Fault diagnosis method and system for wind driven generator system

Abstract

The invention relates to the technical field of fan detection, and particularly provides a fault diagnosis method and system for a wind driven generator system, and the method comprises the steps: firstly obtaining synchronous sampling records of the vibration of a gear box, the temperature of a generator and a blade rotating speed sensor in a continuous time period as system collection data; secondly, extracting an independent feature group reflecting the abnormal tendency of a single component and an associated feature group reflecting abnormal linkage between the components through a wind driven generator feature learning network, and then calling a pre-trained component fault diagnosis model to perform joint judgment on the two groups of features; the method comprises the steps of generating a diagnosis intermediate result containing component abnormal confidence and abnormal linkage strength, performing dynamic calibration, identifying a main fault component, an accompanying fault component and a fault propagation path based on the diagnosis intermediate result, finally generating a fault early warning instruction containing component positioning information, and sending the fault early warning instruction to an operation and maintenance system to trigger targeted maintenance. Therefore, faults can be accurately diagnosed, maintenance efficiency is improved, and stable operation of the system is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of wind turbine testing technology, and more specifically, to a fault diagnosis method and system for wind turbine generator systems. Background Technology

[0002] In the field of wind power generation, the stable operation of wind turbine systems is crucial for ensuring the reliability and economy of power supply. However, due to the fact that wind turbine systems operate in complex natural environments for extended periods, and have numerous internal components and complex operating conditions, failures are difficult to avoid.

[0003] Currently, traditional methods for diagnosing wind turbine systems primarily rely on data analysis from single sensors and human experience. For critical components such as gearboxes, generators, and blades, traditional methods often focus only on measurement data of individual components, such as gearbox vibration data, generator temperature data, or blade speed data, neglecting the inherent connections and interactions between different components. For example, when a gearbox malfunctions, its abnormal vibration may gradually transmit to the generator, causing the generator temperature to rise, but traditional methods struggle to detect such cross-component abnormal linkages.

[0004] Furthermore, human experience-based judgment is highly subjective and lacks a systematic approach. Different maintenance personnel may have inconsistent standards for judging faults, which can easily lead to misjudgments or omissions. Moreover, fault warning information generated by traditional methods is usually quite simple, only indicating components that may be faulty, and failing to accurately identify the primary faulty component, accompanying faulty components, and the propagation path of the fault between components. This makes subsequent targeted maintenance operations difficult, increases maintenance costs, and reduces downtime. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, the present invention provides a fault diagnosis method for a wind turbine generator system, the method comprising:

[0006] Acquire system data of the wind turbine system. The system data includes synchronous sampling records of gearbox vibration sensor, generator temperature sensor and blade speed sensor in a continuous period of time. Each sampling record contains the original measurement value of the corresponding component at a single moment.

[0007] The system collects data by using a wind turbine feature learning network to extract fault-sensitive features, generating independent feature groups that reflect the abnormal tendencies of a single component and associated feature groups that reflect the abnormal linkages between components. The independent feature groups include the time-series fluctuation patterns of component measurements, and the associated feature groups include the synchronous abnormal patterns of cross-component measurements.

[0008] The pre-trained component fault diagnosis model is invoked to jointly discriminate the independent feature group and the associated feature group, generating an intermediate diagnostic result containing the component anomaly confidence and the anomaly linkage strength. The intermediate diagnostic result is dynamically calibrated through the component association knowledge base associated with the component fault diagnosis model.

[0009] Based on the intermediate diagnostic results, the main faulty components and accompanying faulty components in the wind turbine system are identified, and the propagation path of the fault between components and the causal relationship of the abnormality of each component are determined.

[0010] Based on the main fault component, accompanying fault components and fault propagation path, a fault warning instruction containing component location information is generated, and the fault warning instruction is sent to the wind turbine operation and maintenance system to trigger targeted maintenance operations. The component location information is associated with the physical installation location of the component inside the wind turbine tower.

[0011] Furthermore, the present invention also provides a fault diagnosis system for wind turbine generator systems, the system comprising:

[0012] The data acquisition module is used to acquire system acquisition data of the wind turbine system. The system acquisition data includes synchronous sampling records of gearbox vibration sensor, generator temperature sensor and blade speed sensor in a continuous period of time. Each sampling record contains the original measurement value of the corresponding component at a single moment.

[0013] The data preprocessing module is used to extract fault-sensitive features from the data collected by the system through the wind turbine feature learning network, and generate independent feature groups reflecting the abnormal tendency of a single component and related feature groups reflecting the abnormal linkage between components. The independent feature groups include the time-series fluctuation pattern of component measurement values, and the related feature groups include the synchronous abnormal pattern of cross-component measurement values.

[0014] The diagnostic intermediate result generation module is used to call the pre-trained component fault diagnosis model to jointly discriminate the independent feature group and the associated feature group, and generate diagnostic intermediate results including component anomaly confidence and anomaly linkage strength. The diagnostic intermediate results are dynamically calibrated through the component association knowledge base associated with the component fault diagnosis model.

[0015] The fault diagnosis module identifies the main faulty components and accompanying faulty components in the wind turbine system based on the intermediate diagnostic results, and determines the propagation path of the fault between components and the causal relationship of the abnormality of each component.

[0016] The fault early warning module generates a fault early warning command containing component location information based on the main fault component, the accompanying fault components and the fault propagation path, and sends the fault early warning command to the wind turbine operation and maintenance system to trigger targeted maintenance operations. The component location information is associated with the physical installation location of the component inside the wind turbine tower.

[0017] Based on the above, this invention acquires system data collected synchronously from multiple sensors over a continuous period. It then utilizes a wind turbine feature learning network to extract fault-sensitive features from the system data, generating independent feature groups reflecting the abnormal tendencies of individual components and associated feature groups reflecting the abnormal linkages between components. This allows for in-depth analysis of the abnormal characteristics of the components themselves and the relationships between them, providing a more comprehensive understanding of the wind turbine system's operating status. A pre-trained component fault diagnosis model is then used to jointly discriminate the independent and associated feature groups, generating intermediate diagnostic results containing component anomaly confidence and anomaly linkage strength. Dynamic calibration is performed using a component association knowledge base, improving the accuracy and reliability of fault diagnosis. Based on the intermediate diagnostic results, the main faulty component and accompanying faulty components are identified, determining the fault propagation path between components and the causal relationships of each component's anomalies. Finally, based on this information, a fault warning command containing component location information is generated and sent to the operation and maintenance system. This enables rapid and accurate triggering of targeted maintenance operations, improving maintenance efficiency, reducing maintenance costs and downtime, and ensuring the stable operation of the wind turbine system. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the execution flow of the fault diagnosis method for wind turbine systems provided in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of exemplary hardware and software components of a fault diagnosis device for a wind turbine system provided in an embodiment of the present invention.

[0020] Figure 3 This is an exemplary schematic diagram of a fault diagnosis system for wind turbine generator systems provided in an embodiment of the present invention.

[0021] Among them, 100 is the fault diagnosis device, 110 is the network port, 120 is the processor, 130 is the communication bus, 140 is the storage medium, and 150 is the I / O interface. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a fault diagnosis method for a wind turbine system according to an embodiment of the present invention. The following is a detailed description of the fault diagnosis method for a wind turbine system.

[0023] Step S110: Acquire system data of the wind turbine system. The system data includes synchronous sampling records of the gearbox vibration sensor, generator temperature sensor, and blade speed sensor over a continuous period of time. Each sampling record contains the original measurement value of the corresponding component at a single moment.

[0024] In this embodiment, the wind turbine system is equipped with several key sensors, including a gearbox vibration sensor, a generator temperature sensor, and a blade speed sensor. These sensors continuously and synchronously sample data over a continuous time period. The gearbox vibration sensor primarily monitors the vibration of the gearbox during operation. When the wind turbine is running, the gears inside the gearbox mesh and rotate, generating vibrations. Under normal circumstances, this vibration has a relatively stable pattern and frequency. However, when the gearbox malfunctions, such as gear wear or tooth surface peeling, the vibration pattern and frequency change. Therefore, the gearbox vibration sensor samples the vibration at regular time intervals, and the record obtained from each sample is the original vibration measurement value of the gearbox at that single moment.

[0025] Generator temperature sensors are specifically designed to monitor changes in generator temperature. Generators generate heat during the conversion of mechanical energy into electrical energy, and their temperature fluctuates with changes in load. However, under normal operating conditions, the generator temperature remains within a reasonable range. If a generator malfunctions, such as a winding short circuit or a cooling system failure, the temperature will rise abnormally. Therefore, generator temperature sensors collect the generator's temperature value in real time at every moment; these raw measurements reflect the generator's thermal state.

[0026] The function of a blade speed sensor is to measure the rotational speed of the blades. The blade speed is closely related to factors such as wind speed and generator output power. Under different wind conditions, the blades rotate at corresponding speeds to achieve optimal energy conversion efficiency. When blades malfunction, such as due to deformation or imbalance, the speed will fluctuate abnormally. Therefore, the blade speed sensor continuously records the blade's rotational speed at various times.

[0027] By synchronously sampling the aforementioned sensors over a continuous period of time, the system data collected by the wind turbine system can be obtained, which can reflect the operating status of each component of the wind turbine at different times.

[0028] Step S120: Extract fault-sensitive features from the system data collected by the wind turbine feature learning network to generate independent feature groups reflecting the abnormal tendency of a single component and related feature groups reflecting the abnormal linkage between components. The independent feature groups include the time-series fluctuation pattern of component measurement values, and the related feature groups include the synchronous abnormal pattern of cross-component measurement values.

[0029] Step S121: Divide the system-collected data into synchronous periodic segments according to the power generation cycle of the wind turbine. Each periodic segment contains sampling records of the gearbox, generator, and blades within a complete power generation cycle.

[0030] In this embodiment, in order to more effectively analyze and process the data collected by the system, it is necessary to divide it according to the power generation cycle of the wind turbine. The power generation process of the wind turbine is periodic, usually starting with the blades capturing wind energy, which is then increased in speed and transmitted through the gearbox, and finally converted into electrical energy by the generator, and finally output to the power grid, completing a complete power generation cycle.

[0031] In practice, the first step is to determine the power generation cycle of the wind turbine. This can be achieved by monitoring and analyzing the generator's output power. The generator's output power fluctuates periodically with the changes in the power generation cycle; by analyzing these fluctuations, the start and end points of the power generation cycle can be accurately determined.

[0032] Then, based on the determined power generation cycle, the system's collected data is divided into synchronous periodic segments. Each periodic segment contains sampling records from the gearbox, generator, and blades within a complete power generation cycle. For example, within a power generation cycle, the gearbox vibration sensor records the gearbox's vibration at different times, the generator temperature sensor records the generator's temperature changes, and the blade speed sensor records the blade speed fluctuations. These data collectively constitute a periodic segment.

[0033] By dividing the system's collected data into periodic segments, the operating status of each component within each power generation cycle can be observed more clearly, facilitating subsequent extraction and analysis of component characteristics. Furthermore, this division method also helps to identify patterns of component variation across different power generation cycles.

[0034] Step S122: Through the wind turbine feature learning network, time-series pattern learning processing is performed on the sampling records of a single component in each cycle segment to generate an independent feature group that reflects the operational stability of the component itself. The independent feature group includes the deviation between the fluctuation pattern and the historical normal fluctuation pattern. The time-series pattern learning processing is implemented through a recurrent neural network to extract the fluctuation pattern of the component measurement value in the rising segment, the stable segment, and the falling segment within the cycle.

[0035] Step S1221: Input the sampling record of the single component into the input layer of the recurrent neural network of the wind turbine feature learning network. The input layer receives the time series of gearbox vibration value, generator temperature value or blade speed value.

[0036] In this embodiment, when performing time-series pattern learning processing on the sampling records of a single component, these records are first input into the input layer of the recurrent neural network of the wind turbine feature learning network. The input layer is the entry point of the recurrent neural network, which is responsible for receiving the time series of gearbox vibration values, generator temperature values, or blade speed values.

[0037] The time series of gearbox vibration values ​​consists of a series of vibration measurements collected by a gearbox vibration sensor within a periodic segment. These vibration measurements are arranged in chronological order, forming a continuous sequence. The input layer receives each value in this continuous sequence sequentially as initial data for processing by the recurrent neural network.

[0038] Similarly, the time series of generator temperature and blade speed values ​​are input to the input layer in a similar manner. The time series of generator temperature values ​​consists of temperature measurements collected by the generator temperature sensor within one cycle segment, while the time series of blade speed values ​​consists of speed measurements collected by the blade speed sensor within the same cycle segment.

[0039] The role of the input layer is to perform preliminary processing and formatting of these time series data so that the subsequent recurrent neural network can process them better. It ensures that the sampling records of each component can enter the recurrent neural network in the correct order and format.

[0040] Step S1222: Process the sampled values ​​step by step through the hidden layer of the recurrent neural network, and record the hidden state of each time step. The hidden state includes the relationship between the current value and the value of the previous time step.

[0041] After the sampled records of a single component are input into the input layer of a recurrent neural network, the hidden layer processes these sampled values ​​step by step, and the hidden layer can capture the temporal dependencies in the data.

[0042] At each time step, the hidden layer can calculate the hidden state for the current time step based on the sampled value of the current input and the hidden state of the previous time step. This calculation process is implemented through a series of neurons and weight parameters. Specifically, the neurons in the hidden layer perform a weighted summation of the sampled value of the current input and the hidden state of the previous time step, and then perform a nonlinear transformation through an activation function to obtain the hidden state for the current time step.

[0043] The hidden state contains the relationship between the current value and the value at the previous time step. For example, for vibration measurements of a gearbox, if the vibration value at the current time step is greater than the vibration value at the previous time step, the hidden state will reflect this increase. By recording the hidden state at each time step, the recurrent neural network can track the changing trends of component measurements and capture temporal patterns in the data.

[0044] As time progresses, the hidden layers continuously update their hidden states, thereby processing the entire time series. This step-by-step processing method enables recurrent neural networks to effectively process sequence data with time dependencies.

[0045] Step S1223: After the complete periodic segment processing is completed, the key feature sequence is extracted through the output layer of the recurrent neural network. The key feature sequence includes: the slope feature of the rising segment, the variance feature of the stationary segment, and the decay feature of the falling segment.

[0046] After the recurrent neural network completes the processing of a single component sampling record within a complete periodic segment, the output layer extracts the key feature sequence. The output layer is the last layer of the recurrent neural network, responsible for organizing and outputting the information learned by the hidden layers.

[0047] The key feature sequence includes the slope characteristics of the rising segment, the variance characteristics of the stationary segment, and the decay characteristics of the falling segment. These features can accurately describe the fluctuation patterns of component measurements at different stages.

[0048] The slope characteristic of the rising segment reflects the rate of increase of the component's measured value during the rising phase. For example, for vibration measurements of a gearbox, the slope characteristic of the rising segment can represent the rate of change of vibration amplitude during the rising phase. A large slope indicates that the vibration amplitude is increasing rapidly, which may indicate an anomaly.

[0049] The variance characteristics of the steady-state segment are used to measure the degree of fluctuation in the component's measurements during the steady-state phase. A smaller variance indicates that the fluctuation range is relatively stable, suggesting that the component is operating normally during this phase; while a larger variance indicates that the fluctuation range is larger, which may indicate potential anomalies.

[0050] The decay characteristic of the fading phase describes the rate at which a component's measured value decays during the decreasing phase. For example, for a generator's temperature measurement, the decay characteristic of the fading phase can represent the rate at which the temperature decreases during this phase. If the decay rate does not conform to a normal pattern, it may indicate a problem with the component.

[0051] Step S1224: Calculate the Euclidean distance between the slope feature, variance feature, attenuation feature and the corresponding feature of the historical normal cycle, and generate an independent feature group that reflects the deviation of the component's own fluctuation law.

[0052] After extracting the key feature sequences, the next step is to calculate the Euclidean distance between these features and the corresponding features of the historical normal cycle in order to generate independent feature sets that reflect the deviation of the component's own fluctuation pattern.

[0053] Euclidean distance is a commonly used distance metric to measure the similarity between two vectors. In this embodiment, the slope feature, variance feature, and decay feature are combined into a vector, and similarly, the corresponding features of the historical normal cycle are also combined into a vector. By calculating the Euclidean distance between these two vectors, the degree of deviation between the actual fluctuation pattern and the historical normal fluctuation pattern can be obtained.

[0054] For example, for vibration measurements of a gearbox, the actual slope characteristics of the rising segment, the variance characteristics of the steady segment, and the attenuation characteristics of the falling segment form a vector A, while the corresponding characteristics of the historical normal cycle form a vector B. Calculating the Euclidean distance between vector A and vector B, if the distance is large, indicates a significant deviation between the actual fluctuation pattern and the historical normal fluctuation pattern, suggesting that the component may exhibit abnormal tendencies.

[0055] By performing the above calculations on the slope, variance, and attenuation characteristics of each component, an independent feature set reflecting the deviation of the component's own fluctuation pattern can be generated. This independent feature set can accurately reflect the operational stability of the component itself.

[0056] Step S123: Through the wind turbine feature learning network, the cross-component sampling records in each cycle segment are processed for linkage mode learning to generate an associated feature group reflecting abnormal linkage between components. The associated feature group includes the deviation between the synchronous change pattern and the historical normal synchronous pattern. The linkage mode learning is implemented through a graph neural network. The component linkage graph is constructed with components as nodes and mechanical connection relationships between components as edges to extract the synchronous change patterns of gearbox vibration and generator temperature, and blade speed and gearbox load.

[0057] Step S1231: According to the component linkage diagram defined by the mechanical structure definition of the wind turbine generator, the gearbox node is connected to the generator node, the gearbox node is connected to the blade node, and the generator node is connected to the converter node.

[0058] In this embodiment, when performing linkage mode learning processing, the component linkage diagram must first be defined based on the mechanical structure of the wind turbine. There are clear mechanical connections between the various components of the wind turbine, and these connections form the basis of the component linkage diagram.

[0059] The gearbox node connects to the generator node because the gearbox's main function is to convert the low-speed rotation of the blades into the high-speed rotation required by the generator, achieving the connection between the two through mechanical components such as the drive shaft. The gearbox node connects to the blade node; the blades rotate under the influence of wind, converting wind energy into mechanical energy, which is then transferred to the gearbox through a mechanical structure. The generator node connects to the converter node; the converter's function is to convert and regulate the AC power output from the generator to meet the requirements of the power grid.

[0060] By considering the aforementioned connections, components such as the gearbox, generator, blades, and converter as nodes, and the connections between them as edges, a component linkage diagram is constructed. This component linkage diagram accurately reflects the mechanical connection relationships between components in a wind turbine generator system.

[0061] Step S1232: Initialize a feature vector for each component node using the graph neural network of the wind turbine feature learning network. The feature vector consists of the slope feature, variance feature, and attenuation feature of the corresponding component within the period segment.

[0062] After constructing the component linkage diagram, the next step is to initialize feature vectors for each component node using a graph neural network of the wind turbine feature learning network. A feature vector is a vector used in a graph neural network to represent the features of a node.

[0063] For each component node, its feature vector consists of the slope, variance, and decay features of the corresponding component within the periodic segment. These features are obtained through temporal pattern learning processing on the sampling records of a single component.

[0064] For example, for a gearbox node, its eigenvector contains the slope characteristics of the rising segment, the variance characteristics of the stationary segment, and the attenuation characteristics of the falling segment of the gearbox vibration measurement value within a periodic segment. Similarly, the eigenvector of a generator node contains the corresponding characteristics of the generator temperature measurement value, and the eigenvector of a blade node contains the corresponding characteristics of the blade speed measurement value.

[0065] Through the above initialization, each component node has a feature vector that reflects its operating characteristics within a periodic segment.

[0066] Step S1233: In each layer of the graph neural network, perform a neighborhood information aggregation operation on each node: collect the current layer feature vectors of all neighboring nodes, perform feature transformation through a fully connected layer, and then add them to the feature vector of the current node to obtain the updated node feature vector.

[0067] In the processing of graph neural networks, each layer performs neighborhood information aggregation for each node. Specifically, when at a certain layer of the graph neural network, for each node in the graph, it is necessary to collect the feature vectors of all its neighboring nodes in the current layer. Taking the gearbox node as an example, its neighboring nodes may be the generator node and the blade node, so the feature vectors of these two neighboring nodes in the current layer will be collected.

[0068] After collecting the feature vectors of neighboring nodes, these feature vectors need to be transformed using fully connected layers. A fully connected layer contains a series of neurons and weight parameters, which can perform weighted summation and nonlinear transformations on the input feature vectors. This process can be understood as recombining and abstracting the features of neighboring nodes to extract more valuable information. Through the processing of fully connected layers, the feature vectors of neighboring nodes are mapped to a new feature space, making them more suitable for subsequent computation and analysis.

[0069] After feature transformation, the transformed feature vectors of neighboring nodes are added to the current node's own feature vector. The purpose of this addition is to fuse the information of the current node with that of its neighboring nodes, thus obtaining an updated node feature vector. For example, for a gearbox node, the feature vectors of the generator and blade nodes transformed by the fully connected layer are added to the gearbox node's own feature vector. This updated gearbox node feature vector includes information from its neighboring nodes, providing a more comprehensive reflection of the relationship between the gearbox and other components.

[0070] Since feature vectors are multi-dimensional, and each dimension represents different feature information, it is necessary to ensure that the dimensions of each feature vector are consistent and the units of measurement are uniform when performing feature transformation and addition operations. For example, if a dimension in the feature vector of an adjacent node represents the amplitude of vibration, then the corresponding dimension in the feature vector of the current node should also be the amplitude of vibration.

[0071] As the number of layers in the graph neural network increases, the neighborhood information aggregation operation is repeatedly performed. Each operation adds more features from its multi-hop neighboring nodes to the node's feature vector. Multi-hop neighboring nodes refer to nodes that include not only directly adjacent nodes but also nodes further away, such as the neighbors of those neighbors. For example, after multiple layers of neighborhood information aggregation, the feature vector of a gearbox node may contain not only information about the generator and blade nodes but also information about other component nodes adjacent to the generator or blade nodes, thus enabling a more comprehensive capture of the complex interrelationships between components.

[0072] Step S1234: Repeat the neighborhood information aggregation operation until a preset number of layers is reached, so that the feature vector of each node contains the linkage features of its multi-hop neighborhood nodes.

[0073] In graph neural networks, repeatedly performing neighborhood information aggregation is a process of progressively deepening the understanding of relationships between components. The preset number of layers determines the depth and breadth of relationships between components that the graph neural network can capture.

[0074] As the neighborhood information aggregation operation is repeated, the feature vector of each node gradually becomes richer, from initially containing only its own feature information, to gradually incorporating information from directly adjacent nodes, and then further incorporating information from multi-hop neighboring nodes. For example, in the first layer of neighborhood information aggregation, the feature vector of the gearbox node mainly incorporates information from the generator node and the blade node; while in the second layer operation, since the generator node and the blade node have already updated their feature vectors in the first layer, including some information from their respective neighboring nodes, the feature vector of the gearbox node in this layer will further incorporate information from these indirectly adjacent nodes.

[0075] Step S1235: Extract the linkage features between the gearbox node and the generator node, and the linkage features between the gearbox node and the blade node. Calculate the cosine similarity between the linkage features and the historical normal linkage features, and generate a set of associated features that reflects the deviation of the synchronous change law between components.

[0076] After the graph neural network completes the neighborhood information aggregation operation of the preset number of layers, the feature vector of each node already contains the linkage features of its multi-hop neighboring nodes. Next, it is necessary to extract the linkage features between specific node pairs, namely the linkage features between the gearbox node and the generator node, and the linkage features between the gearbox node and the blade node.

[0077] These linkage features are obtained by processing the component linkage diagram through a graph neural network, and they reflect the synchronous change relationships between different components. For example, the linkage features between the gearbox node and the generator node may reflect how the vibration changes of the gearbox affect the temperature changes of the generator, or how the load changes of the generator are fed back to the operating status of the gearbox.

[0078] After extracting these linkage features, it is necessary to calculate their cosine similarity with historical normal linkage features. Cosine similarity is a commonly used method to measure the similarity between two vectors, which determines their similarity by calculating the cosine value of the angle between the two vectors. Both linkage feature vectors and historical normal linkage feature vectors are multi-dimensional vectors, with each dimension representing different linkage information.

[0079] By calculating the cosine similarity between the linkage characteristics of the gearbox node and the generator node, and the gearbox node and the blade node, and comparing them with the historical normal linkage characteristics, these cosine similarity values ​​are combined to generate a set of associated features that reflects the deviation of the synchronization change pattern between components. This set of associated features can accurately reflect abnormal linkage between different components.

[0080] Step S124: Through the wind turbine feature learning network, the independent feature group and the associated feature group are input into the feature standardization layer for standardization processing to obtain the final independent feature group and associated feature group.

[0081] After generating independent and associated feature sets, they need to be input into the feature normalization layer of the wind turbine feature learning network for normalization. The purpose of feature normalization is to eliminate differences in dimensions and numerical ranges between different features, ensuring that all features have the same scale and distribution. In a wind turbine system, independent and associated feature sets contain various types of features, such as time-series fluctuation patterns of component measurements and synchronous anomaly patterns of cross-component measurements. These features may have different dimensions and numerical ranges. Directly inputting these features into the fault diagnosis model may cause the model to be overly sensitive to some features while ignoring others, thus affecting the accuracy of fault diagnosis. After processing by the feature normalization layer, the resulting independent and associated feature sets have better comparability and consistency, and can be used more effectively in the subsequent fault diagnosis process.

[0082] Step S130: Call the pre-trained component fault diagnosis model to jointly discriminate the independent feature group and the associated feature group, and generate a diagnostic intermediate result containing the component anomaly confidence and the anomaly linkage strength. The diagnostic intermediate result is dynamically calibrated through the component association knowledge base associated with the component fault diagnosis model.

[0083] Step S131: Process the independent feature group through the first fully connected layer of the component fault diagnosis model to output the preliminary anomaly confidence of each component.

[0084] The component fault diagnosis model comprises multiple layers and modules, with the primary function of the first fully connected layer being the processing of independent feature sets. An independent feature set is a multi-dimensional collection of features containing information such as the temporal fluctuation patterns of measurements from various components.

[0085] The first fully connected layer consists of multiple neurons and weight parameters. When independent feature groups are input into the first fully connected layer, the neurons within the layer perform weighted summation of the input features and apply a non-linear transformation through an activation function. This process integrates and abstracts the information from the independent feature groups, thereby outputting the preliminary anomaly confidence score for each component.

[0086] The initial anomaly confidence score for each component is a numerical value representing the probability that the component will exhibit an anomaly in its current state. For example, for the gearbox component, the first fully connected layer, based on information such as the gearbox's rising slope feature, stationary variance feature, and falling decay feature from the independent feature group, performs a series of calculations and processing to output a numerical value representing the gearbox's initial anomaly confidence score. Similarly, for other components such as generators and blades, corresponding initial anomaly confidence scores will also be output.

[0087] The weight parameters of the first fully connected layer are learned during the training of the component fault diagnosis model. These weight parameters are adjusted and optimized based on a large amount of training data so that the initial anomaly confidence of the output can accurately reflect the abnormal situation of the component.

[0088] Step S132: Process the associated feature group through the second fully connected layer of the component fault diagnosis model to output the preliminary abnormal linkage strength of each pair of components.

[0089] In addition to processing independent feature groups, the second fully connected layer of the component fault diagnosis model is responsible for processing associated feature groups. These associated feature groups contain information such as synchronization anomaly patterns between components, reflecting the abnormal linkage between different components.

[0090] The second fully connected layer also consists of multiple neurons and weight parameters. When the associated feature set is input into the second fully connected layer, the neurons within the layer perform a weighted summation of the input features and a non-linear transformation through an activation function. Similar to the first fully connected layer, this process integrates and abstracts the information in the associated feature set, thereby outputting the initial abnormal linkage strength of each pair of components.

[0091] The initial abnormal linkage strength for each pair of components is a numerical value, representing the degree of abnormal linkage between the components. For example, for the gearbox and generator pair, the second fully connected layer will output a numerical value representing the initial abnormal linkage strength between the gearbox and generator after a series of calculations and processing based on the linkage characteristics of the gearbox and generator in the associated feature group, such as the deviation of their synchronization change patterns. Similarly, for other component pairs, such as gearbox and blades, generator and converter, the corresponding initial abnormal linkage strength will also be output.

[0092] The weight parameters of the second fully connected layer are also learned during the training of the component fault diagnosis model. The adjustment and optimization of these weight parameters are to ensure that the initial abnormal linkage strength of the output can accurately reflect the abnormal linkage between components.

[0093] Step S133: Call the component association knowledge base associated with the component fault diagnosis model. The component association knowledge base stores the fault influence weights between each component of the wind turbine. Based on the fault influence weights, the preliminary anomaly confidence and preliminary anomaly linkage strength are weighted and corrected to generate calibrated component anomaly confidence and anomaly linkage strength as intermediate diagnostic results.

[0094] After obtaining the initial anomaly confidence level of each component and the initial anomaly linkage strength of each pair of components, the next step is to call the component association knowledge base associated with the component fault diagnosis model for dynamic calibration.

[0095] The component association knowledge base stores the fault impact weights between various components of a wind turbine generator. This weighting is derived through in-depth analysis and statistics of a large amount of historical fault data, reflecting the degree of mutual influence between faults in different components. For example, in a wind turbine generator system, a gearbox fault may have a significant impact on generator operation; therefore, the fault impact weight between the gearbox and the generator would be relatively large in the component association knowledge base. Conversely, some components have less mutual influence, and their respective fault impact weights would be relatively small.

[0096] Based on the fault impact weights in the component association knowledge base, the initial anomaly confidence level and initial anomaly linkage strength can be weighted and adjusted. For the initial anomaly confidence level, each component can be adjusted according to its fault impact weights with other components. For example, if the fault impact weight between the gearbox and generator is significant, and the generator's initial anomaly confidence level is high, then the gearbox's anomaly confidence level can be appropriately increased to reflect the impact of the generator fault on the gearbox.

[0097] The initial abnormal linkage strength will also be corrected based on the weight information in the component association knowledge base. For example, the initial abnormal linkage strength of the gearbox and generator pair of components can be adjusted by combining the fault impact weights between them, so that the corrected abnormal linkage strength can more accurately reflect the actual linkage between them.

[0098] By using the weighted correction method described above, the calibrated component anomaly confidence and anomaly linkage strength are obtained. Combining these results forms the final diagnostic intermediate result, which more accurately reflects the abnormal conditions of components in the wind turbine system and the abnormal linkage relationships between components.

[0099] Step S140: Based on the intermediate diagnostic results, identify the main faulty components and accompanying faulty components in the wind turbine generator system, and determine the propagation path of the fault between components and the causal relationship of the abnormality of each component.

[0100] Step S141: Extract the target component with the highest confidence level of component abnormality from the intermediate diagnostic results, and determine it as the primary fault component with the earliest fault start time by combining the time point when the deviation of the fluctuation pattern of the target component in the independent feature group first exceeds the historical normal range.

[0101] In the intermediate diagnostic results, the component anomaly confidence level reflects the likelihood of each component exhibiting anomalies. First, the target component with the highest anomaly confidence level is identified from the intermediate diagnostic results. This target component is most likely the source of the fault in the current diagnostic results, but it cannot be directly identified as the primary fault component; its fluctuation pattern deviation within the independent characteristic group needs to be considered.

[0102] The independent feature group contains information such as the time-series fluctuation patterns of component measurements. The deviation of the fluctuation pattern reflects the difference between the actual operating condition of the component and its historical normal operating condition. By analyzing the time point when the fluctuation pattern deviation of the target component in the independent feature group first exceeds the historical normal range, it is possible to further determine whether the component is the one with the earliest failure onset time.

[0103] In practice, for an independent feature set of a target component, the deviation of its fluctuation pattern can be checked step by step over time. For example, for an independent feature set of a gearbox, this includes the rising slope characteristic, the stable variance characteristic, and the falling attenuation characteristic of its vibration measurements. The actual values ​​of these features are compared with the historical normal range, and the time point when the deviation of a certain feature first exceeds the historical normal range is recorded.

[0104] Step S142: Based on the mechanical transmission structure of the wind turbine, construct a physical transmission chain starting from the main faulty component.

[0105] After identifying the primary faulty component, the next step is to construct a physical transmission chain starting from the primary faulty component, based on the mechanical transmission structure of the wind turbine. A wind turbine is a complex mechanical system, with its various components interconnected through specific mechanical structures, forming an ordered transmission relationship.

[0106] Taking a common wind turbine as an example, its basic mechanical transmission process is as follows: the blades rotate under the action of wind, converting wind energy into mechanical energy. The rotation of the blades is transmitted to the gearbox via the main shaft. The gearbox's function is to convert the low-speed rotation of the blades into high-speed rotation to meet the generator's operating requirements. The power, increased by the gearbox, is transmitted to the generator via the drive shaft, where the generator converts mechanical energy into electrical energy. The electrical energy output by the generator is then processed by equipment such as converters before finally being transmitted to the power grid.

[0107] Once the primary faulty component is identified, a physical transmission chain can be constructed starting from that component and following the direction of mechanical transmission. For example, if the primary faulty component is a gearbox, then the starting point of the physical transmission chain is the gearbox, and the next node along the transmission direction is the generator, because the power from the gearbox will be transmitted to the generator. If there are other components related to the power generation process downstream of the generator, such as converters, they will also be added to the physical transmission chain in sequence.

[0108] Step S143: In the physical transmission chain, components with abnormal linkage strength higher than a preset threshold that are directly connected to the main fault component are selected in sequence according to the direct connection relationship between components, and are regarded as first-level accompanying fault components.

[0109] After constructing the physical transmission chain starting from the main faulty component, it is necessary to select first-level accompanying faulty components within this physical transmission chain. The selection is based on the direct connection relationship between the components and the comparison between the abnormal linkage strength and the preset threshold.

[0110] In a physical transmission chain, components are arranged sequentially according to the mechanical transmission order, and each component is directly connected to its adjacent components. When screening for components with accompanying failures at the first level, the abnormal linkage strength between them and the main failure component is checked sequentially, starting with the components directly connected to the main failure component.

[0111] The intensity of abnormal linkage, obtained from previous diagnostic intermediate results, reflects the degree of correlation between abnormal conditions in two components. The preset threshold is a pre-defined standard used to determine whether the abnormal linkage between components has reached a level requiring attention. If the intensity of abnormal linkage between a component and the primary faulty component exceeds the preset threshold, it indicates a strong abnormal correlation between them, and the component is likely affected by the failure propagation from the primary faulty component.

[0112] For example, in a physical transmission chain, if the primary faulty component is the gearbox, its directly connected components include the generator and blades. For the generator, the strength of its abnormal linkage with the gearbox is checked. If this strength exceeds a preset threshold, the generator is identified as a primary accompanying faulty component. Similarly, the same check is performed on the blades; if their abnormal linkage strength with the gearbox exceeds a preset threshold, the blades are also classified as primary accompanying faulty components.

[0113] Step S144: Downstream of the physical connection of the first-level accompanying fault component, select components whose abnormal linkage strength with the first-level accompanying fault component is higher than a preset threshold as second-level accompanying fault components.

[0114] After identifying the primary associated faulty component, the next step is to screen for secondary associated faulty components downstream of the physical connections of the primary associated faulty component. This step is based on the principle that faults in a wind turbine system can propagate along the mechanical transmission structure.

[0115] In a physical transmission chain, a first-level component with a associated fault has upstream and downstream connections with other components. Starting from the first-level component with the associated fault, locate the components physically connected downstream along the direction of mechanical transmission. Then, examine the abnormal linkage strength between these downstream components and the first-level component with the associated fault.

[0116] Similarly, the abnormal linkage strength is obtained from the diagnostic intermediate results, and it measures the degree of correlation between abnormal conditions between two components. These abnormal linkage strengths are compared with preset thresholds. If the abnormal linkage strength between a downstream component and a first-level accompanying fault component is higher than the preset threshold, then that component will be screened as a second-level accompanying fault component.

[0117] For example, suppose the primary accompanying fault component is the generator. Downstream of the generator's physical connection, there may be components such as a converter. For the converter, check the strength of its abnormal linkage with the generator. If this strength exceeds a preset threshold, the converter will be identified as a secondary accompanying fault component.

[0118] Step S145: Connect the main fault component, the first-level accompanying fault component, and the second-level accompanying fault component in series according to the mechanical transmission sequence to form a fault propagation path that reflects the physical transmission relationship.

[0119] After identifying the main fault component, the primary accompanying fault component, and the secondary accompanying fault component, they are connected in series according to the mechanical transmission sequence to form a fault propagation path that reflects the physical transmission relationship.

[0120] The mechanical transmission structure of a wind turbine determines the power transmission sequence between its components. The primary faulty component, as the source of the fault, is at the beginning of the fault propagation path. Primary accompanying faulty components are those first affected by the propagation of the primary faulty component's fault and are immediately following the primary faulty component in the mechanical transmission sequence. Secondary accompanying faulty components are those affected after the primary accompanying faulty components, due to the further propagation of the primary accompanying faulty component's fault, and are located further downstream in the mechanical transmission sequence.

[0121] For example, if the primary faulty component is the gearbox, the primary accompanying faulty component is the generator, and the secondary accompanying faulty component is the converter, then according to the mechanical transmission sequence, the rotation of the blades transmits power through the gearbox. A fault in the gearbox will affect the generator, and a fault in the generator may in turn affect the converter. Connecting the gearbox, generator, and converter in series in sequence forms the fault propagation path: gearbox -> generator -> converter.

[0122] Step S146: Compare the abnormal start time of the main fault component and each of the accompanying fault components. If the abnormal start time of the accompanying fault component is later than that of the main fault component and the time difference is within the standard response delay range between components, it is determined that there is a causal relationship between the main fault component and the accompanying fault component. If the time difference exceeds the standard response delay range, it is determined to be an independent abnormality.

[0123] Step S1461: Obtain the specific moment when the fluctuation pattern deviation of the main fault component in the independent feature group first exceeds the historical normal range, and take it as the start time of the main fault.

[0124] To accurately determine the causal relationship between the primary faulty component and its accompanying faulty components, it is first necessary to obtain the onset time of the primary fault of the primary faulty component. The onset time of the primary fault is determined by analyzing the specific moment when the deviation of the fluctuation pattern of the primary faulty component in the independent characteristic group first exceeds the historical normal range.

[0125] The independent feature set contains the time-series fluctuation patterns of the measured values ​​of the primary faulty component, such as the slope characteristics of the rising segment, the variance characteristics of the stationary segment, and the decay characteristics of the falling segment. The actual values ​​of these features change with the operating status of the primary faulty component. The historical normal range is obtained through statistical analysis of a large amount of normal operating data; it is a multi-dimensional interval, with each dimension corresponding to a feature in the independent feature set.

[0126] In practice, the independent characteristic groups of the main faulty component can be checked step by step over time. For each characteristic value at each time step, it is compared with the historical normal range. When the deviation of a certain characteristic first exceeds the historical normal range, the specific moment is recorded and taken as the start time of the main fault.

[0127] For example, if the gearbox is the main faulty component, the slope characteristic of the rising segment of the vibration measurement value in its independent characteristic group, if at a certain time step, the deviation of the actual value of this characteristic from the historical normal range exceeds the preset range for the first time, then the specific time step corresponding to that time step is the start time of the main fault.

[0128] To ensure the accuracy of the onset time of the main fault, the continuity and accuracy of the time series must be guaranteed when recording the deviation of the fluctuation pattern. Simultaneously, the determination of the historical normal range should be based on a sufficient amount of normal operation data to ensure that it truly reflects the operating status of the main faulty component under normal conditions.

[0129] Step S1462: Obtain the specific moment when the deviation of the accompanying fault component in the independent feature group first exceeds the historical normal range, and take it as the starting moment of the accompanying fault.

[0130] Similar to determining the onset time of the main fault, for components with accompanying faults, it is also necessary to obtain the specific moment when the deviation of its fluctuation pattern in the independent characteristic group first exceeds the historical normal range, as the onset time of the accompanying fault.

[0131] The independent feature set accompanying the faulty component also includes the time-series fluctuation pattern of the component's measurements, such as the slope characteristics of the rising segment, the variance characteristics of the stationary segment, and the decay characteristics of the falling segment. These features reflect the actual operating state of the component accompanying the fault.

[0132] In practice, the independent feature groups of the accompanying faulty component are checked step by step over time. The feature value at each time step is compared with the historical normal range. The historical normal range is obtained through statistical analysis of a large amount of normal operating data of the accompanying faulty component, and is a multi-dimensional interval, with each dimension corresponding to a feature in the independent feature group.

[0133] When the deviation of a certain feature in the independent feature group of the accompanying fault component first exceeds the historical normal range, record the specific moment and take it as the start time of the accompanying fault.

[0134] For example, if the accompanying fault component is a generator, the stationary segment variance characteristic of the temperature measurement value in its independent characteristic group, when the deviation of the actual value of this characteristic from the historical normal range first exceeds the preset range, the specific time corresponding to this time step is recorded, which is the start time of the accompanying fault of the generator.

[0135] Step S1463: Calculate the time interval between the start time of the main fault and the start time of the accompanying fault.

[0136] After obtaining the start times of the primary fault and the accompanying fault, the next step is to calculate the time interval between these two times. Calculating this time interval is crucial for determining the causal relationship between the primary faulty component and the accompanying faulty component.

[0137] The process of calculating the time interval involves subtracting the start time of the primary fault from the start time of the accompanying fault. It is important to ensure consistency in the time format during the calculation. For example, if both the start times of the primary and accompanying faults are represented by a specific timestamp, then the time interval can be obtained directly through subtraction.

[0138] This time interval reflects the time elapsed from the occurrence of an anomaly in the primary faulty component to the occurrence of an anomaly in the accompanying faulty component. If the time interval falls within the standard response delay range between the primary and accompanying faulty components, then a causal relationship may exist; if the time interval does not fall within the standard response delay range, then a causal relationship may not exist. It should be noted that this standard response delay range is generated by the mechanical transmission or energy transfer between the two components.

[0139] For example, if the primary faulty component is the gearbox and the secondary faulty component is the generator, and the primary fault in the gearbox begins at time T1, while the secondary fault in the generator begins at time T2, then the time interval ΔT = T2 - T1. By calculating this time interval, we can further combine it with the standard response delay range to determine whether the generator anomaly is caused by the propagation of a fault from the gearbox.

[0140] Step S1464: Query the response characteristic database of wind turbine components to obtain the standard response delay range between the main fault component and the accompanying fault component.

[0141] In order to accurately determine whether there is a causal relationship between the primary faulty component and the accompanying faulty component, after calculating the time interval, it is necessary to query the response characteristic database of the wind turbine components to obtain the standard response delay range between the primary faulty component and the accompanying faulty component.

[0142] The wind turbine component response characteristic database is a database built upon a large amount of experimental data and practical operating experience. During the research and development and operation of wind turbines, numerous experiments and monitoring were conducted on the mechanical transmission and energy transfer between different components. This recorded the time range required for one component to malfunction and subsequently cause another component to exhibit abnormalities under normal conditions.

[0143] For each pair of components, such as gearbox and generator, generator and converter, the database stores the standard response delay range between them. This standard response delay range is a multi-dimensional interval that takes into account time variations under different operating conditions.

[0144] When it is necessary to determine the causal relationship between a primary faulty component and a secondary faulty component, the standard response delay range between them is retrieved from the response characteristic database based on the types of the primary and secondary faulty components. For example, if the primary faulty component is a gearbox and the secondary faulty component is a generator, the standard response delay range between the gearbox and the generator is found in the database.

[0145] Step S1465: If the time interval is within the standard response delay range, it is determined that the abnormality of the accompanying faulty component is caused by the fault propagation of the main faulty component.

[0146] After calculating the time interval between the start time of the primary fault and the start time of the accompanying fault, and querying the standard response delay range between the primary fault component and the accompanying fault component, the next step is to determine the causal relationship.

[0147] If the time interval is within the standard response delay range, it can be determined that the anomaly of the accompanying faulty component is caused by the fault propagation of the primary faulty component. This is based on the mechanical transmission and energy transfer principles of wind turbine systems. When the primary faulty component fails, its fault will affect the connected accompanying faulty components through mechanical structure or energy transfer.

[0148] Under normal circumstances, the time required from the failure of the primary faulty component to the occurrence of an anomaly in the accompanying faulty component follows a certain pattern, which is reflected within the standard response delay range. If the actual time interval conforms to this range, it indicates that the anomaly of the accompanying faulty component occurred within a reasonable timeframe due to the impact of the primary faulty component's failure.

[0149] For example, the primary faulty component might be the gearbox, and the secondary faulty component might be the generator. When the gearbox fails, its fault will affect the generator through mechanical transmission. If the generator's anomaly starts later than the gearbox's, and the time interval is within the standard response delay range between the gearbox and generator, then the generator's anomaly can be considered to be caused by the gearbox's fault. This determination of causality clarifies the fault propagation path and the scope of its impact. Once it is determined that the anomaly of the secondary faulty component is caused by the propagation of the primary faulty component, the focus of maintenance can be placed on the primary faulty component, while the secondary faulty component undergoes corresponding inspection and maintenance.

[0150] For example, after determining that a gearbox failure is causing the generator malfunction, maintenance personnel can first conduct a comprehensive inspection and repair of the gearbox to address the root cause of the problem. During gearbox repair, the generator should also be inspected simultaneously to check for further damage caused by the gearbox failure. If minor damage is found in the generator, even if it may be a consequence of the gearbox malfunction, it should still be repaired promptly to prevent the problem from escalating.

[0151] From a technical perspective, when a major faulty component malfunctions, its physical state changes. Taking a gearbox as an example, if the gears in the gearbox wear out or are damaged, the meshing between the gears becomes irregular, leading to increased vibration and unstable torque transmission. This unstable vibration and torque are transmitted to the generator via the drive shaft, causing uneven forces on the generator rotor, thus affecting the generator's normal operation and leading to abnormal conditions such as increased generator temperature and fluctuating output power. The standard response delay range is determined based on this physical transmission process and the equipment's response characteristics. When the time interval is within this range, it indicates that the physical process of fault propagation is reasonable and as expected.

[0152] Step S1466: If the time interval is not within the standard response delay range, it is determined that the abnormality of the accompanying faulty component is not causally related to the main faulty component and is an independently occurring abnormality.

[0153] When the calculated time interval between the start time of the primary fault and the start time of the secondary fault is not within the standard response delay range between the primary fault component and the secondary fault component, it is necessary to determine that the anomaly of the secondary fault component is not causally related to the primary fault component and is an anomaly that occurs independently.

[0154] From a temporal logic perspective, if the time interval is less than the standard response delay range, it means that the accompanying faulty component has already exhibited abnormalities before the effects of the primary faulty component's failure have properly propagated to it. For example, assuming the primary faulty component is the gearbox and the accompanying faulty component is the generator, according to normal mechanical transmission and fault propagation processes, it takes a certain amount of time for the gearbox failure to affect the generator. If the generator exhibits abnormalities before this time has elapsed, then the generator's abnormality is likely caused by its own internal issues, such as short circuits in the generator's windings or insulation aging, which are not directly related to the gearbox failure.

[0155] If the time interval exceeds the standard response delay range, it indicates that an excessively long period elapsed after the primary faulty component failed before the accompanying faulty component exhibited abnormalities. In this case, it is unlikely that the failure was directly propagated from the primary faulty component. For example, during this extended period, the accompanying faulty component might have been affected by other external factors, or some of its own latent faults might have only gradually manifested. Taking a generator as an example, if the generator only exhibits abnormalities a long time after a gearbox failure, it's possible that the generator was affected by environmental factors during operation, such as sudden changes in humidity or temperature, or that the generator's cooling system malfunctioned during this period, causing the generator temperature to rise and triggering the abnormality.

[0156] After determining that the malfunction of the accompanying component is an independent anomaly, it is necessary to conduct separate troubleshooting and analysis of the accompanying component. This includes a detailed inspection and monitoring of various operating parameters of the accompanying component, such as generator voltage, current, temperature, and speed. By analyzing these parameters, the specific cause of the accompanying component's malfunction can be identified. For example, if an abnormally high generator temperature is found, and inspection reveals that the cooling system fan is not rotating, then a cooling system malfunction can be determined as the cause of the generator malfunction.

[0157] Furthermore, for cases of independent anomalies, it is also necessary to consider the possibility of multiple independent faults occurring simultaneously. Sometimes, multiple components in a wind turbine system may fail simultaneously for different reasons. For example, the gearbox may fail due to its own mechanical wear, while the generator may simultaneously malfunction due to problems with the electrical system. In such cases, it is necessary to address the fault of each component separately and develop corresponding maintenance plans.

[0158] Step S1467: Generate a description of the causal relationships between components in the fault propagation path based on the above determination results.

[0159] After determining the causal relationship between the main faulty component and the accompanying faulty components, it is necessary to generate a description of the causal relationship between each component in the fault propagation path based on the determination results.

[0160] Causal relationship descriptions help maintenance and technical personnel better understand the occurrence and development of faults. For component pairs determined to have a causal relationship, the causal relationship description clearly indicates which component is the source of the fault, which component is affected, and how the fault propagates from the source component to the affected component.

[0161] For example, if it is determined that the gearbox fault propagation caused the generator abnormality, the causal relationship can be described as follows: "As the main faulty component, the gearbox, due to the wear or damage of its internal gears, leads to increased vibration and unstable torque transmission. This unstable vibration and torque are transmitted to the generator through the drive shaft, causing the generator rotor to be subjected to uneven forces, which in turn leads to abnormal conditions such as increased generator temperature and fluctuating output power."

[0162] For components determined to be independently malfunctioning, the causal description will explain that the malfunction of that component is unrelated to the failure of other components, and the possible causes of the malfunction. For example, if a generator malfunction is determined to be an independently malfunction caused by a cooling system failure, the causal description could be: "The generator malfunction is not causally related to the gearbox failure, but is due to a failure in the generator's own cooling system, such as a non-rotating cooling fan or insufficient coolant, which causes the generator temperature to rise, thus triggering the generator malfunction."

[0163] Step S150: Generate a fault warning instruction containing component location information based on the main fault component, the accompanying fault component and the fault propagation path, and send the fault warning instruction to the wind turbine operation and maintenance system to trigger targeted maintenance operations. The component location information is associated with the physical installation location of the component inside the wind turbine tower.

[0164] Step S151: Query the installation location information of the main fault component, the first-level accompanying fault component, and the second-level accompanying fault component from the physical structure database of the wind turbine. The installation location information includes the tower height and the coordinates inside the nacelle.

[0165] To generate accurate fault warning commands, it is necessary to query the installation location information of the primary fault component, the first-level accompanying fault component, and the second-level accompanying fault component from the wind turbine's physical structure database. The physical structure database is an important data record in the design and construction process of wind turbines, storing in detail the specific location of each component within the wind turbine tower.

[0166] In this embodiment, tower height information helps maintenance personnel understand the vertical location of components. For example, some components may be installed at the top of the tower, requiring special climbing equipment to reach; while other components may be installed in the middle or bottom of the tower, making them relatively easier to access. Cabin coordinate information helps maintenance personnel accurately locate components within the cabin. The cabin typically contains multiple pieces of equipment and complex wiring; accurate coordinate information can save time in locating components.

[0167] When querying installation location information, it is necessary to retrieve the information from the physical structure database based on the component's name or number. The data in the database is categorized and stored according to certain rules, typically assigning a unique identifier to each component. For example, a gearbox might have a specific number, which allows for quick location of the gearbox's installation location within the database.

[0168] Step S152: Based on the fault propagation path, mark the primary fault component as the priority maintenance location, the first-level accompanying fault component as the secondary priority maintenance location, and the second-level accompanying fault component as the associated maintenance location.

[0169] After obtaining the installation location information of the primary faulty component, the first-level accompanying faulty components, and the second-level accompanying faulty components, the maintenance priorities of these components are marked based on the fault propagation path. The fault propagation path clearly shows how the fault propagates from the primary faulty component to the first-level and second-level accompanying faulty components, and the maintenance priorities can be reasonably determined based on this process.

[0170] The primary faulty component, as the source of the failure, has the greatest impact on the entire wind turbine system. If the problem with the primary faulty component is not repaired promptly, the fault may continue to spread, causing more components to fail, and potentially even leading to a shutdown of the entire wind turbine system. Therefore, the primary faulty component should be marked as a priority for maintenance, and maintenance personnel must conduct a comprehensive inspection and repair of it first.

[0171] Primary auxiliary fault components are directly connected to the primary fault component and are most directly and obviously affected by the fault propagation from the primary fault component. Although their faults may be caused by the primary fault component, they may also suffer more severe damage during the fault propagation process due to their own potential problems. Therefore, primary auxiliary fault components are marked as secondary priority maintenance locations, and their inspection and maintenance should be carried out simultaneously or after the maintenance of the primary fault component.

[0172] Secondary accompanying fault components are less affected by faults; their abnormalities may arise from further fault propagation after a primary accompanying fault component malfunctions. Although secondary accompanying fault components also require inspection, their priority is relatively low. Marking secondary accompanying fault components as associated maintenance locations allows maintenance personnel to perform appropriate inspections of these components as needed during the maintenance of the primary and primary accompanying fault components.

[0173] When prioritizing maintenance, the priority information needs to be linked to the installation location information of the components. This allows maintenance personnel to quickly develop a maintenance plan upon arrival at the wind turbine site, based on the priority and installation location information, thus improving maintenance efficiency. For example, if the main faulty component is installed at the top of the tower, maintenance personnel can prepare their climbing equipment first and prioritize the maintenance of the main faulty component; during the maintenance of the main faulty component, they can then arrange the maintenance of secondary and primary components as needed.

[0174] Step S153: Match the corresponding maintenance strategy from the fault case library according to the type of the main fault component and the deviation of its independent feature group.

[0175] After determining the maintenance priorities of the components, specific maintenance strategies need to be developed for the maintenance work. Based on the type of the main faulty component and the deviation of its independent characteristic groups, corresponding maintenance strategies are matched from the fault case database.

[0176] The fault case library is a database that stores a large number of wind turbine fault cases and corresponding maintenance strategies. During the operation of wind turbines, various types of fault cases can be accumulated. By analyzing and summarizing these cases, effective maintenance strategies for different fault situations can be obtained.

[0177] Identifying the type of the primary faulty component is crucial for determining the maintenance strategy. Different types of components have different structures and operating principles, leading to different causes and manifestations of failure. For example, gearbox failures may primarily involve gear wear and bearing damage, while generator failures may involve winding short circuits and insulation aging. Therefore, different maintenance methods are required for different types of primary faulty components.

[0178] The deviation of the independent characteristic group reflects the severity of the main faulty component. The larger the deviation, the more serious the component's fault, requiring more rigorous and comprehensive maintenance measures. For example, if the vibration deviation in the gearbox's independent characteristic group is large, it indicates that the gearbox's vibration has seriously deviated from the normal range, possibly due to gear damage or loose installation. In this case, a complete disassembly and inspection of the gearbox is necessary, including replacing damaged gears or readjusting their installation positions.

[0179] When matching maintenance strategies from the fault case database, the type of the main faulty component and the deviation of independent feature groups need to be used as search criteria. The database will filter and match based on these criteria to find the most suitable maintenance strategy. For example, if the main faulty component is a gearbox and the vibration deviation is within a certain range, the database will search all fault cases that meet the search criteria and select the most effective maintenance strategy from them.

[0180] To ensure the accuracy and effectiveness of the matching results, the fault case database needs to be continuously updated and improved. With the continuous development of wind turbine technology and the emergence of new fault cases, it is necessary to promptly add this new information to the fault case database and optimize and adjust the existing maintenance strategies.

[0181] Step S154: Integrate component installation location information, maintenance priority, and maintenance strategy into a fault warning command.

[0182] After obtaining component installation location information, determining maintenance priorities, and matching maintenance strategies, this information needs to be integrated into fault warning commands. A fault warning command is a comprehensive set of information that gathers all important information related to fault repair, enabling the operations and maintenance system to accurately understand the fault situation and take appropriate measures.

[0183] First, component installation location information is a fundamental part of fault warning commands, helping maintenance personnel quickly locate faulty components and reduce the time spent searching for them. During integration, detailed information such as tower height and nacelle coordinates must be clearly included in the commands. For example, the component's installation location can be described in text form, such as "The gearbox is installed at the top of the tower, with nacelle coordinates of (X, Y, Z)".

[0184] Maintenance priority information provides a guiding sequence for maintenance work. The instructions must clearly indicate that the primary faulty component is the highest priority maintenance location, the first-level accompanying faulty component is the next highest priority maintenance location, and the second-level accompanying faulty component is the associated maintenance location. Specific identifiers or text descriptions can be used to highlight these priorities; for example, adding "(Priority)," "(Next Priority)," or "(Associated)" after the component name.

[0185] Maintenance strategy information is the core component of fault warning instructions, detailing the specific maintenance methods for different components. During integration, the maintenance strategy needs to be expressed in a clear and easily understandable manner. For example, if the maintenance strategy for the gearbox is "disassemble and inspect the gears and bearings, replace damaged parts," then this strategy must be fully included in the instruction.

[0186] When integrating the above information, the installation location and maintenance strategy for each component can be listed sequentially according to their priority. Furthermore, to facilitate understanding by the operation and maintenance system and maintenance personnel, necessary instructions and explanations can be added to the instructions. For example, detailed instructions can be provided for specific maintenance tools or operating procedures.

[0187] Step S155: Send the fault warning instruction to the operation and maintenance system so that the operation and maintenance system can dispatch maintenance personnel and allocate maintenance tools according to the fault warning instruction.

[0188] Furthermore, for example, the method may also include:

[0189] Step S210: Obtain the training sample data set and the corresponding training sample label set of the wind turbine system. The training sample data set contains multiple sets of training sample data. Each set of training sample data contains a training independent feature set reflecting the abnormal tendency of a single component and a training correlation feature set reflecting the abnormal linkage between components. The training sample label set contains the true confidence of component abnormality and the true strength of abnormal linkage corresponding to each set of training sample data.

[0190] In this step, the acquisition of the training sample dataset relies on measurement data from sensors on various components of the wind turbine system. As mentioned earlier, synchronous sampling from gearbox vibration sensors, generator temperature sensors, and blade speed sensors over continuous periods is used to process and obtain training independent feature sets and training associated feature sets. The training sample label set is based on annotations from numerous real-world fault cases and professional diagnostic experience, ensuring the accuracy and reliability of the true confidence level of component anomalies and the true strength of anomaly linkages. To guarantee the model's generalization ability, the training samples must cover data under different operating conditions.

[0191] Step S220: Input the training independent feature group and the training associated feature group into the component fault diagnosis model to be trained, process the training independent feature group through the first fully connected layer of the component fault diagnosis model, and output the predicted anomaly confidence of each component; process the training associated feature group through the second fully connected layer of the component fault diagnosis model, and output the predicted anomaly linkage strength of each pair of components.

[0192] The training independent feature sets are input into the first fully connected layer of the component fault diagnosis model. The neurons in this first fully connected layer perform a weighted summation of the input features, with the weight parameters continuously adjusted during training. The weighted summation result undergoes a non-linear transformation using an activation function, outputting the predicted anomaly confidence score for each component, reflecting the model's judgment on the probability of anomalies occurring in each component. For the training associated feature sets, they are input into the second fully connected layer. Similarly, the neurons in the second fully connected layer perform a weighted summation and a non-linear transformation, outputting the predicted anomaly linkage strength for each pair of components, reflecting the degree of predictive anomaly linkage between components.

[0193] Step S230: Call the component association knowledge base associated with the component fault diagnosis model. The component association knowledge base stores the fault influence weights between each component of the wind turbine. Based on the fault influence weights, the predicted anomaly confidence and the predicted anomaly linkage strength are weighted and corrected to generate the calibrated component anomaly confidence and anomaly linkage strength as the model prediction results.

[0194] The component association knowledge base is built upon a large amount of historical fault data and in-depth research on wind turbine systems. The stored fault influence weights among components reflect the degree of mutual influence between component faults. After obtaining the predicted anomaly confidence and predicted anomaly linkage strength, a weighted correction is applied based on the fault influence weights in the knowledge base. For the predicted anomaly confidence, adjustments are made for each component based on its fault influence weights with other components. For example, if the fault influence weights of the gearbox and generator are significant, and the generator's predicted anomaly confidence is high, the gearbox's predicted anomaly confidence can be appropriately increased. Similarly, the predicted anomaly linkage strength is adjusted using corresponding weights to make the corrected linkage strength more consistent with reality. After weighted correction, the calibrated component anomaly confidence and anomaly linkage strength are obtained, constituting the model's prediction results.

[0195] Step S240: Calculate the first difference between the calibrated component anomaly confidence in the model prediction result and the corresponding component anomaly true confidence in the training sample label set; calculate the second difference between the calibrated abnormal linkage strength and the corresponding abnormal linkage true strength in the training sample label set; and combine the first difference and the second difference after standardization to generate the training loss value.

[0196] To evaluate the accuracy of the model's predictions, the difference between the model's predictions and the training sample label set needs to be calculated. For the calibrated component anomaly confidence score and the corresponding true anomaly confidence score, a first difference value is calculated. This calculation compares the calibrated anomaly confidence score and the true anomaly confidence score for each component to obtain a numerical value reflecting the difference. Similarly, a second difference value is calculated between the calibrated anomaly linkage strength and the corresponding true anomaly linkage strength. Since the first and second difference values ​​may have different dimensions and numerical ranges, they need to be standardized for easier processing and comparison. Standardization brings the difference values ​​to the same scale, eliminating the influence of dimensions. Finally, the standardized first and second difference values ​​are combined to generate the training loss value. The training loss value comprehensively reflects the degree of difference between the model's predictions and the true labels, and is an important indicator for measuring model performance.

[0197] Step S250: Optimize the network weight parameters of the first fully connected layer and the second fully connected layer of the component fault diagnosis model and the fault impact weights in the component-related knowledge base using the backpropagation algorithm based on the training loss value.

[0198] Backpropagation is an important algorithm for training neural networks. Its core idea is to adjust the model's parameters based on the training loss value to reduce the loss and improve the model's accuracy. After obtaining the training loss value, the backpropagation algorithm starts from the model's output layer and propagates the loss value back to the first and second fully connected layers. During backpropagation, the gradient of the loss value with respect to the network weight parameters is calculated. The gradient represents the rate and direction of change of the loss value with respect to the weight parameters. Based on the magnitude and direction of the gradient, the network weight parameters of the first and second fully connected layers are adjusted. For example, if the gradient of a certain weight parameter is positive, it means that increasing the weight parameter will increase the loss value, so the weight parameter will be decreased; conversely, if the gradient is negative, the weight parameter will be increased.

[0199] Simultaneously, the backpropagation algorithm also optimizes the fault impact weights in the component association knowledge base. By calculating the gradient of the training loss value relative to the fault impact weights, the fault impact weights are adjusted, allowing the weights in the knowledge base to more accurately reflect the fault impact relationships between components, further improving the model's predictive performance. After multiple backpropagations and parameter adjustments, the model's network weight parameters and fault impact weights gradually converge to their optimal values, resulting in a continuous decrease in the training loss value and a continuous improvement in the model's performance.

[0200] Step S260: Input the optimized component fault diagnosis model into the verification sample data set for verification testing. The verification sample data set includes multiple sets of training sample data that did not participate in the training and the corresponding verification sample label set. The verification sample label set includes the component anomaly verification confidence and anomaly linkage verification strength corresponding to each set of verification training sample data.

[0201] After optimizing the model parameters, the optimized component fault diagnosis model needs to be validated to evaluate its generalization ability on untrained data. The validation sample dataset consists of multiple sets of untrained training samples, including training independent feature sets reflecting the anomaly tendency of individual components and training associated feature sets reflecting the anomaly linkage between components. The corresponding validation sample label set includes component anomaly validation confidence and anomaly linkage validation strength, which are real labels determined based on actual fault conditions and professional diagnosis. The validation sample data is input into the optimized component fault diagnosis model. Following the previous processing flow, the training independent feature sets and training associated feature sets are processed through the first and second fully connected layers, respectively, outputting the predicted anomaly confidence of the components and the predicted anomaly linkage strength for each pair of components. Then, weighted corrections are made based on the component association knowledge base to obtain the calibrated component anomaly confidence and anomaly linkage strength.

[0202] Step S270: Process the training sample data in the verification sample dataset using the component fault diagnosis model to generate verification prediction results.

[0203] Once the validation sample data is input into the component fault diagnosis model, the entire process is similar to the training phase, but validation sample data that was not used in the training is used. The purpose is to test the predictive ability of the component fault diagnosis model on new data.

[0204] Step S280: Calculate the first deviation between the calibrated component anomaly confidence in the verification prediction result and the component anomaly verification confidence in the verification sample label set; calculate the second deviation between the calibrated anomaly linkage strength and the anomaly linkage verification strength in the verification sample label set; and combine the first deviation and the second deviation after standardization to generate the verification loss value.

[0205] To evaluate the accuracy of the validation predictions, it is necessary to calculate the deviation from the validation sample label set, generating a validation loss value. The validation loss value reflects the model's prediction error on the validation sample data and is an important indicator for evaluating the model's generalization ability.

[0206] Step S290: If the verification loss value is less than the preset loss threshold, the component fault diagnosis model training is completed; if the verification loss value is not less than the preset loss threshold, the calculation weight of the training loss value is adjusted and the model optimization step is repeated until the verification condition is met.

[0207] The preset loss threshold is a pre-defined standard used to determine whether the model's performance meets the requirements. If the validation loss value is less than the preset loss threshold, it indicates that the model's prediction error on the validation sample data is small, and the model's generalization ability is good. At this point, the component fault diagnosis model can be considered to have completed training and can be used for actual fault diagnosis tasks.

[0208] If the validation loss value is not less than the preset loss threshold, it indicates that the requirements have not yet been met and further optimization is needed. At this point, the calculation weights of the training loss value are adjusted. The calculation weights of the training loss value affect the proportions of the first and second difference values ​​in the training loss value. By adjusting these weights, the model's focus on different types of errors during optimization can be changed. For example, if the prediction error of component anomaly confidence is found to be large, the calculation weight of the first difference value can be appropriately increased, making the model pay more attention to optimizing component anomaly confidence. After adjusting the calculation weights, the model optimization steps are repeated, i.e., the backpropagation algorithm is used to optimize the network weight parameters of the first and second fully connected layers of the model and the fault impact weights in the component-related knowledge base based on the new training loss value. Then, validation testing is performed again, calculating the validation loss value until the validation loss value is less than the preset loss threshold, satisfying the validation conditions. Through the above iterative optimization process, the performance and accuracy of the component fault diagnosis model are continuously improved.

[0209] Figure 2 The diagram illustrates exemplary hardware and software components of a fault diagnosis device 100 for a wind turbine system, provided in some embodiments of the present invention, which can implement the inventive concept. For example, a processor 120 may be used in the fault diagnosis device 100 for a wind turbine system and to perform the functions of the present invention.

[0210] The fault diagnosis device 100 for wind turbine systems can be a general-purpose server or a special-purpose server; both can be used to implement the fault diagnosis method for wind turbine systems of the present invention. Although only one server is shown in this invention, for convenience, the functions described in this invention can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0211] For example, the fault diagnosis device 100 for a wind turbine system may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the fault diagnosis device 100 for a wind turbine system may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present invention can be implemented according to these program instructions. The fault diagnosis device 100 for a wind turbine system also includes an I / O interface 150 between a computer and other input / output devices.

[0212] For ease of explanation, only one processor is described in the fault diagnosis device 100 for wind turbine systems. However, it should be noted that the fault diagnosis device 100 for wind turbine systems of the present invention may also include multiple processors, and therefore the steps performed by one processor described in the present invention may also be performed jointly by multiple processors or individually. For example, if the processor of the fault diagnosis device 100 for wind turbine systems performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0213] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned fault diagnosis method for wind turbine generator systems is implemented.

[0214] Figure 3 The invention also illustrates a fault diagnosis system for wind turbine systems provided by some embodiments of the invention, which can realize the ideas of the invention. The system includes a data acquisition module for acquiring system acquisition data of the wind turbine system. The system acquisition data includes synchronous sampling records of gearbox vibration sensors, generator temperature sensors, and blade speed sensors over a continuous period of time. Each sampling record includes the original measurement value of the corresponding component at a single moment.

[0215] The data preprocessing module is used to extract fault-sensitive features from the data collected by the system through the wind turbine feature learning network, and generate independent feature groups reflecting the abnormal tendency of a single component and related feature groups reflecting the abnormal linkage between components. The independent feature groups include the time-series fluctuation pattern of component measurement values, and the related feature groups include the synchronous abnormal pattern of cross-component measurement values.

[0216] The diagnostic intermediate result generation module is used to call the pre-trained component fault diagnosis model to jointly discriminate the independent feature group and the associated feature group, and generate diagnostic intermediate results including component anomaly confidence and anomaly linkage strength. The diagnostic intermediate results are dynamically calibrated through the component association knowledge base associated with the component fault diagnosis model.

[0217] The fault diagnosis module identifies the main faulty components and accompanying faulty components in the wind turbine system based on the intermediate diagnostic results, and determines the propagation path of the fault between components and the causal relationship of the abnormality of each component.

[0218] The fault early warning module generates a fault early warning command containing component location information based on the main fault component, the accompanying fault components and the fault propagation path, and sends the fault early warning command to the wind turbine operation and maintenance system to trigger targeted maintenance operations. The component location information is associated with the physical installation location of the component inside the wind turbine tower.

[0219] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A fault diagnosis method for wind turbine generator systems, characterized in that, The method includes: Acquire system data of the wind turbine system. The system data includes synchronous sampling records of gearbox vibration sensor, generator temperature sensor and blade speed sensor in a continuous period of time. Each sampling record contains the original measurement value of the corresponding component at a single moment. The system collects data by using a wind turbine feature learning network to extract fault-sensitive features, generating independent feature groups that reflect the abnormal tendencies of a single component and associated feature groups that reflect the abnormal linkages between components. The independent feature groups include the time-series fluctuation patterns of component measurements, and the associated feature groups include the synchronous abnormal patterns of cross-component measurements. The pre-trained component fault diagnosis model is invoked to jointly discriminate the independent feature group and the associated feature group, generating an intermediate diagnostic result containing the component anomaly confidence and the anomaly linkage strength. The intermediate diagnostic result is dynamically calibrated through the component association knowledge base associated with the component fault diagnosis model. Based on the diagnostic intermediate results, the main faulty components and accompanying faulty components in the wind turbine system are identified, and the propagation path of the fault between components and the causal relationship of the abnormality of each component are determined. Based on the main fault component, accompanying fault components and fault propagation path, a fault warning instruction containing component location information is generated, and the fault warning instruction is sent to the wind turbine operation and maintenance system to trigger targeted maintenance operations. The component location information is associated with the physical installation location of the component inside the wind turbine tower.

2. The fault diagnosis method for wind turbine systems according to claim 1, characterized in that, The method involves extracting fault-sensitive features from the system's collected data using a wind turbine feature learning network, generating independent feature groups reflecting the abnormal tendencies of individual components and associated feature groups reflecting abnormal linkages between components, including: The system collects data and divides it into synchronous periodic segments according to the power generation cycle of the wind turbine. Each periodic segment contains sampling records of the gearbox, generator, and blades during the complete power generation cycle. Through the wind turbine feature learning network, time-series pattern learning processing is performed on the sampling records of a single component in each cycle segment to generate an independent feature group that reflects the operational stability of the component itself. The independent feature group includes the deviation between the fluctuation pattern and the historical normal fluctuation pattern. The time-series pattern learning processing is implemented through a recurrent neural network to extract the fluctuation pattern of the component measurement value in the rising segment, the stable segment, and the falling segment within the cycle. Through the wind turbine feature learning network, the cross-component sampling records in each cycle segment are processed for linkage mode learning to generate a set of associated features reflecting abnormal linkage between components. The set of associated features includes the deviation between the synchronous change pattern and the historical normal synchronous pattern. The linkage mode learning is implemented through a graph neural network, which constructs a component linkage graph with components as nodes and mechanical connection relationships between components as edges to extract the synchronous change patterns of gearbox vibration and generator temperature, and blade speed and gearbox load. The independent feature set and the associated feature set are input into the feature standardization layer through the wind turbine feature learning network for standardization processing to obtain the final independent feature set and associated feature set.

3. The fault diagnosis method for wind turbine systems according to claim 2, characterized in that, Through the wind turbine feature learning network, time-series pattern learning processing is performed on the sampling records of a single component within each cycle segment to generate independent feature sets reflecting the operational stability of the component itself, including: The sampling record of the single component is input into the input layer of the recurrent neural network of the wind turbine feature learning network. The input layer receives the time series of gearbox vibration values, the time series of generator temperature values, or the time series of blade speed values. The sampled values ​​are processed step by step through the hidden layer of the recurrent neural network, and the hidden state of each time step is recorded. The hidden state contains the relationship between the current value and the value of the previous time step. After the complete periodic segment is processed, key feature sequences are extracted through the output layer of the recurrent neural network. The key feature sequences include: the slope features of the rising segment, the variance features of the stationary segment, and the decay features of the falling segment. Calculate the Euclidean distance between the slope feature, variance feature, attenuation feature and the corresponding feature of the historical normal cycle, and generate an independent feature set that reflects the deviation of the component's own fluctuation law.

4. The fault diagnosis method for wind turbine generator systems according to claim 2, characterized in that, Through the wind turbine feature learning network, the cross-component sampling records within each cycle segment are processed for linkage pattern learning to generate a set of associated feature groups reflecting abnormal linkages between components, including: According to the component linkage diagram defined by the mechanical structure of the wind turbine, the gearbox node is connected to the generator node, the gearbox node is connected to the blade node, and the generator node is connected to the converter node. The graph neural network of the wind turbine feature learning network initializes a feature vector for each component node. The feature vector is composed of the slope feature, variance feature and attenuation feature of the corresponding component in the period segment. In each layer of the graph neural network, a neighborhood information aggregation operation is performed on each node: the current layer feature vectors of all neighboring nodes are collected, feature transformation is performed through a fully connected layer, and then added to the feature vector of the current node to obtain the updated node feature vector; Repeat the neighborhood information aggregation operation until a preset number of layers is reached, so that the feature vector of each node contains the linkage features of its multi-hop neighborhood nodes; Extract the linkage features between the gearbox node and the generator node, and the linkage features between the gearbox node and the blade node. Calculate the cosine similarity between the linkage features and the historical normal linkage features to generate a set of associated features that reflect the deviation of the synchronous change pattern between components.

5. The fault diagnosis method for wind turbine systems according to claim 1, characterized in that, The method of calling the pre-trained component fault diagnosis model to jointly discriminate the independent feature groups and the associated feature groups generates intermediate diagnostic results containing component anomaly confidence and anomaly linkage strength, including: The independent feature groups are processed through the first fully connected layer of the component fault diagnosis model to output the preliminary anomaly confidence of each component. The associated feature group is processed by the second fully connected layer of the component fault diagnosis model to output the preliminary abnormal linkage strength of each pair of components. The component association knowledge base associated with the component fault diagnosis model is invoked. The component association knowledge base stores the fault impact weights between each component of the wind turbine. The preliminary anomaly confidence and preliminary anomaly linkage strength are weighted and corrected according to the fault impact weights to generate calibrated component anomaly confidence and anomaly linkage strength as intermediate diagnostic results.

6. The fault diagnosis method for wind turbine systems according to claim 1, characterized in that, Based on the intermediate diagnostic results, the system identifies the primary faulty components and accompanying faulty components in the wind turbine generator system, determines the propagation path of the fault between components and the causal relationship of the abnormalities in each component, including: Extract the target component with the highest confidence level of component abnormality from the intermediate diagnostic results, and determine it as the main fault component with the earliest fault start time by combining the time point when the deviation of the fluctuation pattern of the target component in the independent feature group first exceeds the historical normal range. Based on the mechanical transmission structure of the wind turbine, a physical transmission chain is constructed starting from the main faulty component; In the physical transmission chain, components with abnormal linkage strength higher than a preset threshold that are directly connected to the main fault component are selected in sequence according to the direct connection relationship between the components and designated as first-level accompanying fault components. Downstream of the physical connection of the primary accompanying fault component, components with abnormal linkage strength with the primary accompanying fault component exceeding a preset threshold are selected as secondary accompanying fault components. The main fault component, the first-level accompanying fault component, and the second-level accompanying fault component are connected in series in mechanical transmission order to form a fault propagation path that reflects the physical transmission relationship. By comparing the abnormal start time of the main fault component and each of the accompanying fault components, if the abnormal start time of the accompanying fault component is later than that of the main fault component and the time difference is within the standard response delay range between components, it is determined that there is a causal relationship between the main fault component and the accompanying fault component; if the time difference exceeds the standard response delay range, it is determined to be an independent anomaly.

7. The fault diagnosis method for a wind turbine system according to claim 6, characterized in that, By comparing the abnormal start time of the primary fault component with that of each accompanying fault component, if the abnormal start time of the accompanying fault component is later than that of the primary fault component and the time difference meets the standard response delay range between components, then it is determined that there is a causal relationship between the primary fault component and the accompanying fault components, including: The specific moment when the fluctuation pattern deviation of the main fault component in the independent feature group first exceeds the historical normal range is obtained as the start time of the main fault. The exact moment when the deviation of the accompanying fault component in the independent feature group first exceeds the historical normal range is obtained as the start time of the accompanying fault. Calculate the time interval between the start time of the primary fault and the start time of the accompanying fault; Query the response characteristic database of wind turbine components to obtain the standard response delay range between the main fault component and the accompanying fault component; If the time interval is within the standard response delay range, it is determined that the abnormality of the accompanying faulty component is caused by the failure propagation of the main faulty component; If the time interval is not within the standard response delay range, it is determined that the anomaly of the accompanying faulty component is not causally related to the main faulty component and is an independently occurring anomaly. Based on the above determination, a description of the causal relationships between the components in the fault propagation path is generated.

8. The fault diagnosis method for wind turbine systems according to claim 6, characterized in that, Based on the main fault component, accompanying fault components, and fault propagation path, a fault warning command containing component location information is generated, and the fault warning command is sent to the wind turbine operation and maintenance system to trigger targeted maintenance operations, including: The installation location information of the main failure component, the first-level accompanying failure component, and the second-level accompanying failure component is retrieved from the physical structure database of the wind turbine. The installation location information includes the tower height and the coordinates inside the nacelle. Based on the fault propagation path, the primary fault component is marked as the priority maintenance location, the first-level accompanying fault component is marked as the second priority maintenance location, and the second-level accompanying fault component is marked as the associated maintenance location. Based on the type of the main fault component and the deviation of its independent feature group, a corresponding maintenance strategy is matched from the fault case library; The component installation location information, maintenance priority, and maintenance strategy are integrated into a fault early warning command. The fault warning instruction is sent to the operation and maintenance system so that the operation and maintenance system can dispatch maintenance personnel and allocate maintenance tools according to the fault warning instruction.

9. The fault diagnosis method for wind turbine systems according to claim 1, characterized in that, The training method for the component fault diagnosis model includes: The training sample data set and the corresponding training sample label set of the wind turbine system are obtained. The training sample data set contains multiple sets of training sample data. Each set of training sample data contains training independent feature sets reflecting the abnormal tendency of a single component and training correlation feature sets reflecting the abnormal linkage between components. The training sample label set contains the true confidence of component abnormality and the true strength of abnormal linkage corresponding to each set of training sample data. The training independent feature set and the training associated feature set are input into the component fault diagnosis model to be trained. The training independent feature set is processed by the first fully connected layer of the component fault diagnosis model to output the predicted anomaly confidence of each component. The training associated feature set is processed by the second fully connected layer of the component fault diagnosis model to output the predicted anomaly linkage strength of each pair of components. The component association knowledge base associated with the component fault diagnosis model is called. The component association knowledge base stores the fault impact weights between each component of the wind turbine. The predicted anomaly confidence and the predicted anomaly linkage strength are weighted and corrected according to the fault impact weights to generate calibrated component anomaly confidence and anomaly linkage strength as model prediction results. Calculate the first difference between the calibrated component anomaly confidence in the model prediction result and the corresponding component anomaly true confidence in the training sample label set; calculate the second difference between the calibrated abnormal linkage strength and the corresponding abnormal linkage true strength in the training sample label set; and combine the first difference and the second difference after standardization to generate the training loss value. The backpropagation algorithm is used to optimize the network weight parameters of the first fully connected layer and the second fully connected layer of the component fault diagnosis model, as well as the fault impact weights in the component-related knowledge base, based on the training loss value. The optimized component fault diagnosis model is input into the verification sample data set for verification testing. The verification sample data set includes multiple sets of training sample data that did not participate in the training and the corresponding verification sample label set. The verification sample label set includes the component anomaly verification confidence and anomaly linkage verification strength corresponding to each set of verification training sample data. The training sample data in the verification sample dataset is processed by the component fault diagnosis model to generate verification prediction results; Calculate the first deviation between the calibrated component anomaly confidence in the verification prediction result and the component anomaly verification confidence in the verification sample label set; calculate the second deviation between the calibrated abnormal linkage strength and the abnormal linkage verification strength in the verification sample label set; and combine the first deviation and the second deviation after standardization to generate the verification loss value. If the verification loss value is less than the preset loss threshold, the component fault diagnosis model training is completed. If the verification loss value is not less than the preset loss threshold, the calculation weight of the training loss value is adjusted and the model optimization steps are repeated until the verification conditions are met.

10. A fault diagnosis system for wind turbine generator systems, characterized in that, The system includes: The data acquisition module is used to acquire system acquisition data of the wind turbine system. The system acquisition data includes synchronous sampling records of gearbox vibration sensor, generator temperature sensor and blade speed sensor in a continuous period of time. Each sampling record contains the original measurement value of the corresponding component at a single moment. The data preprocessing module is used to extract fault-sensitive features from the data collected by the system through the wind turbine feature learning network, and generate independent feature groups reflecting the abnormal tendency of a single component and related feature groups reflecting the abnormal linkage between components. The independent feature groups include the time-series fluctuation pattern of component measurement values, and the related feature groups include the synchronous abnormal pattern of cross-component measurement values. The diagnostic intermediate result generation module is used to call the pre-trained component fault diagnosis model to jointly discriminate the independent feature group and the associated feature group, and generate diagnostic intermediate results including component anomaly confidence and anomaly linkage strength. The diagnostic intermediate results are dynamically calibrated through the component association knowledge base associated with the component fault diagnosis model. The fault diagnosis module identifies the main faulty components and accompanying faulty components in the wind turbine system based on the intermediate diagnostic results, and determines the propagation path of the fault between components and the causal relationship of the abnormality of each component. The fault early warning module generates a fault early warning command containing component location information based on the main fault component, the accompanying fault components and the fault propagation path, and sends the fault early warning command to the wind turbine operation and maintenance system to trigger targeted maintenance operations. The component location information is associated with the physical installation location of the component inside the wind turbine tower.

Images