based on the trigger sequence of the comprehensive attribute vectors in the joint coding matrix and the preset logical topology relationship between each subsystem in the wind turbine, the causal logical path between the alarm nodes is determined; Using the alarm node as the node and the causal logic path as the directed edge, construct a directed acyclic graph; The global importance of each node is obtained by summing the number of out-degrees of each node in the directed acyclic graph, and the centrality of each node is obtained by averaging the topological distances between each node and the remaining nodes in the directed acyclic graph. Following the order of the alarm nodes in the alarm node sequence, the global importance and centrality of all nodes are concatenated, and the concatenation result is truncated or padded to a fixed length to obtain the structural feature vector.

5. The method according to claim 1, characterized in that, The discriminant function set includes sub-discriminant functions corresponding to physical component over-limit faults, external load impact faults, sensor calibration errors, control software deadlocks, and communication protocol anomalies. With the objective of minimizing the logarithmic loss between the structural feature vector and the preset discriminant function set, the regression coefficient matrix in the discriminant function set is iteratively optimized to determine the target fault category corresponding to the structural feature vector, including: The structural feature vector is mapped to the regression coefficient matrix in the discriminant function set to obtain the category probability of each fault category. The fault categories include physical component over-limit faults, external load impact faults, sensor calibration errors, control software deadlocks, and communication protocol anomalies. The distribution difference between the category probability and the preset label value is compared to obtain the log loss. With the goal of reducing the log loss value, each weight parameter of the regression coefficient matrix is ​​iteratively updated to obtain the optimized regression coefficient matrix. The structural feature vector and the optimized regression coefficient matrix are multiplied by a dot product to obtain the target prediction sequence. The target fault category is then determined based on the classification identifier corresponding to the largest component in the target prediction sequence.

6. The method according to claim 5, characterized in that, By performing a probability mapping between the structural feature vector and the regression coefficient matrix in the discriminant function set, the category probability of each fault category is obtained, including: The structural feature vector is multiplied by the regression coefficient matrix to obtain the initial score for each fault category. The initial score is raised to an exponent to obtain the mapped score. The ratio of each mapped score to the sum of all mapped scores is then calculated to obtain the category probability corresponding to each fault category.

7. The method according to claim 5, characterized in that, The distribution difference between the category probabilities and preset label values ​​is compared to obtain the log loss. To reduce the log loss value, each weight parameter of the regression coefficient matrix is ​​iteratively updated to obtain an optimized regression coefficient matrix, including: The log loss is obtained by calculating the distribution difference between the category probability and the preset label value; Calculate the partial derivative of each weight parameter in the regression coefficient matrix to obtain the rate of change of the logarithmic loss with respect to each weight parameter, and determine the rate of change as the gradient component; Adjust each weight parameter in the opposite direction of the gradient components, and perform a dot product operation between the structural feature vector and the updated regression coefficient matrix until the regression coefficient matrix converges or reaches the preset number of iterations to obtain the optimized regression coefficient matrix.

8. A wind turbine fault classification and discovery system based on artificial intelligence, characterized in that, include: The acquisition module is used to acquire the timing alarm sequence of the wind turbine during the safety chain triggering process and the operating condition dataset within a preset time period before the safety chain triggering process. The timing alarm sequence includes the logical state combination of different subsystems in the wind turbine at multiple time points. The determination module is used to determine the alarm node sequence and the time information corresponding to each alarm node based on the logical state of different location points in the time-series alarm sequence, perform boundary processing on the limit parameters of the wind turbine to obtain a set of limit constraints, and calculate the Euclidean distance between each parameter in the operating condition dataset and the corresponding limit threshold in the set of limit constraints to obtain a state deviation matrix. The association module is used to spatiotemporally associate the alarm node and its corresponding time information with the state deviation matrix to obtain a joint coding matrix. The construction module is used to construct a directed acyclic graph using alarm nodes in the joint coding matrix as nodes and causal logical paths between alarm nodes as directed edges, and to construct a structural feature vector by calculating the global importance and centrality of each node in the directed acyclic graph. The iterative module is used to iteratively optimize the regression coefficient matrix in the discriminant function set with the goal of minimizing the log loss between the structural feature vector and the preset discriminant function set, so as to determine the target fault category corresponding to the structural feature vector.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the artificial intelligence-based wind turbine fault classification and mining method as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the artificial intelligence-based wind turbine fault classification and mining method as described in any one of claims 1 to 7.

Images