Explainable fan fault early warning method and system based on multi-source knowledge
By adaptively fusing multi-source knowledge and making interpretable decisions, the problems of low data utilization and uninterpretable decisions in wind turbine fault early warning are solved, achieving high-accuracy fault early warning and supporting the scientific maintenance of wind turbines.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XI AN JIAOTONG UNIV
- Filing Date
- 2026-03-31
- Publication Date
- 2026-06-05
Smart Images
Figure CN122153538A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of fault monitoring and predictive maintenance of complex electromechanical systems, and specifically to an interpretable wind turbine fault early warning method and system based on multi-source knowledge. Background Technology
[0002] Wind energy, as a core component of clean and renewable energy, occupies a crucial position in the global energy transition. With the continuous increase in the capacity and service life of individual wind turbines, their workload and failure risks are rising simultaneously. Furthermore, wind turbine systems operate under variable loads, strong vibrations, and harsh conditions for extended periods, making them highly susceptible to various failures such as gear tooth breakage, bearing pitting, and shaft imbalance. If effective early warning systems are not implemented in a timely manner, these failures can rapidly escalate, triggering a chain reaction that leads to unit shutdowns or even serious safety accidents, resulting in significant economic losses and energy waste. Therefore, conducting research on accurate and reliable fault early warning systems for wind turbines has significant engineering application value.
[0003] In traditional wind turbine fault early warning technologies, the model-driven fault warning process lacks prior mechanistic knowledge and relies solely on sensor data, easily overlooking the correlation between the inherent attributes of the equipment and operating conditions, thus limiting the model's generalization ability. Some studies use mechanistic knowledge to guide the model's decision-making process, but the effective utilization rate of knowledge remains insufficient. Furthermore, most mainstream deep learning models are "black box" structures, lacking interpretability in their decision-making process and early warning results, leading to high decision-making risks and difficulty in supporting subsequent fault repair work. Current research mostly only analyzes the problem from a feature visualization perspective, addressing the symptoms but not the root cause. In summary, existing wind turbine fault early warning methods face the following bottlenecks that urgently need to be addressed:
[0004] (1) Low utilization of multi-source knowledge, which fails to fully utilize the limited data. Deep and effective fusion of multi-source knowledge is a key link to improve early warning performance. However, due to the significant differences in quality, relevance and effective information density of different source knowledge, traditional methods use fixed weights or simple weighted average fusion strategies, which are difficult to accurately distinguish the hierarchical differences in the contribution of different knowledge sources to fault early warning. This not only fails to fully explore the deep value of high-quality data sources and waste the limited data utility, but also introduces redundant noise from low-quality knowledge sources into the model indiscriminately, interfering with the subsequent feature extraction and fault identification process, thus restricting the overall performance of the early warning model.
[0005] (2) The black box nature of the decision-making process for early warning results and the lack of interpretability of early warning results. Existing interpretability studies mostly stay at the level of feature map visualization, which can only show "which areas of the input data the model focuses on", but cannot go deep into the neuron level to reveal how the model makes decisions based on these areas. This lack of interpretability increases the decision-making risk of the model in high-risk industrial scenarios, making it difficult for maintenance personnel to have sufficient trust in the model output. On the other hand, when the early warning results are biased or false alarms occur, due to the lack of neuron-level decision tracing capabilities, subsequent root cause analysis and maintenance decisions are difficult to carry out effectively, which seriously restricts the engineering practical value of the early warning system. Summary of the Invention
[0006] To overcome at least one deficiency in the prior art, this application provides an interpretable wind turbine fault early warning method and system based on multi-source knowledge.
[0007] Firstly, an interpretable wind turbine fault early warning method based on multi-source knowledge is provided, including: Collect time-series data of the fan, which includes vibration data, temperature data, and speed data; The wind turbine time series data were preprocessed using a time series dimensionality reduction method based on Patch Embedding to obtain standardized time series features. For standardized temporal features, an attention network is used to extract temporal autocorrelation features, and an adaptive kernel network is used to extract temporal cross-correlation features from the temporal autocorrelation features. Based on the equipment parameters of the wind turbine, a structured prior knowledge is constructed; the structured prior knowledge is encoded, and a BiGRU model is used to extract prior features; The temporal cross-correlation features and prior features are input into the multi-source knowledge adaptive fusion module to obtain the multi-source knowledge fusion features; the multi-source knowledge adaptive fusion module includes Depth-Wise convolutional units and Point-Wise convolutional units. The multi-source knowledge fusion features are input into the interpretable decision module to obtain the model decision output; the interpretable decision module includes multiple interpretable neurons based on B-spline curves, and the model decision output includes the predicted probability of each wind turbine fault category; Fault warnings are generated based on the model's decision output.
[0008] In one embodiment, structured prior knowledge is encoded and prior features are extracted using a BiGRU model, employing the following formula:
[0009]
[0010]
[0011]
[0012] in, For the first Each prior knowledge item Encoded prior knowledge vector For linear coding layer, For standardized operation, For the first Each prior knowledge item The corresponding numerical vector composed of device parameters For the first Each prior knowledge item The corresponding numerical vector composed of operating parameters Indicates addition. For the first The hidden states of the forward GRU corresponding to each prior knowledge vector. Indicates forward GRU, For the first -1 hidden states of the forward GRU corresponding to the prior knowledge vectors, For the first The hidden states of the backward GRU corresponding to each prior knowledge vector. Indicates backward GRU, For the first The hidden states of the backward GRU corresponding to each prior knowledge vector. As prior features, Indicates splicing, For the first The hidden states of the backward GRU corresponding to each prior knowledge vector. Indicates the first The hidden states of the forward GRU corresponding to each prior knowledge vector. is the length of the prior knowledge vector sequence.
[0013] In one embodiment, the multi-source knowledge adaptive fusion module is used to implement the following functions: The temporal cross-correlation features and prior features are concatenated to obtain the concatenated features; Depth-Wise convolutional units are used to perform convolution operations on the concatenated features to obtain local features; Local features are globally pooled to obtain channel-level global description vectors. Using the channel-level global description vectors, nonlinear interactions between channels are learned based on a two-layer bottleneck structure to generate adaptive weight vectors. The adaptive weight vectors are then used to adaptively weight local features channel by channel to obtain adaptively weighted multi-source knowledge features. Point-Wise convolutional units are used to perform feature map-level cross-channel knowledge fusion on adaptively weighted multi-source knowledge features to obtain multi-source knowledge fusion features.
[0014] In one embodiment, the adaptively weighted multi-source knowledge features are represented by the following formula:
[0015] in, For adaptively weighted multi-source knowledge features, It is the Sigmoid activation function. , The parameters are for the double-layer bottleneck structure. It is the ReLU activation function. This is a channel-level global description vector. These are local features.
[0016] In one embodiment, the multi-source knowledge fusion feature is represented by the following formula:
[0017] in, As a feature of multi-source knowledge fusion, The weight parameters are those of the Point-Wise convolution kernel. This is a bias term.
[0018] In one embodiment, the interpretability neuron based on B-spline curves specifically implements the following function:
[0019] in, For the output of an interpretable neuron based on B-spline curves, For the first The coefficients of the B-spline basis functions For the first A B-spline basis function, As a feature of multi-source knowledge fusion, For the first The node vectors of the B-spline basis functions The number of B-spline basis functions in each B-spline-based interpretable neuron.
[0020] In one embodiment, fault warning is generated based on the model decision output, including: The wind turbine fault category with the highest predicted probability in the model decision output is taken as the final predicted fault category; The warning level is determined based on the predicted probability of the final predicted fault category.
[0021] Secondly, an interpretable wind turbine fault early warning system based on multi-source knowledge is provided, including: The timing data acquisition module is used to collect timing data of the fan, which includes vibration data, temperature data, and speed data. The data preprocessing module is used to preprocess the wind turbine time series data using a time series dimensionality reduction method based on Patch Embedding to obtain standardized time series features; The temporal cross-correlation feature extraction module is used to extract temporal autocorrelation features from standardized temporal features using an attention network, and to extract temporal cross-correlation features from temporal autocorrelation features using an adaptive kernel network. The prior feature extraction module is used to construct structured prior knowledge based on the equipment parameters of the wind turbine; the structured prior knowledge is encoded and prior features are extracted using a BiGRU model; The multi-source knowledge adaptive fusion module is used to obtain multi-source knowledge fusion features based on temporal cross-correlation features and prior features; the multi-source knowledge adaptive fusion module includes Depth-Wise convolutional units and Point-Wise convolutional units. The interpretable decision module is used to obtain the model decision output based on the multi-source knowledge fusion features. The interpretable decision module includes multiple interpretable neurons based on B-spline curves, and the model decision output includes the predicted probability of each wind turbine fault category. The fault warning module is used to provide fault warnings based on the model's decision output.
[0022] Compared with the prior art, this application has the following beneficial effects: Through the deep fusion of multi-source knowledge such as sensor signals and prior mechanisms under adaptive collaboration, and the neural network architecture of the result interpretable fault early warning model, this application significantly improves the accuracy and reliability of fault early warning, and provides scientific and efficient technical support for the predictive maintenance of wind turbines. Attached Figure Description
[0023] This application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which, together with the detailed description below, are incorporated in and form part of this specification. In the drawings: Figure 1 A flowchart of an interpretable wind turbine fault early warning method based on multi-source knowledge is shown. Figure 2 A schematic diagram of the time series dimensionality reduction method based on patch embedding is shown. Figure 3 A schematic diagram of the confusion matrix results for the training and test sets is shown. Detailed Implementation
[0024] Exemplary embodiments of the present application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of the actual embodiments are described in the specification. However, it should be understood that many embodiment-specific decisions can be made in the development of any such actual embodiment to achieve the developer’s specific objectives, and these decisions may vary as the embodiments differ.
[0025] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the device structure closely related to the solution of this application is shown in the accompanying drawings, while other details that are not closely related to this application are omitted.
[0026] It should be understood that this application is not limited to the described embodiments by virtue of the following description with reference to the accompanying drawings. In this document, embodiments may be combined with each other, features may be substituted or borrowed between different embodiments, and one or more features may be omitted in one embodiment, where feasible.
[0027] This application provides an interpretable wind turbine fault early warning method based on multi-source knowledge. Figure 1 A flowchart of an interpretable wind turbine fault early warning method based on multi-source knowledge is shown. (See [link]) Figure 1 The method mainly includes the following steps: Step S1: Collect the fan timing data, which includes vibration data, temperature data, and speed data. Here, multiple sensors deployed at key parts of the fan are used to collect real-time timing data such as vibration, oil temperature, ambient temperature, and speed.
[0028] Step S2: The wind turbine time series data is preprocessed using a time series dimensionality reduction method based on Patch Embedding to obtain standardized time series features.
[0029] Step S3 involves using an attention network to extract time-series autocorrelation features from the standardized time-series features, and then using an adaptive kernel network to extract time-series cross-correlation features from the time-series autocorrelation features. This enables multi-scale, comprehensive signal feature extraction, providing comprehensive and reliable data support for subsequent fault identification.
[0030] Here, the adaptive kernel network can be SKNet (Selective Kernel Networks).
[0031] In other embodiments, the specific implementation function of the adaptive kernel network can be represented by the following formula:
[0032]
[0033]
[0034] in, For the pooling result, This is a time-series autocorrelation feature. Indicates average pooling. This indicates max pooling. , These are all parameters of the adaptive kernel network. For adaptive kernel weights, These are convolution kernels of different sizes.
[0035] Step S4: Based on the equipment parameters of the wind turbine, construct structured prior knowledge; encode the structured prior knowledge and extract prior features using a BiGRU model.
[0036] Equipment parameters, such as gear module, number of teeth, bearing type, and shaft diameter, are used. Each prior knowledge item is encoded as a high-dimensional vector. The encoded vector sequence is then processed using a BiGRU model to fuse forward and backward hidden states, extracting latent prior features that reveal the correlation between the equipment's inherent attributes and operating conditions. The BiGRU model is a recurrent neural network consisting of two independent GRU units: one processes data forward according to the time series, and the other processes data backward according to the time series.
[0037] Step S5: Input the temporal cross-correlation features and prior features into the multi-source knowledge adaptive fusion module to obtain the multi-source knowledge fusion features; the multi-source knowledge adaptive fusion module includes Depth-Wise convolutional units and Point-Wise convolutional units.
[0038] Deep-Wise convolution is used to extract local spatiotemporal features within a single channel of the signal features, with each channel corresponding to an independent convolution kernel. Based on dynamic weights, multi-source knowledge is adaptively weighted, and Point-Wise convolution is combined to perform cross-channel fusion of local features in a single channel. The output of Point-Wise convolution is used as the final fused feature to achieve deep fusion of signal features and prior features.
[0039] Step S6: Input the multi-source knowledge fusion features into the interpretable decision module to obtain the model decision output; the interpretable decision module includes multiple interpretable neurons based on B-spline curves, and the model decision output includes the predicted probability of each wind turbine fault category.
[0040] Interpretable network neurons are constructed based on learnable B-spline curves and embedded into the output layer of the knowledge fusion module. A single interpretable neuron achieves a nonlinear mapping of input features through a linear combination of multiple B-spline basis functions, with both coefficients and node vectors being learnable parameters. Multiple neurons of this type form the output layer, creating a complete interpretable decision-making module. By analyzing the B-spline coefficients and node vectors, the contribution weights and threshold ranges of features to the decision can be clearly defined. Combined with the visualization of the basis function activation states, the interpretability of fault decision-making is achieved from both theoretical and intuitive perspectives.
[0041] Step S7: Provide fault warning based on the model decision output.
[0042] Here, the interpretable wind turbine fault early warning method based on multi-source knowledge is implemented based on the trained interpretable wind turbine fault early warning model, which includes an attention network, an adaptive kernel network, a BiGRU model, a multi-source knowledge adaptive fusion module, and an interpretable decision module connected in sequence.
[0043] During training, standardized time-series features are obtained by using the time-series data of collected wind turbines based on the Patch Embedding time-series dimensionality reduction method.
[0044] Figure 2 The diagram illustrates the principle of a time series dimensionality reduction method based on patch embedding. First, resampling is performed using a movable sliding window to effectively address the imbalanced sample problem during model training. Then, the resampled one-dimensional time series signal is segmented into continuous patches of fixed length. Each patch is flattened and mapped to a high-dimensional feature space using a linear projection layer. Finally, learnable positional encoding is added to preserve temporal order information, outputting standardized temporal features and significantly improving the model's forward propagation efficiency. The specific formula used is as follows: The sampling formula is:
[0045] in, The original signal length is 1. For the width of the sliding window, The sliding window step size, This represents the number of samples.
[0046] The formula for calculating the number of patches for a single sample is:
[0047] in, The length of a single sample signal. Patch length This represents the number of patches.
[0048] The formula for linear projection is:
[0049] in, For a single timing patch, for A flattened one-dimensional vector, The projection weight matrix is... This is a bias term. To The result after linear projection.
[0050] This embodiment significantly improves the accuracy and reliability of fault warnings by deeply integrating multi-source knowledge such as sensor signals and prior mechanisms in an adaptive and collaborative manner, and by using a neuron architecture for the result-interpretable fault warning model. This provides scientific and efficient technical support for the predictive maintenance of wind turbines.
[0051] In one embodiment, structured prior knowledge is encoded and prior features are extracted using a BiGRU model, employing the following formula:
[0052]
[0053]
[0054]
[0055] in, For the first Each prior knowledge item Encoded prior knowledge vector For linear coding layer, For standardized operation, For the first Each prior knowledge item The corresponding numerical vector composed of device parameters For the first Each prior knowledge item The corresponding numerical vector composed of operating parameters Indicates addition. For the first The hidden states of the forward GRU corresponding to each prior knowledge vector. Indicates forward GRU, For the first -1 hidden states of the forward GRU corresponding to the prior knowledge vectors, For the first The hidden states of the backward GRU corresponding to each prior knowledge vector. Indicates backward GRU, For the first The hidden states of the backward GRU corresponding to each prior knowledge vector. As prior features, Indicates splicing, For the first The hidden states of the backward GRU corresponding to each prior knowledge vector. Indicates the first The hidden states of the forward GRU corresponding to each prior knowledge vector. is the length of the prior knowledge vector sequence.
[0056] In one embodiment, the multi-source knowledge adaptive fusion module is used to implement the following functions: The temporal cross-correlation features and prior features are concatenated to obtain the concatenated features; Depth-Wise convolutional units are used to perform convolution operations on the concatenated features to obtain local features; Local features are globally pooled to obtain channel-level global description vectors. Using the channel-level global description vectors, nonlinear interactions between channels are learned based on a two-layer bottleneck structure to generate adaptive weight vectors. The adaptive weight vectors are then used to adaptively weight local features channel by channel to obtain adaptively weighted multi-source knowledge features. The adaptively weighted multi-source knowledge features are represented by the following formula:
[0057] in, For adaptively weighted multi-source knowledge features, It is the Sigmoid activation function. , The parameters are for the double-layer bottleneck structure. It is the ReLU activation function. This is a channel-level global description vector. These are local features.
[0058] Point-Wise convolutional units are used to perform feature map-level cross-channel knowledge fusion on adaptively weighted multi-source knowledge features to obtain multi-source knowledge fusion features.
[0059] The characteristics of multi-source knowledge fusion are expressed by the following formula:
[0060] in, As a feature of multi-source knowledge fusion, The weight parameters are those of the Point-Wise convolution kernel. This is a bias term.
[0061] In one embodiment, the interpretability neuron based on B-spline curves specifically implements the following function:
[0062] in, For the output of an interpretable neuron based on B-spline curves, For the first The coefficients of the B-spline basis functions For the first One B-spline basis function, As a feature of multi-source knowledge fusion, For the first The node vectors of the B-spline basis functions The number of B-spline basis functions in each B-spline-based interpretable neuron.
[0063] The outputs of multiple interpretable neurons based on B-spline curves collectively constitute the model's decision output. ,Right now , These represent the outputs of each interpretable neuron based on B-spline curves. This represents the number of interpretable neurons based on B-spline curves.
[0064] Specifically, fault warnings are generated based on the model's decision output, including: The wind turbine fault category with the highest predicted probability in the model decision output is taken as the final predicted fault category; Based on the predicted probability of the final predicted fault category Determine the warning level.
[0065] For example, For emergency warning, It is a level 2 warning. Level 1 warning If the result is normal, an early warning result will be output.
[0066] Figure 3 This diagram illustrates the confusion matrix results for the training and test sets. Figure 3 Using the method of this application, the accuracy rate of fault early warning can reach 99%.
[0067] This application also provides an interpretable wind turbine fault early warning system based on multi-source knowledge, including: The timing data acquisition module is used to collect timing data of the fan, which includes vibration data, temperature data, and speed data. The data preprocessing module is used to preprocess the wind turbine time series data using a time series dimensionality reduction method based on Patch Embedding to obtain standardized time series features; The temporal cross-correlation feature extraction module is used to extract temporal autocorrelation features from standardized temporal features using an attention network, and to extract temporal cross-correlation features from temporal autocorrelation features using an adaptive kernel network. The prior feature extraction module is used to construct structured prior knowledge based on the equipment parameters of the wind turbine; the structured prior knowledge is encoded and prior features are extracted using a BiGRU model; The multi-source knowledge adaptive fusion module is used to obtain multi-source knowledge fusion features based on temporal cross-correlation features and prior features; the multi-source knowledge adaptive fusion module includes Depth-Wise convolutional units and Point-Wise convolutional units. The interpretable decision module is used to obtain the model decision output based on the multi-source knowledge fusion features. The interpretable decision module includes multiple interpretable neurons based on B-spline curves, and the model decision output includes the predicted probability of each wind turbine fault category. The fault warning module is used to provide fault warnings based on the model's decision output.
[0068] The interpretable wind turbine fault early warning system based on multi-source knowledge in this embodiment has the same inventive concept as the interpretable wind turbine fault early warning method based on multi-source knowledge described above. Therefore, the specific implementation of this device can be found in the embodiment section of the interpretable wind turbine fault early warning method based on multi-source knowledge described above, and its technical effects correspond to the technical effects of the above method, so it will not be repeated here.
[0069] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A wind turbine fault early warning method based on multi-source knowledge, characterized in that, include: Collect time-series data of the fan, including vibration data, temperature data, and speed data; The wind turbine time series data are preprocessed using a time series dimensionality reduction method based on Patch Embedding to obtain standardized time series features; For the standardized temporal features, an attention network is used to extract temporal autocorrelation features, and an adaptive kernel network is used to extract temporal cross-correlation features from the temporal autocorrelation features; Based on the equipment parameters of the wind turbine, a structured prior knowledge is constructed; the structured prior knowledge is encoded, and a BiGRU model is used to extract prior features; The temporal cross-correlation features and the prior features are input into the multi-source knowledge adaptive fusion module to obtain the multi-source knowledge fusion features; the multi-source knowledge adaptive fusion module includes a Depth-Wise convolutional unit and a Point-Wise convolutional unit. The multi-source knowledge fusion features are input into the interpretable decision module to obtain the model decision output; the interpretable decision module includes multiple interpretable neurons based on B-spline curves, and the model decision output includes the predicted probability of each wind turbine fault category; Fault warnings are generated based on the model's decision output.
2. The method as described in claim 1, characterized in that, in, The structured prior knowledge is encoded, and prior features are extracted using a BiGRU model, employing the following formula: in, For the first Each prior knowledge item Encoded prior knowledge vector For linear coding layer, For standardized operation, For the first Each prior knowledge item The corresponding numerical vector composed of device parameters For the first Each prior knowledge item The corresponding numerical vector composed of operating parameters Indicates addition. For the first The hidden states of the forward GRU corresponding to each prior knowledge vector. Indicates forward GRU, For the first -1 hidden states of the forward GRU corresponding to the prior knowledge vectors, For the first The hidden states of the backward GRU corresponding to each prior knowledge vector. Indicates backward GRU, For the first The hidden states of the backward GRU corresponding to each prior knowledge vector. As prior features, Indicates splicing, For the first The hidden states of the backward GRU corresponding to each prior knowledge vector. Indicates the first The hidden states of the forward GRU corresponding to each prior knowledge vector. is the length of the prior knowledge vector sequence.
3. The method as described in claim 1, characterized in that, The multi-source knowledge adaptive fusion module is used to implement the following functions: The temporal cross-correlation features and the prior features are concatenated to obtain the concatenated features; The concatenated features are convolved using Depth-Wise convolutional units to obtain local features. The local features are globally pooled to obtain a channel-level global description vector. Using the channel-level global description vector, the nonlinear interaction between channels is learned based on a two-layer bottleneck structure to generate an adaptive weight vector. The local features are adaptively weighted channel by channel using the adaptive weight vector to obtain adaptively weighted multi-source knowledge features. Point-Wise convolutional units are used to perform feature map-level cross-channel knowledge fusion on the adaptively weighted multi-source knowledge features to obtain multi-source knowledge fusion features.
4. The method as described in claim 3, characterized in that, The adaptively weighted multi-source knowledge features are represented by the following formula: in, For adaptively weighted multi-source knowledge features, It is the Sigmoid activation function. , The parameters are for the double-layer bottleneck structure. It is the ReLU activation function. This is a channel-level global description vector. These are local features.
5. The method as described in claim 3, characterized in that, The multi-source knowledge fusion feature is expressed by the following formula: in, As a feature of multi-source knowledge fusion, The weight parameters are those of the Point-Wise convolution kernel. This is a bias term.
6. The method as described in claim 1, characterized in that, The specific function of the B-spline curve-based interpretable neuron is as follows: in, For the output of an interpretable neuron based on B-spline curves, For the first The coefficients of the B-spline basis functions For the first One B-spline basis function, As a feature of multi-source knowledge fusion, For the first The node vectors of the B-spline basis functions The number of B-spline basis functions in each B-spline-based interpretable neuron.
7. The method as described in claim 1, characterized in that, in, Fault warnings are generated based on the model's decision output, including: The wind turbine fault category with the highest predicted probability in the model decision output is taken as the final predicted fault category; The warning level is determined based on the predicted probability of the final predicted fault category.
8. An interpretable wind turbine fault early warning system based on multi-source knowledge, characterized in that, include: The timing data acquisition module is used to acquire timing data of the fan, which includes vibration data, temperature data, and speed data. The data preprocessing module is used to preprocess the wind turbine time series data using a time series dimensionality reduction method based on Patch Embedding to obtain standardized time series features; The temporal cross-correlation feature extraction module is used to extract temporal autocorrelation features from the standardized temporal features using an attention network, and to extract temporal cross-correlation features from the temporal autocorrelation features using an adaptive kernel network. The prior feature extraction module is used to construct structured prior knowledge based on the equipment parameters of the wind turbine; encode the structured prior knowledge and extract prior features using a BiGRU model; A multi-source knowledge adaptive fusion module is used to obtain multi-source knowledge fusion features based on the temporal cross-correlation features and the prior features; the multi-source knowledge adaptive fusion module includes a Depth-Wise convolutional unit and a Point-Wise convolutional unit. An interpretable decision module is used to obtain model decision output based on the multi-source knowledge fusion features; the interpretable decision module includes multiple interpretable neurons based on B-spline curves, and the model decision output includes the predicted probability of each wind turbine fault category; The fault warning module is used to provide fault warnings based on the model's decision output.