Trusted fan early warning method and system based on domain self-adaptive fusion
By using domain adaptive fusion technology, multi-domain features of wind turbine operating signals are filtered and weighted. Combined with channel decoupling convolution and feature fusion, the problem of early fault identification of wind turbines is solved, and more efficient early warning and interpretability are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XI AN JIAOTONG UNIV
- Filing Date
- 2026-03-26
- Publication Date
- 2026-06-05
AI Technical Summary
Existing wind turbine fault early warning models struggle to effectively identify early, subtle faults when faced with non-stationary, nonlinear, and noisy wind turbine operating signals, and lack interpretability, resulting in insufficient timeliness and accuracy of early warnings.
A domain-adaptive fusion method is adopted. By using complete empirical mode decomposition and wavelet packet decomposition of adaptive noise, key feature matrices are selected and adaptively weighted. Combined with channel decoupling convolution and feature fusion, the matrix is input into a classifier for early fault warning.
It improves the characterization ability of early fault features of wind turbines and the accuracy of early warning, enhances the model's ability to identify early faults, and provides interpretability through SHAP value evaluation, thereby improving the credibility of early warning results.
Smart Images
Figure CN122153537A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of reliability assessment of complex equipment operation and maintenance processes, specifically to a reliable early warning method and system for wind turbines based on domain adaptive fusion. Background Technology
[0002] With the development of deep learning technology, intelligent fault diagnosis and condition assessment methods based on models such as convolutional neural networks are gradually being applied to the operation monitoring of rotating machinery equipment such as wind turbine generators. Compared with traditional methods that rely on human experience, deep learning models can automatically extract features through end-to-end training, which improves the accuracy of fault identification and condition judgment to a certain extent, providing a new technical means for wind turbine operation condition monitoring and fault early warning.
[0003] However, wind turbine operating signals are typically non-stationary, nonlinear, and noisy, and early-stage weak signals are difficult to identify. Inputs in a single time domain or feature space are insufficient to comprehensively represent the multi-scale, multi-frequency operating status information contained within the signal. To improve the model's ability to perceive complex signals, existing research has attempted to introduce multi-domain features (time domain, frequency domain, and time-frequency domain) into deep learning models, achieving multi-domain information fusion through multi-channel input or feature concatenation, thus enriching feature representation to some extent. However, the contributions of different domain features to fault representation are significantly uneven, and early fault signals are weak and susceptible to noise interference. Existing methods mostly employ fixed-weight superposition or concatenation fusion strategies, failing to highlight key domain features and easily introducing redundant information or domain noise interference, thereby weakening the model's ability to identify early faults and affecting the timeliness and accuracy of early warnings.
[0004] On the other hand, most existing wind turbine fault prediction models are black-box models based on complex neural network structures, lacking interpretability and reliability in their fault prediction results, making it difficult to support subsequent operation and maintenance tasks. Some studies have attempted to interpret the model output using feature visualization methods; however, this approach has significant drawbacks. It can only perform a superficial analysis of the model's surface features and cannot clearly reveal how different inputs influence the model's final decision. Specifically, for multi-source input data, this method cannot quantify the degree of influence of different feature sources on the model's final decision, thus limiting the application value of deep learning-based fault prediction models in actual wind turbine operation and maintenance decision-making. Summary of the Invention
[0005] To overcome at least one deficiency in the prior art, this application provides a reliable early warning method and system for wind turbines based on domain adaptive fusion.
[0006] Firstly, a reliable early warning method for wind turbines based on domain adaptive fusion is provided, including: The original signal of wind turbine operation is decomposed using a complete empirical mode decomposition method based on adaptive noise to obtain multiple intrinsic mode function components. The multiple intrinsic mode function components are then screened using the kurtosis similarity criterion to construct a modal domain signal feature matrix. The original wind turbine operation signal is decomposed into wavelet packets to obtain multiple wavelet packet subband coefficient sequences. The multiple wavelet packet coefficient sequences are then filtered using a cumulative probability distribution sampling strategy to construct a wavelet domain signal feature matrix. The original wind turbine operation signal is normalized to obtain the time-domain signal feature matrix; adaptive weighting is performed on the modal domain signal feature matrix, wavelet domain signal feature matrix, and time-domain signal feature matrix respectively to obtain the weighted feature matrix of each domain; the weighted feature matrices of each domain are fused to obtain the weighted multi-domain joint feature representation. Channel-decoupled convolution is performed on the weighted multi-domain joint feature representation to obtain the spatial response results of each channel; the spatial response results of each channel are then fused to obtain the fused feature. The fused features are input into the classifier to obtain early fault warning results for wind turbines.
[0007] In one embodiment, multiple intrinsic mode function components are filtered using a kurtosis similarity criterion to construct a modal domain signal feature matrix, including: Calculate the kurtosis similarity between the original operating signal of the wind turbine and each intrinsic mode function component; The intrinsic mode function components with kurtosis similarity greater than a set threshold are retained to form the modal domain signal feature matrix.
[0008] In one embodiment, kurtosis similarity is calculated using the following formula:
[0009] in, The original signal for wind turbine operation and the first Kurtosis similarity between intrinsic mode function components For the first The intrinsic mode function components at the th eigenmode function component in the ... Kurtosis values on a time-domain sampled signal For the first The kurtosis value of a time-domain sampled signal. This represents the number of time-domain sampled signals.
[0010] In one embodiment, a wavelet domain signal feature matrix is constructed by filtering multiple wavelet subband coefficient sequences using a cumulative probability distribution sampling strategy, including: Calculate the energy eigenvalues of each wavelet subband coefficient sequence; The energy characteristic values are normalized to obtain the energy percentage; Sort the energy percentages from largest to smallest and calculate the cumulative probability; When the cumulative probability is greater than the cumulative probability threshold, all wavelet subband coefficient sequences corresponding to the cumulative probability are selected to form the wavelet domain signal feature matrix.
[0011] In one embodiment, the adaptive weighting process includes: For each domain's signal feature matrix, the global response intensity of each feature channel is calculated. The global response intensities of all feature channels constitute the channel description vector. The global response intensity is calculated using the following formula:
[0012] in, This represents the global response intensity of the c-th feature channel in the d-th domain. The height of the signal feature matrix. The width of the signal feature matrix. This represents the signal feature matrix of the d-th domain in the c-th feature channel. The value of the element at that position; The channel description vector is input into the nonlinear threshold generation structure composed of two-level mapping units to obtain the adaptive weight vector. The adaptive weight vector is applied to the signal feature matrix according to the channel dimension to obtain the weighted feature matrix.
[0013] In one embodiment, the method further includes: performing an interpretability assessment of the early fault warning results of wind turbine equipment based on the SHAP value.
[0014] In one embodiment, the interpretability assessment of early fault warning results for wind turbine equipment based on SHAP values includes: For each signal feature involved in the early warning, calculate the overall contribution of the signal feature to the early warning result; Based on the overall contribution, determine the positive or negative impact of each signal feature on the early warning results of wind turbine faults.
[0015] In one embodiment, the overall contribution is calculated using the following formula:
[0016]
[0017] in, For the first individual signal characteristics The corresponding overall contribution, Representing a feature subset The number of elements, The feature set formed by the various signal characteristics involved in the early warning Excluding the individual signal characteristics The feature subset formed later This represents the total number of all signal features involved in the early warning system. For the first The marginal effect of a signal feature for The corresponding warning results, for The corresponding warning results.
[0018] Secondly, a reliable early warning system for wind turbines based on domain adaptive fusion is provided, including: The modal domain signal feature matrix construction module is used to decompose the original wind turbine operation signal using a complete empirical mode decomposition method based on adaptive noise to obtain multiple intrinsic mode function components; the multiple intrinsic mode function components are then filtered using the kurtosis similarity criterion to construct the modal domain signal feature matrix. The wavelet domain signal feature matrix construction module is used to perform wavelet packet decomposition on the original wind turbine operation signal to obtain multiple wavelet packet subband coefficient sequences; the multiple wavelet packet coefficient sequences are filtered through a cumulative probability distribution sampling strategy to construct the wavelet domain signal feature matrix. The domain adaptive fusion module is used to normalize the original wind turbine operation signal to obtain the time-domain signal feature matrix; adaptive weighting is performed on the modal domain signal feature matrix, wavelet domain signal feature matrix and time-domain signal feature matrix respectively to obtain the weighted feature matrix of each domain; the weighted feature matrices of each domain are fused to obtain the weighted multi-domain joint feature representation. The channel decoupling and fusion module is used to perform channel decoupling convolution on the weighted multi-domain joint feature representation to obtain the spatial response results of each channel; and to fuse the spatial response results of each channel to obtain the fused feature. The classification module is used to input the fused features into the classifier to obtain early fault warning results for wind turbines.
[0019] Thirdly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the aforementioned trusted early warning method for wind turbines based on domain adaptive fusion.
[0020] Compared with the prior art, this application has the following beneficial effects: 1. By constructing modal domain features and wavelet domain features of wind turbine operating signals, and introducing modal screening and cumulative probability distribution sampling mechanisms respectively, effective screening and recombination of multi-domain features can be achieved, which can suppress noise interference and redundant information and improve the characterization ability of early weak fault features of wind turbines.
[0021] 2. By compressing and transforming multi-domain features and generating adaptive weighted vectors, adaptive weighting is applied to different feature domains, thereby enhancing the feature domains that contribute significantly to the characterization of early-stage wind turbine faults. This solves the problem that fixed weights or simple superposition in existing multi-domain fusion methods make it difficult to reflect differences in feature contributions.
[0022] 3. Based on multi-domain feature weighting, a channel decoupled convolution and group feature extraction mechanism is introduced to independently model features of different domains and achieve joint representation through a feature fusion module, thereby improving the pertinence and stability of the early weak fault feature extraction process of wind turbines.
[0023] 4. The early warning results are explained by using feature contribution analysis, which improves the credibility and interpretability of the early warning results for wind turbine failure risks. Attached Figure Description
[0024] 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 a reliable early warning method for wind turbines based on domain adaptive fusion is shown. Figure 2 A schematic diagram of wavelet packet decomposition and wavelet packet energy is shown; Figure 3 This illustrates the intent of the SHAP interpretation method framework. Detailed Implementation
[0025] 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.
[0026] 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.
[0027] 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.
[0028] This application provides a reliable early warning method for wind turbines based on domain adaptive fusion. Figure 1 A flowchart of a trusted early warning method for wind turbines based on domain adaptive fusion is shown. (See [link]) Figure 1 The method mainly includes the following steps: Step S1: The original wind turbine operation signal is decomposed using a complete empirical mode decomposition method based on adaptive noise to obtain multiple intrinsic mode function components; the multiple intrinsic mode function components are screened using the kurtosis similarity criterion to construct the mode domain signal feature matrix.
[0029] Step S2 involves performing wavelet packet decomposition on the original wind turbine operating signal to obtain multiple wavelet packet subband coefficient sequences. These sequences are then filtered using a cumulative probability distribution sampling strategy to construct a wavelet domain signal feature matrix. Here, a wavelet packet decomposition method based on Daubechies wavelet basis functions can be used; this is an existing technique and will not be described in detail. Figure 2 A schematic diagram of wavelet packet decomposition and wavelet packet energy is shown.
[0030] Step S3: Normalize the original wind turbine operation signal to obtain the time-domain signal feature matrix; perform adaptive weighting on the modal domain signal feature matrix, wavelet domain signal feature matrix, and time-domain signal feature matrix respectively to obtain the weighted feature matrix of each domain; fuse the weighted feature matrices of each domain to obtain the weighted multi-domain joint feature representation.
[0031] Step S4: Perform channel-decoupled convolution on the weighted multi-domain joint feature representation to obtain the spatial response results of each channel; fuse the spatial response results of each channel to obtain the fused feature.
[0032] Step S5: Input the fused features into the classifier to obtain the early fault warning result for the wind turbine. The early fault warning result for the wind turbine refers to the type of wind turbine fault.
[0033] Here, steps S4 and S5 are based on the constructed early fault warning model for wind turbines, which is a model trained on a training dataset. The samples in the training dataset are the raw operating signals of the wind turbines collected, and the samples are labeled with categories, including: blade faults, coupling faults, gear faults, etc.
[0034] To address the issues of limited sample quantity and uneven category distribution of early-stage wind turbine faults, a continuous time-domain window overlapping sampling method is used to augment the collected raw wind turbine operating signals. By setting a fixed-length sliding time window on the time-domain signal and successively shifting and truncating it according to a preset step size, multiple overlapping sub-signal samples are formed. The augmented sample quantity satisfies the following conditions:
[0035] in, Indicates the total length of the original signal. This represents the width of the sliding time window and is used as the signal length for a single training sample. This indicates the step size of the sliding window. This indicates the number of samples obtained after data augmentation.
[0036] In one embodiment, step S1 involves filtering multiple intrinsic mode function components using a kurtosis similarity criterion to construct a modal domain signal feature matrix, including: The kurtosis similarity between the original operating signal of the wind turbine and each intrinsic mode function component is calculated. Here, the symmetric mean absolute percentage error (SMAPE) is used to calculate the kurtosis similarity, specifically using the following formula:
[0037] in, The original signal for wind turbine operation and the first Kurtosis similarity between intrinsic mode function components For the first The intrinsic mode function components at the th eigenmode function component in the ... The kurtosis value is a kurtosis value on a time-domain sampled signal. The original signal of wind turbine operation includes multiple time-domain sampled signals. The kurtosis value is used to characterize the impulsiveness and non-Gaussian characteristic intensity of the signal. For the first The kurtosis value of a time-domain sampled signal. This represents the number of time-domain sampled signals.
[0038] The intrinsic mode function components with kurtosis similarity greater than a set threshold are retained to form the modal domain signal feature matrix.
[0039] In one embodiment, step S2 involves filtering multiple wavelet subband coefficient sequences using a cumulative probability distribution sampling strategy to construct a wavelet domain signal feature matrix, including: Calculate the energy eigenvalues of each wavelet subband coefficient sequence using the following formula:
[0040] in, For the first Energy eigenvalues of wavelet packet coefficient sequences For the first The first wavelet packet coefficient sequence in the th wavelet packet subband coefficient sequence One value; The energy characteristic values are normalized to obtain the energy percentage; the following formula is used:
[0041] in, For the first Energy percentage, For the first Energy eigenvalues of wavelet packet coefficient sequences denoted as the number of wavelet packet coefficient sequences.
[0042] Sort the energy percentages from largest to smallest and calculate the cumulative probability using the following formula:
[0043] in, Let r be the cumulative probability of the energy percentage of the first r items in the sorting results. This indicates the sorting result. item.
[0044] When the cumulative probability is greater than the cumulative probability threshold At that time, all wavelet subband coefficient sequences corresponding to the cumulative probability are selected to form the wavelet domain signal feature matrix.
[0045] In one embodiment, step S3, adaptive weighting processing, includes: For the signal feature matrix of each domain Where d represents different feature domain types, Indicates the corresponding channel number. The height of the signal feature matrix. Let be the width of the signal feature matrix; calculate the global response intensity of each feature channel, and the global response intensities of all feature channels constitute the channel description vector. The global response strength is calculated using the following formula:
[0046] in, This represents the global response intensity of the c-th feature channel in the d-th domain. The height of the signal feature matrix. The width of the signal feature matrix. This represents the signal feature matrix of the d-th domain in the c-th feature channel. The value of the element at that position; Channel description vector An adaptive weight vector is obtained by inputting a nonlinear threshold generation structure composed of two-level mapping units. The following formula is used:
[0047] in, and For learnable mapping matrix, This indicates the introduction of a non-linear activation mapping function. This represents a normalization mapping function used to restrict weights within a preset interval, generating... That is, an adaptive weight vector for different domain features; The adaptive weight vector is applied to the signal feature matrix according to the channel dimension to obtain the weighted feature matrix. The following formula is used:
[0048] in, This represents the feature of the c-th feature channel in the d-th domain after adaptive weighting. express In position The element at that location, This represents the adaptive weight corresponding to the c-th feature channel in the d-th domain.
[0049] Then, the weighted feature matrices of each domain are fused according to channel or spatial dimension to obtain a weighted multi-domain joint feature representation.
[0050] In one embodiment, step S4 involves performing channel-decoupled convolution on the weighted multi-domain joint feature representation to obtain the spatial response results for each channel; using the following formula:
[0051] in, This represents the feature of the c-th feature channel in the weighted multi-domain joint feature representation. This represents the independent convolutional kernel corresponding to the c-th feature channel. This represents the spatial response result of the c-th feature channel.
[0052] Then, the spatial response results of each channel are fused to obtain the fused features; here, pointwise convolution is used for fusion.
[0053] Use the following formula:
[0054] in, This represents the channel mapping coefficients of the pointwise convolution of the c-th feature channel. Indicates fusion characteristics; This represents the number of feature channels.
[0055] In one embodiment, the method further includes: performing an interpretability assessment of the early fault warning results of wind turbine equipment based on the SHAP value.
[0056] Figure 3 The framework of the SHAP interpretation method is illustrated. A quantitative analysis method based on SHAP values is used to interpret the early warning results of wind turbine failure risks, improving the reliability of the warning results. Specifically, this includes: First, for each signal feature involved in the early warning, the overall contribution of the signal feature to the early warning result is calculated; the overall contribution is calculated using the following formula:
[0057]
[0058] in, For the first individual signal characteristics The corresponding overall contribution, Representing a feature subset The number of elements, The feature set formed by the various signal characteristics involved in the early warning Excluding the individual signal characteristics The feature subset formed later This represents the total number of all signal features involved in the early warning system. For the first The marginal effect of a signal feature for The corresponding warning results, for The corresponding warning results.
[0059] Here, a mechanism for measuring the contribution of feature vectors from different spatial domains involved in early warning judgment to the model output is constructed. By comprehensively considering the marginal impact of each feature on the early warning output under different combination conditions, the importance of different features to the final early warning result is characterized.
[0060] Then, based on the overall contribution, the positive or negative impact of each signal feature on the early warning result of wind turbine failure is determined.
[0061] Here, a contribution threshold can be set, and the calculated comprehensive contribution is compared with the contribution threshold. If it is greater than the contribution threshold, the signal characteristics are considered to have a positive impact on the early warning results of wind turbines; otherwise, the signal characteristics are considered to have a negative impact on the early warning results of wind turbines.
[0062] Furthermore, the role of signal features in early warning judgment can be visualized or quantitatively explained based on the comprehensive contribution, clarifying the positive or negative impact of each feature on the early warning results of wind turbine failures, thereby achieving interpretability of the early warning results and improving the credibility of early warning decisions.
[0063] Employing the same inventive concept as the domain-adaptive fusion-based trusted wind turbine early warning method, this embodiment also provides a corresponding domain-adaptive fusion-based trusted wind turbine early warning system, including: The modal domain signal feature matrix construction module is used to decompose the original wind turbine operation signal using a complete empirical mode decomposition method based on adaptive noise to obtain multiple intrinsic mode function components; the multiple intrinsic mode function components are then filtered using the kurtosis similarity criterion to construct the modal domain signal feature matrix. The wavelet domain signal feature matrix construction module is used to perform wavelet packet decomposition on the original wind turbine operation signal to obtain multiple wavelet packet subband coefficient sequences; the multiple wavelet packet coefficient sequences are filtered through a cumulative probability distribution sampling strategy to construct the wavelet domain signal feature matrix. The domain adaptive fusion module is used to normalize the original wind turbine operation signal to obtain the time-domain signal feature matrix; adaptive weighting is performed on the modal domain signal feature matrix, wavelet domain signal feature matrix and time-domain signal feature matrix respectively to obtain the weighted feature matrix of each domain; the weighted feature matrices of each domain are fused to obtain the weighted multi-domain joint feature representation. The channel decoupling and fusion module is used to perform channel decoupling convolution on the weighted multi-domain joint feature representation to obtain the spatial response results of each channel; and to fuse the spatial response results of each channel to obtain the fused feature. The classification module is used to input the fused features into the classifier to obtain early fault warning results for wind turbines.
[0064] The trusted wind turbine early warning system based on domain adaptive fusion in this embodiment has the same inventive concept as the trusted wind turbine early warning method based on domain adaptive fusion described above. Therefore, the specific implementation of this device can be found in the embodiment section of the trusted wind turbine early warning method based on domain adaptive fusion described above, and its technical effects correspond to the technical effects of the above method, so it will not be repeated here.
[0065] In summary, this application uses knowledge importance to deeply mine the weak fault features contained in different feature domains, thereby realizing the identification and early warning of early fault risks of wind turbines. Furthermore, it evaluates the influence of features in different domains on the early warning results of wind turbines through feature contribution, thus achieving quantitative interpretability of the model decision results.
[0066] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the aforementioned trusted wind turbine early warning method based on domain adaptive fusion.
[0067] 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 reliable early warning method for wind turbines based on domain adaptive fusion, characterized in that, include: The original signal of wind turbine operation is decomposed using a complete empirical mode decomposition method based on adaptive noise to obtain multiple intrinsic mode function components; The modal domain signal feature matrix is constructed by filtering the multiple intrinsic mode function components using the kurtosis similarity criterion. The original wind turbine operation signal is decomposed using wavelet packet decomposition to obtain multiple wavelet packet subband coefficient sequences; The multiple wavelet subband coefficient sequences are filtered using a cumulative probability distribution sampling strategy to construct a wavelet domain signal feature matrix; The original operating signal of the wind turbine is normalized to obtain the time-domain signal feature matrix; Adaptive weighting is performed on the modal domain signal feature matrix, the wavelet domain signal feature matrix, and the time domain signal feature matrix to obtain the weighted feature matrices of each domain; the weighted feature matrices of each domain are then fused to obtain a weighted multi-domain joint feature representation. Channel-decoupled convolution is performed on the weighted multi-domain joint feature representation to obtain the spatial response results of each channel; the spatial response results of each channel are then fused to obtain the fused feature. The fused features are input into the classifier to obtain early fault warning results for wind turbines.
2. The method as described in claim 1, characterized in that, in, The multiple intrinsic mode function components are filtered using the kurtosis similarity criterion to construct a modal domain signal feature matrix, including: Calculate the kurtosis similarity between the original operating signal of the wind turbine and each intrinsic mode function component; The intrinsic mode function components with kurtosis similarity greater than a set threshold are retained to form the modal domain signal feature matrix.
3. The method as described in claim 2, characterized in that, The kurtosis similarity is calculated using the following formula: in, The original signal for wind turbine operation and the first Kurtosis similarity between intrinsic mode function components For the first The intrinsic mode function components at the th eigenmode function component in the ... Kurtosis values on a time-domain sampled signal For the first The kurtosis value of a time-domain sampled signal. This represents the number of time-domain sampled signals.
4. The method as described in claim 1, characterized in that, in, The multiple wavelet subband coefficient sequences are filtered using a cumulative probability distribution sampling strategy to construct a wavelet domain signal feature matrix, including: Calculate the energy eigenvalues of each wavelet subband coefficient sequence; The energy characteristic values are normalized to obtain the energy percentage; Sort the energy percentages from largest to smallest and calculate the cumulative probability; When the cumulative probability is greater than the cumulative probability threshold, all wavelet subband coefficient sequences corresponding to the cumulative probability are selected to form the wavelet domain signal feature matrix.
5. The method as described in claim 1, characterized in that, The adaptive weighting process includes: For each domain's signal feature matrix, the global response intensity of each feature channel is calculated, and the global response intensities of all feature channels constitute the channel description vector; the global response intensity is calculated using the following formula: in, This represents the global response intensity of the c-th feature channel in the d-th domain. The height of the signal feature matrix. The width of the signal feature matrix. This represents the signal feature matrix of the d-th domain in the c-th feature channel. The value of the element at that position; The channel description vector is input into a nonlinear threshold generation structure composed of two-level mapping units to obtain an adaptive weight vector. The adaptive weight vector is applied to the signal feature matrix according to the channel dimension to obtain the weighted feature matrix.
6. The method as described in claim 1, characterized in that, The method further includes: performing an interpretability assessment of the early fault warning results of the wind turbine equipment based on the SHAP value.
7. The method as described in claim 5, characterized in that, in, An interpretability assessment of the early fault warning results of the wind turbine equipment based on the SHAP value is performed, including: For each signal feature involved in the early warning, calculate the overall contribution of the signal feature to the early warning result; Based on the overall contribution, the positive or negative impact of each signal feature on the early fault warning result of the wind turbine is determined.
8. The method as described in claim 7, characterized in that, The overall contribution is calculated using the following formula: in, For the first individual signal characteristics The corresponding overall contribution, Representing a feature subset The number of elements, The feature set formed by the various signal characteristics involved in the early warning Excluding the individual signal characteristics The feature subset formed later This represents the total number of all signal features involved in the early warning system. For the first The marginal effect of a signal feature for The corresponding warning results, for The corresponding warning results.
9. A reliable early warning system for wind turbines based on domain adaptive fusion, characterized in that, include: The modal domain signal feature matrix construction module is used to decompose the original wind turbine operation signal using a complete empirical mode decomposition method based on adaptive noise to obtain multiple intrinsic mode function components. The modal domain signal feature matrix is constructed by filtering the multiple intrinsic mode function components using the kurtosis similarity criterion. The wavelet domain signal feature matrix construction module is used to perform wavelet packet decomposition on the original wind turbine operation signal to obtain multiple wavelet packet subband coefficient sequences; and to filter the multiple wavelet packet subband coefficient sequences through a cumulative probability distribution sampling strategy to construct a wavelet domain signal feature matrix. The domain adaptive fusion module is used to normalize the original wind turbine operation signal to obtain the time-domain signal feature matrix; Adaptive weighting is performed on the modal domain signal feature matrix, the wavelet domain signal feature matrix, and the time domain signal feature matrix to obtain the weighted feature matrices of each domain; the weighted feature matrices of each domain are then fused to obtain a weighted multi-domain joint feature representation. The channel decoupling and fusion module is used to perform channel decoupling convolution on the weighted multi-domain joint feature representation to obtain the spatial response results of each channel; and to fuse the spatial response results of each channel to obtain the fused feature. The classification module is used to input the fused features into the classifier to obtain early fault warning results for wind turbines.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the trusted wind turbine early warning method based on domain adaptive fusion as described in any one of claims 1-8.