Wind turbine generator converter fault prediction method and system based on data driving

Abstract

The invention discloses a wind turbine generator converter fault prediction method and system based on data driving, and the method comprises the steps: firstly, extracting deep features from high-frequency signal data and low-frequency working condition data through independent branches, and capturing the fine state and macroscopic operation background of equipment; the extracted working condition features are used as dynamic modulation reference, and features extracted from signals are adaptively adjusted and reconstructed through a working condition attention modulation layer. According to the modulation process, the model can learn to distinguish which signal fluctuation is normal and which signal fluctuation is abnormal under a specific working condition, so that feature coupling between working condition change and equipment degradation is effectively decoupled. And finally, the model performs health index prediction based on the feature vector purified by the working condition, so that the false alarm caused by the sudden change of the working condition is fundamentally inhibited, and the prediction accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of fault handling technology, specifically to a data-driven method and system for predicting faults in wind turbine converters. Background Technology

[0002] Wind power, as a clean and renewable energy source, plays a crucial role in the global energy structure transformation. As the core equipment of wind power generation, the reliability of wind turbines directly affects the power generation efficiency and economic benefits of the entire wind farm. Among the many subsystems of a wind turbine, the converter is a key power electronic device connecting the generator and the grid, responsible for the conversion and control of electrical energy; however, it is also one of the most vulnerable links with a high failure rate. Unexpected converter failures will lead to turbine shutdown, resulting not only in significant power generation losses but also incurring high maintenance costs and safety risks. Therefore, to improve the availability of wind turbines and reduce operation and maintenance costs, developing predictive maintenance technologies that can provide early warnings of potential converter failures has become an urgent need for the wind power industry.

[0003] Currently, some data-driven attempts have emerged in the industry to predict wind turbine converter faults. These methods utilize operational data acquired from Supervisory Control and Data Acquisition (SCADA) systems or Condition Monitoring Systems (CMS) to identify abnormal patterns through machine learning or deep learning models. However, existing technologies generally face a core challenge in practical applications: the difficulty in effectively suppressing false alarms caused by changes in wind turbine operating conditions. The operating conditions of wind turbines (such as wind speed, power output, and grid commands) are constantly changing dynamically. These changes cause drastic, but within-normal-range, fluctuations in electrical signals (such as current and voltage) and thermal signals (such as temperature) within the converter. These normal operating condition fluctuations sometimes bear a striking resemblance in shape and amplitude to those of certain early faults (such as power device performance degradation and capacitor parameter drift). Most existing prediction models are static; they attempt to learn a fixed boundary between health and abnormality, but they cannot effectively distinguish between normal fluctuations caused by changes in operating conditions and abnormal signals caused by equipment degradation. This causes the model to misinterpret normal signal fluctuations as precursors to faults when wind speeds change abruptly or the system is operating under high load, resulting in a large number of false alarms and severely reducing the reliability and practical value of the prediction system.

[0004] Therefore, an optimized fault prediction scheme for wind turbine converters is desired. Summary of the Invention

[0005] The embodiments of this application aim to at least solve one of the technical problems existing in the prior art, and provide a data-driven method and system for predicting the faults of wind turbine converters.

[0006] On one hand, embodiments of this application provide a data-driven method for predicting faults in wind turbine converters, including: Acquire high-frequency CMS data and low-frequency SCADA data; Feature window processing is performed on high-frequency CMS data and low-frequency SCADA data to obtain signal window data and operating condition vectors; Dual-stream feature extraction is performed on the signal window data and the operating condition vector to obtain the signal time sequence feature map and the operating condition feature vector; The signal timing feature map and the operating condition feature vector are input into the operating condition attention modulation layer to obtain the operating condition modulated signal timing feature vector. Online health index prediction is performed based on the time-series feature vector of the modulated signal under operating conditions to obtain the predicted health index. Based on the predicted health index, determine whether a fault warning signal has been triggered.

[0007] Optionally, dual-stream feature extraction is performed on the signal window data and the operating condition vector to obtain the signal time-series feature map and the operating condition feature vector, including: Input the signal window data into the signal encoder to obtain the signal timing characteristic map; The operating condition vector is input into the operating condition processor to obtain the operating condition feature vector.

[0008] Optionally, the signal encoder uses a two-dimensional convolutional neural network model as a feature encoder, and the operating condition processor uses a multilayer perceptron as a feature encoder.

[0009] Optionally, the signal timing feature map and the operating condition feature vector are input into the operating condition attention modulation layer to obtain the operating condition modulated signal timing feature vector, including: The signal timing feature map is decomposed into nodes along the channel direction to obtain a set of signal timing feature vectors; Using the operating condition feature vector as a modulation reference, the dynamic context correlation matrix of the set of signal timing feature vectors is calculated to obtain the signal timing time-varying context correlation matrix; The set of signal time-series feature vectors and the signal time-series time-varying context association matrix are input into the graph convolutional neural network model to obtain the working condition prior driving signal time-series node structure relationship matrix; Using the operating condition feature vector as the query vector, cross-modal dynamic fusion is performed on the operating condition prior driving signal time sequence node structure relationship matrix to obtain the operating condition modulation signal time sequence feature vector.

[0010] Optionally, online health index prediction based on the time-series feature vector of the operating condition modulation signal to obtain the predicted health index includes: inputting the time-series feature vector of the operating condition modulation signal into the trained regression head to obtain the predicted health index.

[0011] Optionally, based on the predicted health index, it is determined whether to trigger a fault warning signal, including: in response to the predicted health index being lower than a preset health index threshold, confirming that a fault warning signal has been triggered.

[0012] On the other hand, this application provides a data-driven wind turbine converter fault prediction system, comprising: The data acquisition module is used to acquire high-frequency CMS data and low-frequency SCADA data; The feature window processing module is used to perform feature window processing on high-frequency CMS data and low-frequency SCADA data to obtain signal window data and operating condition vectors. The feature extraction module is used to perform dual-stream feature extraction on signal window data and operating condition vectors to obtain signal time-series feature maps and operating condition feature vectors. The attention modulation module is used to input the signal timing feature map and the operating condition feature vector into the operating condition attention modulation layer to obtain the operating condition modulated signal timing feature vector. The index prediction module is used to perform online health index prediction based on the timing feature vector of the operating condition modulation signal to obtain the predicted health index. The early warning determination module is used to determine whether a fault early warning signal should be triggered based on the predicted health index.

[0013] Optionally, the feature extraction module includes: The signal encoding unit is used to input the signal window data into the signal encoder to obtain the signal timing feature map; The operating condition processing unit is used to input the operating condition vector into the operating condition processor to obtain the operating condition feature vector.

[0014] Optionally, the attention modulation module includes: The node decomposition unit is used to perform node decomposition on the signal timing feature map along the channel direction to obtain a set of signal timing feature vectors; The matrix calculation unit is used to calculate the dynamic context correlation matrix of the set of signal timing feature vectors using the operating condition feature vector as a modulation reference to obtain the signal timing time-varying context correlation matrix. The model input unit is used to input the set of signal time-series feature vectors and the signal time-series time-varying context association matrix into the graph convolutional neural network model to obtain the working condition prior driving signal time-series node structure relationship matrix; The dynamic fusion unit is used to perform cross-modal dynamic fusion of the time sequence node structure relationship matrix of the prior driving signal of the operating conditions, using the feature vector of the operating conditions as a query vector, to obtain the time sequence feature vector of the modulated signal of the operating conditions.

[0015] Compared with existing technologies, this application provides a data-driven method and system for predicting wind turbine converter faults. It extracts deep features from high-frequency signal data and low-frequency operating condition data through independent branches, capturing the fine-grained state of the equipment and its macroscopic operating background. The key lies in using the extracted operating condition features as a dynamic modulation reference, enabling a condition-focused modulation layer to adaptively adjust and reconstruct the features extracted from the signals. This modulation process allows the model to learn to distinguish which signal fluctuations are normal and which are abnormal under specific operating conditions, effectively decoupling the feature coupling between operating condition changes and equipment degradation. Finally, the model predicts a health index based on this feature vector purified by the operating conditions, fundamentally suppressing false alarms caused by drastic changes in operating conditions and improving prediction accuracy. Attached Figure Description

[0016] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 This is a flowchart of a data-driven wind turbine converter fault prediction method according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow in the data-driven wind turbine converter fault prediction method according to an embodiment of this application; Figure 3 This is a flowchart illustrating the data-driven wind turbine converter fault prediction method according to an embodiment of this application, which extracts dual-stream features from signal window data and operating condition vectors to obtain signal timing feature maps and operating condition feature vectors. Figure 4 This is a flowchart illustrating the input of a signal timing feature map and a condition feature vector into a condition attention modulation layer to obtain a condition modulated signal timing feature vector, according to an embodiment of the present application, of a data-driven wind turbine converter fault prediction method. Figure 5 This is a block diagram of a data-driven wind turbine converter fault prediction and control system according to an embodiment of this application. Detailed Implementation

[0018] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0019] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0020] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0021] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0022] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0023] To address the technical problems mentioned in the background, this application proposes a data-driven fault prediction scheme for wind turbine converters. This scheme solves the problem in existing technologies where normal signal fluctuations are misjudged as fault precursors due to drastic changes in operating conditions such as wind speed and power, resulting in numerous false alarms. Specifically, this scheme first starts from the data source. It simultaneously acquires high-frequency CMS data reflecting the fine-grained state of the equipment and low-frequency SCADA data characterizing the macroscopic operating background, processing them into signal window data and operating condition vectors. Then, using a dual-stream architecture, a two-dimensional convolutional neural network is used to perform deep feature extraction on the signal window data, generating a signal time-series feature map rich in spatiotemporal information. Simultaneously, a multilayer perceptron is used to encode the operating condition vector into a concise operating condition feature vector. Further, through an operating condition attention modulation layer: first, the signal time-series feature map is decomposed into a set of feature nodes. Then, using the operating condition feature vector as a reference, the correlation matrix between these nodes is dynamically calculated, constructing a graph structure that reflects the inherent relationships between signal features under the current operating conditions. Then, a graph convolutional neural network is used to propagate and aggregate information on the dynamic graph, resulting in a node relationship matrix guided by prior knowledge of the operating conditions. Finally, the operating condition feature vector is used as the query again, and the relationship matrix is ​​dynamically fused through a cross-modal attention mechanism to ultimately generate a timing feature vector of the operating condition modulation signal that has fully absorbed the operating condition information and effectively suppressed normal fluctuations.

[0024] The technical solution of this application proposes a data-driven method for predicting the faults of wind turbine converters. Figure 1 This is a flowchart of a data-driven wind turbine converter fault prediction method according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in a data-driven wind turbine converter fault prediction method according to an embodiment of this application. Figure 1 and Figure 2 As shown, the data-driven wind turbine converter fault prediction method according to an embodiment of this application includes the following steps: S100, acquiring high-frequency CMS data and low-frequency SCADA data. S200, performing feature window processing on the high-frequency CMS data and low-frequency SCADA data to obtain signal window data and operating condition vector. S300, performing dual-stream feature extraction on the signal window data and operating condition vector to obtain a signal timing feature map and an operating condition feature vector. S400, inputting the signal timing feature map and operating condition feature vector into the operating condition attention modulation layer to obtain an operating condition modulation signal timing feature vector. S500, performing online health index prediction based on the operating condition modulation signal timing feature vector to obtain a predicted health index. S600, determining whether a fault warning signal should be triggered based on the predicted health index.

[0025] Specifically, in step S100, high-frequency CMS data and low-frequency SCADA data are acquired. It should be understood that early fault characteristics of wind turbine converters often manifest as weak disturbances in high-frequency electrical signals, which are difficult to effectively capture relying solely on low-frequency SCADA data. Furthermore, analyzing only high-frequency CMS data cannot distinguish between equipment degradation and normal signal fluctuations caused by drastic changes in external operating conditions, leading to insufficient reliability of diagnostic results. Therefore, in the technical solution of this application, high-frequency CMS data and low-frequency SCADA data are acquired to construct a complete dataset that can simultaneously characterize the fine internal electrical state of the converter and the external macroscopic operating conditions. This provides the necessary data foundation for subsequent model learning and decoupling the impact of operating condition changes on signal characteristics, thereby creating the preconditions for achieving highly accurate fault prediction.

[0026] More specifically, in a concrete example of this application, firstly, high-frequency data that can accurately reflect the electrical characteristics of the converter are acquired from the wind turbine's condition monitoring system, such as the converter's three-phase current and DC bus voltage signals acquired at a sampling frequency of 10 kHz. Simultaneously, low-frequency data that characterizes the macroscopic operating conditions of the unit is acquired from the wind turbine's monitoring and data acquisition system, such as active power and wind speed data with a 10-minute period. Both types of data are transmitted to a data server via an industrial bus and stored uniformly. Next, to align data of different frequencies, the low-frequency SCADA data is time-synchronized based on the timestamp sequence of the high-frequency CMS data. This processing uses a forward padding method, assigning the value of each SCADA sampling point to all high-frequency data timestamps up to the next SCADA sampling point. Finally, a unified time-series data stream is formed, where each row of data contains a high-frequency signal measurement value at a given moment and its corresponding macroscopic operating condition parameter value, providing formatted input for subsequent feature windowing processing steps.

[0027] Specifically, in step S200, feature windowing is performed on the high-frequency CMS data and low-frequency SCADA data to obtain signal window data and operating condition vectors. It should be understood that continuous, unstructured time-series data streams cannot be directly used as input to deep learning models, and analyzing instantaneous data points makes it difficult to capture dynamic patterns that reflect the evolution of equipment states. Therefore, in the technical solution of this application, feature windowing is further performed on the high-frequency CMS data and low-frequency SCADA data to obtain signal window data and operating condition vectors. This divides the infinitely long continuous data stream into a series of discrete samples with fixed dimensions and lengths that can be processed by the model, and separates the signals representing details from the operating conditions representing the background at the data organization level. This generates structured input pairs, providing a standardized data format for the subsequent parallel and targeted extraction of signal timing features and operating condition features using a dual-stream model.

[0028] More specifically, in a concrete example of this application, firstly, a fixed-length sliding window with 2048 sampling points and a sliding step size of 1024 sampling points are set. This sliding window moves along a time-aligned uniform data stream. At each window position, 2048 consecutive data points of all high-frequency CMS data channels within the window—namely, the converter three-phase current and DC bus voltage—are extracted. These data collectively form a two-dimensional signal window data matrix. Simultaneously, for the corresponding low-frequency SCADA data within the same time window—namely, active power and wind speed—their respective arithmetic mean is calculated. These two averages together form a one-dimensional operating condition vector. By continuously moving the sliding window according to the set step size and repeating the above extraction and calculation process, the entire continuous data stream can be converted into a structured dataset composed of a large number of (signal window data, operating condition vector) sample pairs. Finally, all generated signal window data and operating condition vectors are normalized, linearly scaling their values ​​to the [0,1] interval, to serve as input for subsequent dual-stream feature extraction steps.

[0029] Specifically, in step S300, dual-stream feature extraction is performed on the signal window data and the operating condition vector to obtain a signal time-series feature map and an operating condition feature vector. It should be understood that signal window data, as a two-dimensional matrix containing multi-channel time-series information, differs fundamentally from the one-dimensional operating condition vector, which describes a macroscopic state, in terms of data structure, information density, and physical meaning. If a single network structure is used for processing, it is impossible to perform optimal feature extraction for each data characteristic, easily leading to information loss or feature confusion. Therefore, in the technical solution of this application, dual-stream feature extraction is further performed on the signal window data and the operating condition vector to obtain a signal time-series feature map and an operating condition feature vector. This utilizes a dedicated network architecture best suited to each data modality to extract deep abstract features from two heterogeneous data sources in parallel. In this way, a signal timing feature map can be generated that can capture the multi-scale temporal sequence and local correlation between channels within the signal through convolution and pooling operations, and can also learn the nonlinear relationship between operating parameters through fully connected layers. This provides high-quality, preliminarily decoupled feature input for the subsequent operating condition attention modulation step.

[0030] Figure 3 This is a flowchart illustrating the process of extracting dual-stream features from signal window data and operating condition vectors to obtain signal timing feature maps and operating condition feature vectors, according to an embodiment of the data-driven wind turbine converter fault prediction method of this application. Figure 3 As shown, step S300 includes: S310, inputting signal window data into a signal encoder to obtain a signal timing feature map; and S320, inputting a working condition vector into a working condition processor to obtain the working condition feature vector.

[0031] Specifically, in step S310, the signal window data is input to the signal encoder to obtain a signal timing feature map. The operating condition vector is input to the operating condition processor to obtain the operating condition feature vector. It is worth noting that the signal encoder uses a two-dimensional convolutional neural network model as the feature encoder. It should be understood that the original signal window data, as a two-dimensional matrix, has a high dimension and contains a large amount of original waveform details. Directly using it for subsequent processing not only incurs a huge computational burden, but also submerges key feature information related to the equipment's health status in redundant data and noise. Therefore, in the technical solution of this application, the signal window data is further input to the signal encoder to obtain a signal timing feature map. This utilizes the powerful local perception and parameter sharing capabilities of the two-dimensional convolutional neural network to automatically learn and extract hierarchical, low-order to high-order abstract features from the signal window data. In this way, the original, massive signal matrix can be compressed and transformed into a signal timing feature map with higher information density and better characterizing the signal's internal structure and dynamic patterns, providing a purified and dimensionality-reduced high-quality input for subsequent feature modulation.

[0032] More specifically, in a concrete example of this application, a 4x2048 signal window data matrix is ​​input into a two-dimensional convolutional neural network model acting as a signal encoder. This network model consists of multiple stacked convolutional and pooling layers. First, the input matrix passes through a first convolutional layer, which uses, for example, 32 3x3 kernels to convolve the input matrix, capturing local fundamental features of the signal in the temporal and channel directions, such as waveform edges and inflection points, and generating 32 initial feature maps. These feature maps are then passed through a ReLU activation function and downsampled by a max-pooling layer to preserve features and reduce data dimensionality. Next, the output is fed into subsequent convolutional layers, such as a second convolutional layer using 64 3x3 kernels, which learns more complex and abstract patterns based on the initial features extracted by the previous layer. This convolution, activation, and pooling process is repeated several times, with each layer extracting higher-order features than the previous one. Finally, the tensor output after the last layer of processing is the signal temporal feature map, which is a compact representation with compressed dimensions but increased number of channels (e.g., 64 channels), where each channel represents a specific spatiotemporal pattern learned from the original signal.

[0033] Specifically, in step S320, the operating condition vector is input into the operating condition processor to obtain the operating condition feature vector. It is worth noting that the operating condition processor uses a multilayer perceptron as the feature encoder. It should be understood that the original operating condition vector is merely a simple combination of several macroscopic physical quantities; it does not explicitly contain the complex nonlinear relationships between these physical quantities, nor can it directly serve as the high-dimensional reference or query information required by the subsequent attention modulation mechanism. Therefore, in the technical solution of this application, the operating condition vector is further input into the operating condition processor to obtain the operating condition feature vector, thereby utilizing the powerful nonlinear fitting capability of the multilayer perceptron to map the low-dimensional original operating condition parameters into a higher-dimensional, more expressive feature space. This generates a compact operating condition feature vector that contains deep interaction information between operating condition parameters. This vector not only matches the signal features in dimension but is also semantically sufficient to guide the subsequent complex modulation process of the signal features.

[0034] More specifically, in a concrete example of this application, a 2-dimensional load condition vector, whose elements are the normalized mean active power and mean wind speed, is input into a multilayer perceptron model acting as a load condition processor. This model consists of an input layer, two fully connected hidden layers, and an output layer. First, the 2D vector enters a first hidden layer containing, for example, 16 neurons, and undergoes a nonlinear transformation using a ReLU activation function to initially learn the basic relationships between load condition parameters. Subsequently, the output of the first hidden layer is fed into a second hidden layer containing, for example, 32 neurons, and again processed by a ReLU activation function to combine lower-level features and learn more complex nonlinear mappings. Finally, the output of the second hidden layer enters a linear output layer containing 64 neurons. This layer does not use an activation function but directly performs a linear transformation, ultimately outputting a 64-dimensional load condition feature vector. The dimension of this output vector matches the number of channels in the signal temporal feature map output by the signal encoder, preparing for subsequent modulation operations.

[0035] Specifically, in step S400, the signal timing feature map and the operating condition feature vector are input into the operating condition attention modulation layer to obtain the operating condition modulated signal timing feature vector. It should be understood that although the signal timing feature map and the operating condition feature vector obtained by the dual-stream encoder each contain rich abstract information, they are still separate. If simple splicing or addition is used for fusion, only a shallow feature-level superposition is achieved, which cannot reveal the deep modulation relationship between operating condition changes and the internal structure of the signal features; that is, it cannot answer which combinations of signal feature modes are normal and which are abnormal under a specific operating condition. Therefore, in the technical solution of this application, the signal timing feature map and the operating condition feature vector are further input into the operating condition attention modulation layer to obtain the operating condition modulated signal timing feature vector. This elevates the focus of cross-modal interaction from traditional feature-level fusion to the level of structure-level modulation, allowing the feature vector representing the macroscopic operating condition to serve as a modal prior to dynamically shape and infer the internal topological structure of the feature map representing the microscopic signal. In this way, the model can go beyond pixel-level feature comparison and perform information aggregation and reasoning on a structured path guided by the semantics of operating conditions. This generates a feature vector that has internalized the impact of operating conditions from interference factors into an analytical perspective and can better reflect the true health status of the equipment, thus fundamentally solving the problem of false alarms under varying operating conditions.

[0036] Figure 4 This is a flowchart illustrating the input of a signal timing feature map and a condition feature vector into a condition attention modulation layer to obtain a condition-modulated signal timing feature vector, according to an embodiment of the data-driven wind turbine converter fault prediction method of this application. Figure 4As shown, step S400 includes: S410, performing node decomposition on the signal timing feature map along the channel direction to obtain a set of signal timing feature vectors; S420, using the operating condition feature vector as a modulation reference, calculating the dynamic context association matrix of the set of signal timing feature vectors to obtain a signal timing time-varying context association matrix; S430, inputting the set of signal timing feature vectors and the signal timing time-varying context association matrix into a graph convolutional neural network model to obtain an operating condition prior-driven signal timing node structure relationship matrix; S440, using the operating condition feature vector as a query vector, performing cross-modal dynamic fusion on the operating condition prior-driven signal timing node structure relationship matrix to obtain the operating condition modulated signal timing feature vector.

[0037] Specifically, in step S410, the signal timing feature map is decomposed into nodes along the channel direction to obtain a set of signal timing feature vectors. This is expressed by the following formula: ; in, This is a time-series characteristic map of the signal. For feature decomposition, It is a set of signal time-series feature vectors. It is the first in the set of signal time-series eigenvectors Each signal has a time-series feature vector.

[0038] It should be understood that signal timing feature maps are essentially a continuous grid-based representation. While they contain rich signal patterns, the relationships between these patterns are implicit and cannot be directly used as units for graph-based relational reasoning. To explore the dynamic changes in the correlations between these signal patterns under different operating conditions, the continuous feature map must first be reshaped into a more flexible graph-based relational paradigm composed of discrete nodes. Therefore, in the technical solution of this application, the signal timing feature map is decomposed along the channel direction to obtain a set of signal timing feature vectors. This abstracts the specific timing pattern represented by each channel in the feature map into independent, discrete graph nodes. This lays the structural foundation for the subsequent introduction of relational reasoning models such as graph convolutional neural networks, enabling the model to transcend the pixel-level adjacency relationships within the feature map and capture and reason about the dynamic correlations between these higher-dimensional abstract signal feature patterns, thereby opening up the possibility of structure-level modulation.

[0039] Specifically, in step S420, using the operating condition feature vector as a modulation reference, the dynamic context correlation matrix of the set of signal timing feature vectors is calculated to obtain the signal timing time-varying context correlation matrix. This is expressed by the following formula:

[0040] in, It is the first in the set of signal time-series eigenvectors Each signal time-series feature vector It is the feature vector of operating conditions. , and For learnable linear transformation matrices, This is a vector concatenation operation. For attention parameter vectors, For attention parameter vectors Transpose. for Activation function for Normalization function, For the signal timing dynamic context correlation matrix The characteristic value of the location.

[0041] It is understandable that a static, predefined graph structure cannot reflect the dynamic changes in the correlation between signal characteristic modes under different operating conditions. For example, a strong correlation between a certain current harmonic and voltage ripple is normal under high-power conditions, but it is a fault symptom under low-power conditions. A fixed graph structure cannot capture this context-dependent relationship. Therefore, in the technical solution of this application, the operating condition feature vector is further used as a modulation reference to calculate the dynamic context correlation matrix of the set of signal timing feature vectors to obtain the signal timing time-varying context correlation matrix. This completely overturns the idea of ​​constructing static graphs, and uses the feature vector representing the macroscopic operating condition as a high-level semantic prior or modulation reference, directly injecting it into the definition of the structural relationship between signal feature nodes. This dynamically generates a unique, non-local context correlation matrix for each input (signal window, operating condition) sample pair. In this way, the graph's topology itself becomes part of the contextual information, generating a time-varying contextual correlation matrix that already contains prior information about the operating conditions. This matrix provides the subsequent graph convolutional network with an optimized information propagation path calibrated for the current operating conditions, enabling the model to reason on the correct structural relationships, rather than on fixed, misleading structures.

[0042] Specifically, in step S430, the set of signal time-series feature vectors and the signal time-series time-varying context correlation matrix are input into the graph convolutional neural network model to obtain the operating condition prior driving signal time-series node structure relationship matrix. This is expressed by the following formula:

[0043] in, For the signal timing dynamic context correlation matrix, For graph convolutional neural network models, The matrix represents the temporal node structure relationship of the a priori driving signal under operating conditions.

[0044] It should be understood that merely generating a signal time-varying context association matrix guided by operating conditions only defines the association strength between signal feature nodes under the current operating conditions, while the feature representations of the nodes themselves have not yet been updated and interacted with according to this dynamic structure; the information remains local and isolated. Therefore, in the technical solution of this application, the set of signal time-series feature vectors and the signal time-varying context association matrix are further input into a graph convolutional neural network model to obtain a priori-driven signal time-series node structure relationship matrix under operating conditions. This leverages the powerful information propagation and node feature update capabilities of the graph convolutional neural network to perform iterative message passing and feature aggregation on the graph topology that has just been calibrated by operating conditions. In this way, the representation of each signal feature node incorporates its neighborhood structure information under the current operating conditions, thereby outputting a deeply reasoned, structured, and holistic signal representation, namely, a priori-driven signal time-series node structure relationship matrix under operating conditions, which internally encodes the topological relationships most relevant to the current operating conditions.

[0045] Specifically, in step S440, using the operating condition feature vector as the query vector, cross-modal dynamic fusion is performed on the operating condition prior driving signal timing node structure relation matrix to obtain the operating condition modulation signal timing feature vector. This is expressed by the following formula:

[0046]

[0047]

[0048]

[0049] ； in, , and These are the learnable query mapping matrix, the learnable key mapping matrix, and the learnable value mapping matrix, respectively. , and These are the query vector, key matrix, and value matrix, respectively. For the key matrix Transpose. Scaling factor The time-series vector of the prior modulation signal under operating conditions. It is a feedforward neural network. for function, This is the timing feature vector of the modulated signal under operating conditions.

[0050] It should be understood that the temporal node structure relationship matrix of the operating condition prior driving signal obtained after graph convolutional network processing, although already a holistic representation after structural reorganization of the operating condition prior, is still a high-dimensional feature set. It has not yet undergone a final focusing and information contraction based on the final direction of the operating condition to form a single feature vector that can be directly used for health status prediction. Therefore, in the technical solution of this application, the operating condition feature vector is further used as a query vector to perform cross-modal dynamic fusion on the temporal node structure relationship matrix of the operating condition prior driving signal to obtain the temporal feature vector of the operating condition modulation signal. This is used to perform a final alignment and secondary focusing, that is, again using the feature vector representing the macroscopic operating condition as a query to perform a precise and global information reading and weighted fusion on the highly structured signal feature matrix that has undergone deep context reasoning. In this way, the most critical information can be selectively extracted and fused from the signal representation that has undergone sufficient structured reasoning, based on the final direction of the current operating condition. This information is then condensed into a single, highly condensed operating condition modulation signal timing feature vector. This vector is the inevitable result of the deep and structured interaction between the signal and the operating condition, ensuring that the final health index prediction is based on the purest equipment status representation that has gone through a complete reasoning chain.

[0051] Specifically, in step S500, an online health index prediction is performed based on the time-series feature vector of the operating condition modulation signal to obtain a predicted health index. It should be understood that although the time-series feature vector of the operating condition modulation signal generated in the preceding steps is a pure representation after deep inference and operating condition decoupling, it is essentially still a high-dimensional abstract vector and cannot be directly used as an intuitive quantitative evaluation indicator of the equipment's health status. It needs to be converted into a concrete numerical value that is easy to understand and apply. Therefore, in the technical solution of this application, an online health index prediction is further performed based on the time-series feature vector of the operating condition modulation signal to obtain a predicted health index. This utilizes a simple regression network (i.e., a regression head) to perform a final nonlinear mapping on this highly condensed feature vector, projecting it from a complex feature space to a one-dimensional continuous numerical space representing the equipment's health level. In this way, an intuitive and quantitative predictive health index with a value between [0,1] can be output. This index can not only continuously track the performance degradation trajectory of the converter from health to failure, but also, because its calculation basis is the characteristics of the operating condition modulation, its numerical changes can more realistically reflect the internal health status of the equipment rather than the interference of external operating conditions, providing a reliable basis for setting accurate early warning thresholds.

[0052] More specifically, in this embodiment, the timing feature vector of the operating condition modulation signal is input into the trained regression head to obtain the predicted health index. First, a 64-dimensional timing feature vector of the operating condition modulation signal output by the operating condition attention modulation layer is input to the first fully connected layer of the regression head. This layer contains 32 neurons and uses ReLU as the activation function for a nonlinear transformation. Subsequently, the output of this hidden layer is fed into the second fully connected layer of the regression head, i.e., the output layer. This output layer contains only one neuron and uses the Sigmoid activation function. The Sigmoid function maps any input value to the (0,1) interval, and its output value is the final predicted health index. An index value close to 1 indicates that the converter is in a very healthy state, while when its value gradually decreases and approaches 0, it indicates that the equipment performance is deteriorating and a failure is imminent. The network parameters of the regression head are learned and determined during the end-to-end training process of the entire model by minimizing the error between the predicted health index and the preset true health label.

[0053] Specifically, in step S600, based on the predicted health index, it is determined whether a fault warning signal should be triggered. It should be understood that the output predicted health index is a continuously changing value, which only provides a quantitative assessment of the equipment status and does not directly give a clear instruction on whether maintenance action is needed. This makes it difficult for maintenance personnel to make timely and unified decisions based on this continuous value. Therefore, in the technical solution of this application, it is further determined whether a fault warning signal should be triggered based on the predicted health index, thereby establishing a clear, objective, and automated decision-making mechanism that transforms continuous health status assessments into discrete warning events with clear operational significance. This allows complex prediction results to be transformed into simple and intuitive alarm information, providing the maintenance team with timely and unambiguous action instructions, thereby ensuring that predictive maintenance strategies can be effectively implemented, ultimately achieving the goals of avoiding unplanned downtime and reducing maintenance costs.

[0054] More specifically, in this embodiment, determining whether to trigger a fault warning signal based on a predicted health index includes: confirming the triggering of a fault warning signal in response to the predicted health index falling below a preset health index threshold. This is achieved by setting two independent preset health index thresholds, namely a primary warning threshold and a secondary warning threshold, for example, 0.6 and 0.4 respectively. During online operation, the system continuously compares the latest calculated predicted health index with these two thresholds. When the predicted health index falls below the secondary warning threshold (e.g., 0.4), the system immediately confirms the triggering of a secondary fault warning signal, indicating that the device is in a severely degraded state. The system will automatically generate a high-priority alarm event and push it to the operation and maintenance management platform. When the predicted health index is between the primary and secondary warning thresholds (e.g., within the range of 0.4 to 0.6), the system enters an observation and confirmation state. Only when this state persists for more than a preset duration (e.g., 10 consecutive minutes) does the system formally confirm the triggering of a primary fault warning signal to generate a suggestive maintenance recommendation. This aims to filter out transient index drops caused by momentary disturbances and increase the reliability of the warning. If the predicted health index is consistently higher than the first-level warning threshold, the system determines that the equipment is operating normally and does not trigger any warning signals.

[0055] In summary, the data-driven wind turbine converter fault prediction method according to the embodiments of this application is explained. It extracts deep features from high-frequency signal data and low-frequency operating condition data through independent branches, capturing the fine-grained state of the equipment and the macroscopic operating background. The key lies in using the extracted operating condition features as a dynamic modulation reference, and enabling an operating condition attention modulation layer to adaptively adjust and reconstruct the features extracted from the signal. This modulation process allows the model to learn to distinguish which signal fluctuations are normal and which are abnormal under specific operating conditions, thereby effectively decoupling the feature coupling between operating condition changes and equipment degradation. Finally, the model predicts the health index based on this feature vector purified by the operating conditions, fundamentally suppressing false alarms caused by drastic changes in operating conditions and improving the accuracy of prediction. Furthermore, a data-driven wind turbine converter fault prediction and control system is also provided.

[0056] Figure 5 This is a block diagram of a data-driven wind turbine converter fault prediction and control system according to an embodiment of this application. Figure 5As shown, the data-driven wind turbine converter fault prediction control system 500 according to an embodiment of this application includes: a data acquisition module 510 for acquiring high-frequency CMS data and low-frequency SCADA data; a feature window processing module 520 for performing feature window processing on the high-frequency CMS data and low-frequency SCADA data to obtain signal window data and operating condition vectors; a feature extraction module 530 for performing dual-stream feature extraction on the signal window data and operating condition vectors to obtain signal timing feature maps and operating condition feature vectors; an attention modulation module 540 for inputting the signal timing feature maps and operating condition feature vectors into an operating condition attention modulation layer to obtain an operating condition modulated signal timing feature vector; an index prediction module 550 for performing online health index prediction based on the operating condition modulated signal timing feature vector to obtain a predicted health index; and a warning determination module 560 for determining whether a fault warning signal should be triggered based on the predicted health index.

[0057] As described above, the data-driven wind turbine converter fault prediction and control system 500 according to the embodiments of this application can be implemented in various wireless terminals, such as servers with data-driven wind turbine converter fault prediction and control algorithms. In one possible implementation, the data-driven wind turbine converter fault prediction and control system 500 according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the data-driven wind turbine converter fault prediction and control system 500 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal. Of course, the data-driven wind turbine converter fault prediction and control system 500 can also be one of many hardware modules of the wireless terminal.

[0058] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A data-driven method for predicting faults in wind turbine converters, characterized in that, include: Acquire high-frequency CMS data and low-frequency SCADA data; Feature window processing is performed on high-frequency CMS data and low-frequency SCADA data to obtain signal window data and operating condition vectors; Dual-stream feature extraction is performed on the signal window data and the operating condition vector to obtain the signal time sequence feature map and the operating condition feature vector; The signal timing feature map and the operating condition feature vector are input into the operating condition attention modulation layer to obtain the operating condition modulated signal timing feature vector. Online health index prediction is performed based on the time-series feature vector of the modulated signal under operating conditions to obtain the predicted health index. Based on the predicted health index, determine whether a fault warning signal has been triggered.

2. The data-driven wind turbine converter fault prediction method according to claim 1, characterized in that, Dual-stream feature extraction is performed on the signal window data and operating condition vector to obtain the signal time-series feature map and operating condition feature vector, including: Input the signal window data into the signal encoder to obtain the signal timing characteristic map; The operating condition vector is input into the operating condition processor to obtain the operating condition feature vector.

3. The data-driven wind turbine converter fault prediction method according to claim 2, characterized in that, The signal encoder uses a two-dimensional convolutional neural network model as a feature encoder, and the operating condition processor uses a multilayer perceptron as a feature encoder.

4. The data-driven wind turbine converter fault prediction method according to claim 3, characterized in that, The signal timing feature map and the operating condition feature vector are input into the operating condition attention modulation layer to obtain the operating condition modulated signal timing feature vector, including: The signal timing feature map is decomposed into nodes along the channel direction to obtain a set of signal timing feature vectors; Using the operating condition feature vector as a modulation reference, the dynamic context correlation matrix of the set of signal timing feature vectors is calculated to obtain the signal timing time-varying context correlation matrix; The set of signal time-series feature vectors and the signal time-series time-varying context association matrix are input into the graph convolutional neural network model to obtain the working condition prior driving signal time-series node structure relationship matrix; Using the operating condition feature vector as the query vector, cross-modal dynamic fusion is performed on the operating condition prior driving signal time sequence node structure relationship matrix to obtain the operating condition modulation signal time sequence feature vector.

5. The data-driven wind turbine converter fault prediction method according to claim 1, characterized in that, Online health index prediction based on the time-series feature vector of the operating condition modulation signal to obtain the predicted health index includes: inputting the time-series feature vector of the operating condition modulation signal into a trained regression head to obtain the predicted health index.

6. The data-driven wind turbine converter fault prediction method according to claim 5, characterized in that, Based on the predicted health index, determine whether to trigger a fault warning signal, including: confirming the triggering of a fault warning signal in response to the predicted health index being lower than a preset health index threshold.

7. A data-driven wind turbine converter fault prediction system, characterized in that, include: The data acquisition module is used to acquire high-frequency CMS data and low-frequency SCADA data; The feature window processing module is used to perform feature window processing on high-frequency CMS data and low-frequency SCADA data to obtain signal window data and operating condition vectors. The feature extraction module is used to perform dual-stream feature extraction on signal window data and operating condition vectors to obtain signal time-series feature maps and operating condition feature vectors. The attention modulation module is used to input the signal timing feature map and the operating condition feature vector into the operating condition attention modulation layer to obtain the operating condition modulated signal timing feature vector. The index prediction module is used to perform online health index prediction based on the timing feature vector of the operating condition modulation signal to obtain the predicted health index. The early warning determination module is used to determine whether a fault early warning signal should be triggered based on the predicted health index.

8. The data-driven wind turbine converter fault prediction system according to claim 7, characterized in that, The feature extraction module includes: The signal encoding unit is used to input the signal window data into the signal encoder to obtain the signal timing feature map; The operating condition processing unit is used to input the operating condition vector into the operating condition processor to obtain the operating condition feature vector.

9. The data-driven wind turbine converter fault prediction system according to claim 8, characterized in that, The attention modulation module includes: The node decomposition unit is used to perform node decomposition on the signal timing feature map along the channel direction to obtain a set of signal timing feature vectors; The matrix calculation unit is used to calculate the dynamic context correlation matrix of the set of signal timing feature vectors using the operating condition feature vector as a modulation reference to obtain the signal timing time-varying context correlation matrix. The model input unit is used to input the set of signal time-series feature vectors and the signal time-series time-varying context association matrix into the graph convolutional neural network model to obtain the working condition prior driving signal time-series node structure relationship matrix; The dynamic fusion unit is used to perform cross-modal dynamic fusion of the time sequence node structure relationship matrix of the prior driving signal of the operating conditions, using the feature vector of the operating conditions as a query vector, to obtain the time sequence feature vector of the modulated signal of the operating conditions.

Images