Wind turbine fault diagnosis method and system based on mechanism data fusion

Abstract

The application belongs to the technical field of wind turbine fault diagnosis. A wind turbine fault diagnosis method and system based on mechanism data fusion are provided. Feature sequences, time sequence features, residuals and differential residuals are spliced and fused to obtain a fusion feature vector. The fusion feature vector is linearly mapped to a high-dimensional hidden space. The features in the high-dimensional hidden space are subjected to matrix multiplication operation between features to obtain a similarity distribution between features. After normalization, an adaptive adjacency matrix is obtained. The adaptive adjacency matrix and the features in the high-dimensional hidden space are subjected to matrix multiplication operation between features again. After normalization, a coupling feature matrix is obtained. The adaptive adjacency matrix and the coupling feature matrix are added to obtain an interactive feature vector. According to the interactive feature vector, a wind turbine fault diagnosis result is obtained. The application realizes adaptive modeling of different feature importance, and reduces misjudgment caused by environmental noise and working condition fluctuations.

Description

Technical Field

[0001] This invention relates to the field of wind turbine fault diagnosis technology, specifically to a wind turbine fault diagnosis method and system based on mechanism data fusion. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Because wind turbines are all-weather electromechanical integrated devices, their main components, including blades, gearboxes, and generators, are subjected to alternating loads and various extreme operating conditions for extended periods, making them prone to a series of faults such as bearing damage, insulation aging, and blade cracking. Therefore, fault diagnosis of wind turbines is essential. Traditional fault diagnosis methods, such as vibration analysis and signal processing, typically rely on expert experience and manual feature extraction, which limits their practical application. On the one hand, manual feature extraction struggles to comprehensively capture the complex operating characteristics of wind turbines under varying conditions, leading to subjective influences on diagnostic results. On the other hand, data from a single signal is insufficient to fully reflect the overall health status of the wind turbine, and during actual operation, environmental noise, load fluctuations, and operating condition switching can all blur the feature signals, further reducing the accuracy and reliability of the diagnosis.

[0004] In recent years, with the improvement of the intelligence level of wind turbine units, data-driven fault diagnosis methods based on Supervisory Control and Data Acquisition (SCADA) systems have gradually become a research hotspot. SCADA systems can continuously collect operating parameters such as wind speed, rotational speed, current, voltage, power, and temperature over long periods, providing abundant resources for fault diagnosis. Based on this, deep learning methods have been widely introduced to automatically identify and classify the operating status of wind turbine units by constructing convolutional neural networks, recurrent neural networks, or time-series graphical models. These methods alleviate the limitations of manual feature extraction to some extent, enabling models to directly learn potential patterns from large amounts of historical data.

[0005] However, existing research mainly focuses on modeling the temporal and spatial characteristics of SCADA data, that is, achieving fault diagnosis by capturing the time series patterns of signals or the spatial correlation between different sensor features. Although this type of method has made significant progress, it still has shortcomings: "space-time" modeling that relies solely on statistical correlation easily ignores the complex physical coupling relationships within the wind turbine and lacks integration with mechanistic constraints, resulting in insufficient physical interpretability of the diagnostic results; under complex operating conditions and multi-source interference, purely data-driven models are prone to insufficient generalization ability and misjudgment; some models have complex structures and a large number of parameters, which are highly dependent on the number of samples and computing resources, leading to difficulties in deployment and real-time performance in actual wind farm applications. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a wind turbine fault diagnosis method and system based on mechanism data fusion. It introduces a power mechanism model into the deep learning model and calculates residual and differential residuals with the actual monitored power in the form of residual signals. This not only enables data-driven feature learning but also combines mechanism constraints for anomaly detection. During feature modeling, a relationship graph between features is constructed through dynamic graph feature interaction, achieving adaptive modeling of different feature importance and reducing misjudgments caused by environmental noise and operating condition fluctuations.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a wind turbine fault diagnosis method based on mechanism data fusion.

[0009] A wind turbine fault diagnosis method based on mechanism data fusion includes the following process:

[0010] The operating condition data of the wind turbine is preprocessed, including sliding window processing and standardization processing.

[0011] Temporal features are extracted from the preprocessed feature sequences and obtained through a pre-trained neural network model.

[0012] A power mechanism model is constructed using the preprocessed data. The theoretical power is determined based on the power mechanism model, and the residuals and difference residuals between the theoretical power and the actual power are calculated.

[0013] The feature sequence, temporal features, residuals, and difference residuals are concatenated and fused to obtain a fused feature vector;

[0014] The fused feature vectors are linearly mapped to a high-dimensional latent space. The features in the high-dimensional latent space are then subjected to matrix multiplication to obtain the similarity distribution between features. After normalization, the adaptive adjacency matrix is ​​obtained.

[0015] The adaptive adjacency matrix and the features in the high-dimensional latent space are multiplied again by matrix multiplication. After normalization, the coupled feature matrix is ​​obtained. The adaptive adjacency matrix and the coupled feature matrix are added to obtain the interactive feature vector. The wind turbine fault diagnosis result is obtained based on the interactive feature vector.

[0016] Secondly, the present invention provides a wind turbine fault diagnosis system based on mechanism data fusion.

[0017] A wind turbine fault diagnosis system based on mechanism data fusion includes:

[0018] The data preprocessing unit is configured to preprocess the wind turbine's operating condition data, including sliding window processing and standardization processing.

[0019] The temporal feature extraction unit is configured to: extract temporal features from the preprocessed feature sequence and obtain temporal features through a pre-trained neural network model;

[0020] The mechanism data processing unit is configured to: construct a power mechanism model using the preprocessed data, determine the theoretical power based on the power mechanism model, and calculate the residual and difference residual between the theoretical power and the actual power;

[0021] The multidimensional feature fusion unit is configured to concatenate and fuse feature sequences, temporal features, residuals, and difference residuals to obtain a fused feature vector.

[0022] The adjacency matrix generation unit is configured to: linearly map the fused feature vector to a high-dimensional latent space, perform matrix multiplication operations between features in the high-dimensional latent space to obtain the similarity distribution between features, and obtain an adaptive adjacency matrix after normalization.

[0023] The diagnostic result generation unit is configured to: perform matrix multiplication between the adaptive adjacency matrix and the features in the high-dimensional latent space again, and obtain a coupled feature matrix after normalization; add the adaptive adjacency matrix and the coupled feature matrix to obtain an interactive feature vector; and obtain the wind turbine fault diagnosis result based on the interactive feature vector.

[0024] Thirdly, the present invention provides a computer device, comprising: a processor and a computer-readable storage medium;

[0025] A processor, adapted to execute computer programs;

[0026] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the wind turbine fault diagnosis method based on mechanism data fusion, which is the first aspect of the present invention.

[0027] Fourthly, the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed by the wind turbine fault diagnosis method based on mechanism data fusion of the first aspect of the present invention.

[0028] Fifthly, the present invention provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the wind turbine fault diagnosis method based on mechanism data fusion of the first aspect of the present invention.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] This invention introduces a power mechanism model and calculates residuals and differential residuals with the actual monitored power, effectively combining data-driven feature learning with mechanism-constrained anomaly detection, thus solving the problem of insufficient physical interpretability in traditional methods. It utilizes dynamic graph feature interaction to construct a feature relationship graph, achieving adaptive modeling of different feature importance, greatly reducing misjudgments caused by environmental noise and operating condition fluctuations, and overcoming the shortcomings of previous methods in terms of insufficient generalization ability under varying operating conditions and noise interference. By adopting a lightweight design strategy, it significantly reduces training and inference time while ensuring diagnostic accuracy, solving the problem of existing models being too large and difficult to meet the real-time application requirements of wind farms. Overall, it improves the stability, robustness, and real-time performance of wind turbine fault diagnosis.

[0031] To address the problem that existing deep learning methods, which mostly rely on spatiotemporal modeling based on statistical features, struggle to depict the complex dynamic mechanisms within wind turbines, this invention introduces a power mechanism model into the network structure. It calculates residuals and differential residuals between theoretical and measured power, forming physically constrained information inputs to ensure that diagnostic results reflect the true mechanistic characteristics. To address the issue that purely data-driven models exhibit insufficient generalization ability and are prone to misjudgment under varying operating conditions and noise interference, this invention utilizes LSTM (Long Short-Term Memory) neural networks to model temporal evolution patterns and combines it with DGI to construct a dynamic relationship graph between features. This adaptively highlights key features and suppresses redundant information, thereby improving the stability and robustness of the diagnosis. Finally, to address the problem that existing models are bulky, time-consuming in training and inference, and unable to meet the real-time application requirements of wind farms, the proposed PIA (Physics-Integrated Analysis Network) network adopts a lightweight design strategy, maintaining the shallow structure of LSTM while using DGI (Dynamic Graph) to achieve the desired effect. Interaction (dynamic graph feature interaction) replaces complex attention mechanisms, significantly reducing training and inference time while ensuring diagnostic accuracy.

[0032] The operating condition data of this invention comes from the SCADA system and covers key parameters such as wind speed and rotational speed, ensuring the comprehensiveness and real-time nature of the data source. This provides a rich and accurate information foundation for fault diagnosis. By utilizing this multi-dimensional, high-quality data, the changes in the operating status of wind turbines can be captured more accurately, thereby improving the sensitivity and specificity of fault diagnosis and effectively solving the problems of misdiagnosis and missed diagnosis caused by incomplete or inaccurate data. The specific form of the power mechanism model is clarified, including parameters such as cut-in wind speed and full-load wind speed. The output power characteristics of wind turbines are described through a precise physical model, so that fault diagnosis not only relies on data-driven pattern recognition, but also combines physical laws for in-depth analysis, enhancing the physical interpretability of the diagnostic results and improving the accuracy of fault identification under complex operating conditions.

[0033] The theoretical power calculation method of this invention takes into account the influence of temperature. By introducing the environmental temperature variable, the power model is made closer to the actual operating conditions, improving the accuracy of theoretical power calculation. This helps to more accurately capture power fluctuations caused by environmental changes, thereby improving the sensitivity and reliability of fault diagnosis. This invention defines the calculation methods of residual and differential residual. By quantifying the deviation between theoretical power and actual power, it provides an intuitive and effective anomaly detection index for fault diagnosis, which helps to quickly locate potential faults, reduce misjudgments and missed judgments, and improve diagnostic efficiency.

[0034] This invention describes the process of obtaining wind turbine fault diagnosis results based on interactive feature vectors. By integrating multi-dimensional feature information and using linear layers and activation modules for in-depth analysis, it achieves accurate assessment of the health status of wind turbine units. This not only improves the accuracy of fault diagnosis but also enhances the robustness of the model, enabling it to maintain stable diagnostic performance in complex and ever-changing operating environments.

[0035] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0036] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0037] Figure 1 An architecture diagram of a mechanism fusion analysis network provided as an exemplary embodiment of the present invention;

[0038] Figure 2 A power over-limit diagram of the gearbox oil temperature control system provided as an exemplary embodiment of the present invention is shown, wherein... Figure 2 (A) in the diagram is a schematic diagram showing the relationship between active power and ambient wind speed. Figure 2 (B) in the diagram is a schematic diagram showing the relationship between the power difference and the ambient wind speed;

[0039] Figure 3 A schematic diagram illustrating an excessively low average power at short-time blade pitch rate, provided as an exemplary embodiment of the present invention, wherein... Figure 3 (A) in the diagram is a schematic diagram showing the relationship between active power and ambient wind speed. Figure 3 (B) in the diagram is a schematic diagram showing the relationship between the power difference and the ambient wind speed;

[0040] Figure 4 A power diagram illustrating a mismatch between generator speed and impeller speed, provided as an exemplary embodiment of the present invention, wherein... Figure 4 (A) in the diagram is a schematic diagram showing the relationship between active power and ambient wind speed. Figure 4 (B) in the diagram is a schematic diagram showing the relationship between the power difference and the ambient wind speed;

[0041] Figure 5 A power diagram of a cabin vibration fault is provided as an exemplary embodiment of the present invention, wherein, Figure 5 (A) in the diagram is a schematic diagram showing the relationship between active power and ambient wind speed. Figure 5 (B) in the diagram is a schematic diagram showing the relationship between the power difference and the ambient wind speed;

[0042] Figure 6 A schematic diagram illustrating the principle of dynamic graph feature interaction provided as an exemplary embodiment of the present invention;

[0043] Figure 7 A schematic diagram of the loss curve of the comparative model provided as an exemplary embodiment of the present invention;

[0044] Figure 8 A confusion matrix illustration provided for an exemplary embodiment of the present invention Figure 1 ,in, Figure 8 (A) in the diagram is a schematic of the confusion matrix of the XGBoost model; Figure 8 (B) in the diagram is a schematic diagram of the confusion matrix of the Light GBM model; Figure 8 (C) in the diagram is a schematic representation of the confusion matrix of the HGB model; Figure 8 (D) in the diagram is a schematic representation of the confusion matrix of the CNN model; Figure 8 (E) in the diagram is a schematic representation of the confusion matrix of the CNN-LSTM model; Figure 8 (F) in the diagram is a schematic representation of the confusion matrix of the TCN model;

[0045] Figure 9 A confusion matrix illustration provided for an exemplary embodiment of the present invention Figure 2,in, Figure 9 (A) in the diagram is a schematic diagram of the confusion matrix of the RNN model; Figure 9 (B) in the diagram is a schematic diagram of the confusion matrix of the GCN model; Figure 9 (C) in the diagram is a schematic of the confusion matrix of the EfficientNet model; Figure 9 (D) in the diagram is a schematic representation of the confusion matrix of the Transformer model; Figure 9 (E) in the diagram is a schematic representation of the confusion matrix of the Swin-Transformer model; Figure 9 (F) in the diagram is a schematic representation of the confusion matrix of the NODE model;

[0046] Figure 10 A confusion matrix illustration provided for an exemplary embodiment of the present invention Figure 3 ,in, Figure 10 (A) in the diagram is a schematic diagram of the confusion matrix of the Mamba model; Figure 10 (B) in the diagram is a schematic diagram of the confusion matrix of the ResNet model; Figure 10 (C) in the diagram is a schematic diagram of the confusion matrix of the PIA model;

[0047] Figure 11 A schematic diagram of a wind turbine fault diagnosis system based on mechanism data fusion, provided as an exemplary embodiment of the present invention;

[0048] Figure 12 A schematic diagram of a computer device provided for an exemplary embodiment of the present invention. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0051] This implementation proposes a wind turbine fault diagnosis method based on mechanism-data fusion. A physical mechanism model of segmented wind turbine power is introduced into the deep learning model, and residual and differential residual calculations are performed between the residual signal and the actual monitored power. This allows the model to not only learn features based on data but also detect anomalies by incorporating mechanism constraints. During feature modeling, an LSTM structure is used to extract temporal dynamic features, and a relationship graph between features is constructed through a Dynamic Graph Interaction (DGI) module. This enables adaptive modeling of the importance of different features, thereby reducing misjudgments caused by environmental noise and operating condition fluctuations. To balance accuracy and practical application requirements, a lightweight approach is adopted in the overall model design, proposing a Physics-Integrated Analysis Network (PIA).

[0052] like Figure 1 As shown, the wind turbine fault diagnosis method based on mechanism data fusion of the present invention includes the following process:

[0053] Step S101: In the data processing stage, the data processing of each SCADA sensor is performed using a sliding window process;

[0054] Step S102: In the temporal feature extraction stage, the data after the sliding window is extracted temporally, and temporal features are obtained through the LSTM network;

[0055] Step S103: In the physical constraint construction stage, the mechanism power mechanism model is constructed using the data after sliding windowing, the theoretical power is calculated, and the residual and difference residual between the theoretical power and the actual power are obtained;

[0056] Step S104: In the spatial matrix generation stage, the eigenvalues ​​obtained from the temporal features and physical constraints are fused into vectors for linear mapping and “graphical” interaction to obtain the correlation between each eigenvector, while suppressing redundant correlations.

[0057] In step S101 of the present invention, the data preprocessing includes sliding window construction and normalization transformation, specifically including:

[0058] First, considering the temporal characteristics of the monitored signals, a sliding window mechanism is used for sample partitioning. Let the original signal sequence be as shown in equation (1):

[0059] (1);

[0060] in, For the length of time, For feature dimension, These represent the signal values ​​in the original signal sequence, Represents the first in the original signal sequence One signal value, It represents the real number field.

[0061] By setting the window length and step length The sequence can be divided into several training samples, as shown in equation (2):

[0062] (2);

[0063] in, Indicates the first Samples from one window, Representing the first The, the The, ..., the One window; The running status label at the end of the window. This approach, which uses labels, not only preserves the dynamic evolution of signals within local time periods but also enhances the number of samples and improves the robustness of temporal feature learning by constructing overlapping windows.

[0064] Secondly, considering the differences in dimensions and numerical ranges of features collected by different sensors, this invention employs standardization processing to map each feature to a unified numerical space, for any feature component. Its standardized form is as shown in equation (3):

[0065] (3);

[0066] in, and Features of the training set The mean and standard deviation, The transformation represents the total number of feature components, and the processed features satisfy equations (4) and (5):

[0067] (4);

[0068] (5);

[0069] in, represent Expectations represent The variance.

[0070] During the testing phase, the data obtained during the training phase will still be used. and Normalization is performed to ensure the consistency of training and testing data distribution and avoid information leakage.

[0071] In step S102 of the present invention, after completing the preprocessing of the feature data, the dynamic features of the time series are further modeled to capture the dependence of the wind turbine's operating state between consecutive time points. A Long Short Term Memory (LSTM) network is used as the time series modeling unit to process the samples constructed by the sliding window.

[0072] Specifically, assuming the first window samples If equation (2) is satisfied, then the input corresponding to the LSTM is the length. The feature sequence is used as the feature sequence, and the output is the hidden state sequence at each time step. The hidden state at the final moment Used as a temporal characteristic representation for this window.

[0073] Inside the LSTM unit, the state update is jointly determined by the gating mechanism of equations (6)–(11):

[0074] (6);

[0075] (7);

[0076] (8);

[0077] (9);

[0078] (10);

[0079] (11);

[0080] in, These represent the input gate, forget gate, and output gate, respectively. For the Sigmoid function, This is element-wise multiplication; This represents the input vector at time step; Representative at Hidden state at time step; Representative at Hidden state at time step; Representative at The state of the memory unit at each time step; Representative at The state of the memory unit at each time step; , , , These represent the input weight matrices respectively; , , and These represent the cyclic weight matrix, respectively. , , and These represent the bias vectors respectively; tanh(•) represents the hyperbolic tangent function.

[0081] Through the above structure, LSTM can retain key information over long time spans and forget unimportant historical signals, thereby effectively extracting temporal dependencies in sequences. The final result... As a window Temporal feature representation As shown in equation (12):

[0082] (12);

[0083] The above-mentioned time series characteristics are represented It can comprehensively reflect the dynamic pattern of the input sequence within the window in the time dimension, and provide temporal feature support for subsequent fusion with the mechanism residual signal and spatial interaction module.

[0084] In step S103 of the present invention, the output power is determined according to the operating mechanism of the wind turbine. The mechanism exhibits obvious piecewise characteristics across different wind speed ranges. Specifically, the piecewise function shown in Equation (13) is used to describe the mechanism power mechanism model:

[0085] (13);

[0086] in, The cut-in wind speed, full-load wind speed, and cut-out wind speed are 3 m / s, 10.8 m / s, and 25 m / s, respectively, in this wind turbine. The rated power of the wind turbine is 2050kW; air density; The sweeping area of ​​the wind turbine. Wind speed; The power factor is typically in the range of 0 to 0.59.

[0087] From a physical perspective: when the wind speed is lower than the cut-in wind speed, the airflow kinetic energy is insufficient to drive the impeller to rotate, the unit does not generate electricity, and the power is zero; when the wind speed is higher than the cut-out wind speed, the unit automatically shuts down to protect the equipment, and the power is zero; between the full-load wind speed and the cut-out wind speed, the power is limited to the rated value. Between the cut-off speed and the full-load wind speed, the power increases cubically with the wind speed. This range is the area where the difference between the power in the mechanistic model and the actual power is most sensitive.

[0088] like Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, the power curves of normal and fault conditions are relatively similar in the low wind speed and full-load wind speed ranges. However, between the cut-in wind speed and the full-load wind speed, the power curve of the fault condition is significantly different from that of the normal condition, reflecting the impact of different types of faults on the energy conversion efficiency of wind turbines. This difference is exactly consistent with the core interval of the mechanism segmentation model. Therefore, by modeling the residuals of the mechanism constraints, the fault sensitivity in this interval can be significantly improved.

[0089] More specifically, Figure 2 (A) in the figure shows the change of active power of wind turbine under different ambient wind speeds. It can be seen that between the cut-in wind speed and the full-load wind speed, the power increases significantly with the increase of wind speed, and the power remains stable after reaching the full-load wind speed. Figure 2 (B) shows the difference between actual power and theoretical power as a function of ambient wind speed. In the case of a fault where the power of the gearbox oil temperature control system exceeds the limit, the power difference in certain wind speed ranges is significantly larger, indicating that the fault causes abnormal power output. Figure 3 (A) in the figure shows the normal output of active power of the wind turbine under different wind speed conditions; Figure 3 (B) reflects the change in the difference between actual power and theoretical power when the average pitch rate of the blade is too low in a short time. Under fault conditions, the power difference is particularly significant in the low wind speed range, indicating that the low pitch rate affects power capture. Figure 4 (A) in the figure depicts the trend of the active power of the wind turbine with wind speed under normal operating conditions; Figure 4 (B) shows the change in power difference when the generator speed and impeller speed are mismatched. Under fault conditions, the power difference in multiple wind speed ranges exceeds the normal range, indicating that the speed mismatch seriously affects the power conversion efficiency.

[0090] The output power between the cut-in wind speed and the full-load wind speed is obtained according to equation (13). The formula. Furthermore, where the swept area... From equation (14), we can derive:

[0091] (14);

[0092] in, Where is the radius of the wind turbine.

[0093] air density From equation (15), we can derive:

[0094] (15);

[0095] in, The gas constant is... Kelvin temperature, Atmospheric pressure.

[0096] Atmospheric pressure As altitude changes, as shown in equation (16):

[0097] (16);

[0098] in, This is the standard atmospheric pressure at sea level; It is the temperature lapse rate; Altitude; It is the standard temperature at sea level. It is gravitational acceleration; It is the molar mass of air.

[0099] According to equations (13) to (16), it can be obtained that the wind turbine unit at a fixed position only has temperature. With wind speed The variable is 'one', and the rest are constants. From this, we can obtain the theoretical power of the wind turbine between the cut-in wind speed and the full-load wind speed. As shown in equation (17):

[0100] (17);

[0101] Where, constant It is determined by the unit's structural parameters, altitude, and climate conditions.

[0102] To characterize the difference between theoretical power and actual power, the residual signal is defined in this invention as shown in equation (18):

[0103] (18);

[0104] in, For actual monitored power, This is the theoretical power, which is the output power that the mechanism model should have under normal operation.

[0105] Meanwhile, the differential residuals are further constructed as shown in equation (19) to characterize the sensitivity of the residuals to wind speed and temperature:

[0106] (19);

[0107] Through residuals With difference residual By constructing the structure, this invention can quantify the deviation between theoretical power and measured power. In particular, when entering the full-power wind speed range, energy conversion anomalies caused by different types of faults can be effectively captured. At the same time, the residual and differential residual characteristics are fused with the previous time series characteristics and a spatial matrix is ​​constructed on the DGI module.

[0108] In step S104 of the present invention, after completing the extraction of time-series features and the construction of residual signals, a DGI module is further proposed to model the dynamic coupling relationship between different features. By constructing an adaptive adjacency matrix, cross-feature dimension information fusion is achieved, thereby enhancing the ability to distinguish fault features.

[0109] like Figure 6 As shown, the input to the DGI module includes: raw features (i.e., the preprocessed data feature sequence), residual signal , differential residuals and temporal features extracted by LSTM The above multi-source features are concatenated to obtain a fusion vector, as shown in equation (20):

[0110] (20);

[0111] The concatenated input vector is linearly mapped to the high-dimensional latent space through equation (21):

[0112] (twenty one);

[0113] in, For trainable weight matrix, This is the hidden space dimension.

[0114] In the latent space, matrix multiplication (i.e. dot product similarity calculation) is first performed between features to obtain the similarity distribution between features. Then, after Softmax normalization, an adaptive adjacency matrix A is generated, as shown in Equation (22). This adjacency matrix can reflect the coupling strength between different features and is dynamically adjusted according to the changes in the input samples.

[0115] (twenty two);

[0116] Among them, in the adaptive adjacency matrix A Indicates the first The first feature and the second The adjacency matrix, which assigns correlation weights to each feature, can be dynamically adjusted based on the input samples, exhibiting strong adaptability.

[0117] The adaptive adjacency matrix and the features in the high-dimensional latent space are multiplied again by matrix multiplication. After normalization, the coupled feature matrix is ​​obtained. The adaptive adjacency matrix and the coupled feature matrix are added to obtain the interactive feature vector. The wind turbine fault diagnosis result is obtained based on the interactive feature vector.

[0118] The constructed adjacency matrix is ​​multiplied again with the latent space features, as shown in equation (23):

[0119] (twenty three);

[0120] The result of equation (23) After normalization, a coupled feature matrix is ​​obtained, realizing cross-feature information propagation and aggregation, and enhancing the expression of correlation between features. That is, each feature dimension is composed of weighted combinations of other features related to it, thereby explicitly capturing cross-feature coupling relationships. At the same time, in order to maintain numerical stability and prevent overfitting, this invention introduces residual connections on the basis of aggregated output (that is, adding the adaptive adjacency matrix and the coupled feature matrix to obtain the interaction feature vector). The interaction feature vector is processed by a linear layer and layer normalization, and then processed by multiple identical activation modules in sequence. Finally, it passes through a linear layer to obtain the wind turbine fault diagnosis result. The activation module includes a linear layer and a modified linear unit activation function set in sequence.

[0121] The following is a specific example. The dataset is selected from the SCADA system data of a wind farm. The resolution of the SCADA system data is 1 minute. The characteristic variables of 12 sensors were selected from it, namely ambient temperature, ambient wind speed, active power, gearbox cooling system pump temperature, generator speed, converter grid voltage, first blade pitch motor temperature, second blade pitch motor temperature, third blade pitch motor temperature, impeller speed, impeller speed, and grid-side C-phase current.

[0122] The fault types selected were four types: power over-limit of gearbox oil temperature control system, short-term average low blade pitch rate, mismatch between generator speed and impeller speed, and nacelle vibration fault. The constructed dataset is shown in Table 1.

[0123] Table 1: Dataset Settings

[0124]

[0125] To comprehensively evaluate the effectiveness of the wind turbine fault diagnosis method proposed in this invention, this invention selects multiple macro-average evaluation indicators to comprehensively examine the model's classification ability under different categories. Let the... The number of true positives corresponding to the class is The number of true negatives is The number of false positives is The number of false negatives is The total number of categories is The definitions of each indicator are as shown in equations (24) to (30):

[0126] (twenty four);

[0127] (25);

[0128] (26);

[0129] (27);

[0130] (28);

[0131] (29);

[0132] (30);

[0133] in, It is used to measure the proportion of samples predicted as positive that actually belong to the positive class. Its significance lies in reflecting the reliability of the model's fault identification results; the higher the accuracy, the lower the false alarm rate. This represents the average classification accuracy across all categories, preventing the overall accuracy from being dominated by the major categories. It is used to measure the proportion of correct identifications among all real fault samples, and can reflect the coverage of the model. In wind turbine diagnosis, a high recall rate means fewer false negatives, ensuring the safe operation of the equipment. The geometric mean, which combines recall and specificity, can evaluate the model's ability to balance across different classes. It is the harmonic mean of precision and recall, which takes into account both accuracy and recall, and is suitable for evaluating the overall performance of the model. It reflects the proportion of normal samples that are incorrectly identified as faults. The lower the value, the lower the risk of false alarms in the system. In actual wind farm applications, a low FAR can reduce unnecessary inspection and maintenance costs. It reflects the proportion of fault samples that are misclassified as normal. The lower the value, the lower the risk of false negatives in the system. In wind power operation scenarios, a low MAR means that the model can effectively reduce potential safety hazards.

[0134] To verify the effectiveness of the fault diagnosis method proposed in this invention, the invention was trained and tested on a real wind turbine operation dataset, and compared with various existing deep learning models.

[0135] During the training phase, this invention employs the cross-entropy loss function as the optimization objective and utilizes the Adam optimizer for parameter updates. The learning rate is set to 0.01, the batch size to 64, and the number of training epochs to 100. To ensure temporal dependencies, the input samples are constructed using a sliding window, and scaling across feature dimensions is achieved through standardization. During training, macro-accuracy is used as the optimization criterion, performance is monitored in real-time on the test set, and the optimal model weights are saved for subsequent evaluation.

[0136] To comprehensively evaluate the diagnostic performance of this invention, a variety of representative deep learning and traditional machine learning methods, including Transformer, Swin-Transformer, Mamba, NODE, EfficientNet, ResNet, 1D-CNN, CNN-LSTM, GCN, RNN, TCN, HGB, GBDT-Light, GBM, and XGBoost, were selected as comparison models. These models cover a wide range of mainstream technologies, including convolutional structures, recursive structures, graph neural networks, temporal modeling networks, and ensemble learning methods.

[0137] As shown in Table 2, macro-Accuracy, macro-Precision and macro-Recall In terms of multiple metrics, the PIA model proposed in this invention achieves optimal results. Although its inference speed is not optimal, 2.73µs is sufficient to support real-time detection of wind turbine faults, achieving a lightweight model. Other large models such as Transformer, Swin-Transformer, and ResNet also achieve good results, but their inference speed is an order of magnitude faster than the PIA model, making it difficult to meet the requirements of real-time detection.

[0138] Table 2: Comparison of Fault Diagnosis Models

[0139]

[0140] like Figure 7 The diagram shows a comparison of the loss curves of different models during the training process. The PIA model proposed in this invention exhibits a rapid decrease in loss during the initial stage and converges within 10 iterations, with an overall loss value significantly lower than the comparison models. This indicates that the PIA model possesses more efficient convergence speed and more stable optimization performance in feature extraction and fault mode discrimination.

[0141] also, Figure 8 , Figure 9 and Figure 10The confusion matrix for each model on the test set (representing the relationship between predicted labels and true labels, including labels 0, 1, 2, 3, and 4). Figure 8 , Figure 9 and Figure 10 The subplots show the classification performance of different models on the test set, including the distribution of true positives, false positives, true negatives, and false negatives. By comparing the confusion matrices of each model, the recognition accuracy and misclassification rate of each model on different categories can be seen intuitively.

[0142] The PIA model proposed in this invention exhibits high recognition accuracy across all fault categories. In particular, it effectively reduces the false alarm rate and false negative rate in complex fault categories such as "power over-limit of gearbox oil temperature control system" and "mismatch between generator speed and impeller speed". Compared with other models, PIA is more outstanding in category discrimination, which verifies the significant advantages of introducing physical constraints and dynamic feature interaction mechanism.

[0143] Figure 11 A wind turbine fault diagnosis system based on mechanism data fusion is shown, including:

[0144] The data preprocessing unit is configured to preprocess the wind turbine's operating condition data, including sliding window processing and standardization processing.

[0145] The temporal feature extraction unit is configured to: extract temporal features from the preprocessed feature sequence and obtain temporal features through a pre-trained neural network model;

[0146] The mechanism data processing unit is configured to: construct a power mechanism model using the preprocessed data, determine the theoretical power based on the power mechanism model, and calculate the residual and difference residual between the theoretical power and the actual power;

[0147] The multidimensional feature fusion unit is configured to concatenate and fuse feature sequences, temporal features, residuals, and difference residuals to obtain a fused feature vector.

[0148] The adjacency matrix generation unit is configured to: linearly map the fused feature vector to a high-dimensional latent space, perform matrix multiplication operations between features in the high-dimensional latent space to obtain the similarity distribution between features, and obtain an adaptive adjacency matrix after normalization.

[0149] The diagnostic result generation unit is configured to: perform matrix multiplication between the adaptive adjacency matrix and the features in the high-dimensional latent space again, and obtain a coupled feature matrix after normalization; add the adaptive adjacency matrix and the coupled feature matrix to obtain an interactive feature vector; and obtain the wind turbine fault diagnosis result based on the interactive feature vector.

[0150] It is understood that the aforementioned units can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of the present invention. The aforementioned units are based on logical functional division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of the present invention, the system may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0151] According to another embodiment of the present invention, the system of this embodiment can be constructed by running a computer program (including program code) capable of performing the steps involved in the corresponding method of the present invention on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). The computer program can be recorded on, for example, a computer-readable recording medium, loaded into the aforementioned computing device through the computer-readable recording medium, and run therein.

[0152] Figure 12 A computer device is shown, comprising a processor, a communication interface, and a computer-readable storage medium. The processor, communication interface, and computer-readable storage medium are connected via a bus or other means.

[0153] The communication interface is used to receive and send data. The computer-readable storage medium can be stored in the memory of the electronic device. The computer-readable storage medium is used to store computer programs, which include program instructions. The processor is used to execute the program instructions stored in the computer-readable storage medium.

[0154] A processor is the computing and control core of an electronic device. It is suitable for implementing one or more instructions, specifically for loading and executing one or more instructions to achieve the corresponding method flow or function.

[0155] The processor is configured to perform the following procedure:

[0156] The operating condition data of the wind turbine is preprocessed, including sliding window processing and standardization processing.

[0157] Temporal features are extracted from the preprocessed feature sequences and obtained through a pre-trained neural network model.

[0158] A power mechanism model is constructed using the preprocessed data. The theoretical power is determined based on the power mechanism model, and the residuals and difference residuals between the theoretical power and the actual power are calculated.

[0159] The feature sequence, temporal features, residuals, and difference residuals are concatenated and fused to obtain a fused feature vector;

[0160] The fused feature vectors are linearly mapped to a high-dimensional latent space. The features in the high-dimensional latent space are then subjected to matrix multiplication to obtain the similarity distribution between features. After normalization, the adaptive adjacency matrix is ​​obtained.

[0161] The adaptive adjacency matrix and the features in the high-dimensional latent space are multiplied again by matrix multiplication. After normalization, the coupled feature matrix is ​​obtained. The adaptive adjacency matrix and the coupled feature matrix are added to obtain the interactive feature vector. The wind turbine fault diagnosis result is obtained based on the interactive feature vector.

[0162] This invention also provides a computer-readable storage medium, which is a memory device in an electronic device for storing programs and data. It is understood that the computer-readable storage medium here may include both built-in storage media in the electronic device and extended storage media supported by the electronic device. The computer-readable storage medium provides storage space for storing the processing system of the electronic device.

[0163] Furthermore, this storage space also contains one or more instructions suitable for loading and execution by the processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM memory or unstable memory, such as at least one disk storage device; optionally, it can also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0164] In one embodiment, the computer-readable storage medium stores one or more instructions; the processor loads and executes the one or more instructions stored in the computer-readable storage medium to perform the following process:

[0165] The operating condition data of the wind turbine is preprocessed, including sliding window processing and standardization processing.

[0166] Temporal features are extracted from the preprocessed feature sequences and obtained through a pre-trained neural network model.

[0167] A power mechanism model is constructed using the preprocessed data. The theoretical power is determined based on the power mechanism model, and the residuals and difference residuals between the theoretical power and the actual power are calculated.

[0168] The feature sequence, temporal features, residuals, and difference residuals are concatenated and fused to obtain a fused feature vector;

[0169] The fused feature vectors are linearly mapped to a high-dimensional latent space. The features in the high-dimensional latent space are then subjected to matrix multiplication to obtain the similarity distribution between features. After normalization, the adaptive adjacency matrix is ​​obtained.

[0170] The adaptive adjacency matrix and the features in the high-dimensional latent space are multiplied again by matrix multiplication. After normalization, the coupled feature matrix is ​​obtained. The adaptive adjacency matrix and the coupled feature matrix are added to obtain the interactive feature vector. The wind turbine fault diagnosis result is obtained based on the interactive feature vector.

[0171] The present invention also provides a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the following process:

[0172] The operating condition data of the wind turbine is preprocessed, including sliding window processing and standardization processing.

[0173] Temporal features are extracted from the preprocessed feature sequences and obtained through a pre-trained neural network model.

[0174] A power mechanism model is constructed using the preprocessed data. The theoretical power is determined based on the power mechanism model, and the residuals and difference residuals between the theoretical power and the actual power are calculated.

[0175] The feature sequence, temporal features, residuals, and difference residuals are concatenated and fused to obtain a fused feature vector;

[0176] The fused feature vectors are linearly mapped to a high-dimensional latent space. The features in the high-dimensional latent space are then subjected to matrix multiplication to obtain the similarity distribution between features. After normalization, the adaptive adjacency matrix is ​​obtained.

[0177] The adaptive adjacency matrix and the features in the high-dimensional latent space are multiplied again by matrix multiplication. After normalization, the coupled feature matrix is ​​obtained. The adaptive adjacency matrix and the coupled feature matrix are added to obtain the interactive feature vector. The wind turbine fault diagnosis result is obtained based on the interactive feature vector.

[0178] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can implement the described functions using different methods for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0179] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic cable, digital cable) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0180] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A wind turbine fault diagnosis method based on mechanism data fusion, characterized in that, The method comprises the following steps: The operating condition data of the wind turbine is preprocessed, and the preprocessing includes sliding window processing and standardization processing, wherein the feature sequence of the original signal is represented as: ， wherein, is a time length, is a feature dimension, respectively represent a signal value in the original signal feature sequence, represents the i-th signal value in the original signal feature sequence, represents the i-th signal value in the original signal feature sequence, represents a real number field; The preprocessed feature sequence is time series extracted, and the time series features are obtained through a pre-trained neural network model; A power mechanism model is constructed using the preprocessed data, and the theoretical power is determined according to the power mechanism model, and the residual error and the differential residual error between the theoretical power and the real power are calculated; The power mechanism model is: wherein, Vci, Vr, and Vco are respectively the cut-in wind speed, the full-load wind speed, and the cut-out wind speed, Pn is the rated power of the wind turbine, is the air density, A is the swept area of the wind wheel, is the wind speed, is the power coefficient; and the theoretical power Pth is determined according to the power mechanism model , comprising: wherein, T represents the temperature; The preprocessed feature sequence, the time series features, the residual error and the differential residual error are spliced and fused to obtain a fused feature vector; The fused feature vector is linearly mapped to a high-dimensional hidden space, and the features in the high-dimensional hidden space are subjected to matrix multiplication operation between the features to obtain the similarity distribution between the features, and the adaptive adjacency matrix is obtained after normalization; The adaptive adjacency matrix and the features in the high-dimensional hidden space are subjected to matrix multiplication operation between the features again, and the coupled feature matrix is obtained after normalization, and the interactive feature vector is obtained by adding the adaptive adjacency matrix and the coupled feature matrix, and the wind turbine fault diagnosis result is obtained according to the interactive feature vector, including: The interactive feature vector is subjected to linear layer and layer normalization processing, and then is sequentially subjected to multiple same activation module processing, and then is subjected to linear layer to obtain the wind turbine fault diagnosis result, wherein the activation module comprises a linear layer and a rectified linear unit activation function arranged in sequence.

2. The wind turbine fault diagnosis method based on mechanism data fusion according to claim 1, wherein The operating condition data is the data collected by the data acquisition and monitoring system, including wind speed, rotating speed, current, voltage, power and temperature.

3. The wind turbine fault diagnosis method based on mechanism data fusion according to claim 1, comprising: The residual is: The differential residual is: , is the real power, is the theoretical power, is the residual, is the differential residual.

4. A wind turbine fault diagnosis system based on mechanism data fusion, characterized in that, A data preprocessing unit configured to preprocess the operating condition data of the wind turbine, and the preprocessing includes sliding window processing and standardization processing, wherein the feature sequence of the original signal is represented as: A time series feature extraction unit configured to time series extract the preprocessed feature sequence, and obtain time series features through a pre-trained neural network model; ， wherein, is a time length, is a feature dimension, respectively represent a signal value in the original signal feature sequence, represents the i-th signal value in the original signal feature sequence, represents the i-th signal value in the original signal feature sequence, represents a real number field; A mechanism data processing unit configured to construct a power mechanism model using the preprocessed data, determine the theoretical power according to the power mechanism model, and calculate the residual error and the differential residual error between the theoretical power and the real power; A multi-dimensional feature fusion unit configured to splice and fuse the preprocessed feature sequence, the time series features, the residual error and the differential residual error to obtain a fused feature vector; The power mechanism model is: wherein, Vci, Vr, and Vco are respectively the cut-in wind speed, the full-load wind speed, and the cut-out wind speed, Pn is the rated power of the wind turbine, ρ is the air density, A is the swept area of the wind wheel, V is the wind speed, Cp is the power coefficient; and the theoretical power Pth is determined according to the power mechanism model , comprising: wherein, T represents the temperature; An adjacency matrix generation unit configured to linearly map the fused feature vector to a high-dimensional hidden space, and perform matrix multiplication operation between the features in the high-dimensional hidden space to obtain the similarity distribution between the features, and obtain an adaptive adjacency matrix after normalization; A diagnosis result generation unit configured to perform matrix multiplication operation between the adaptive adjacency matrix and the features in the high-dimensional hidden space again, obtain a coupled feature matrix after normalization, obtain an interactive feature vector by adding the adaptive adjacency matrix and the coupled feature matrix, and obtain the wind turbine fault diagnosis result according to the interactive feature vector, including: ​ The interaction feature vector is processed by a linear layer and a layer normalization, and then is processed by a plurality of same activation modules in sequence, and finally is processed by a linear layer to obtain a wind turbine fault diagnosis result, wherein the activation module comprises a linear layer and a rectified linear unit activation function arranged in sequence.

5. A computer device, comprising: The method comprises: a processor and a computer readable storage medium; a processor adapted to execute a computer program; a computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and the computer program is executed by the processor to implement the wind turbine fault diagnosis method based on mechanism data fusion according to any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is adapted to be loaded and executed by the processor to implement the wind turbine fault diagnosis method based on mechanism data fusion according to any one of claims 1 to 3.

7. A computer program product, characterised in that, The computer program product comprises a computer program, and the computer program is executed by the processor to implement the wind turbine fault diagnosis method based on mechanism data fusion according to any one of claims 1 to 3.

Images