A wind turbine generator equipment fault diagnosis method and system

Abstract

The application discloses a kind of wind turbine equipment fault diagnosis method and system, belong to the technical field of fault diagnosis, extract the excitation signal characteristics of excitation signal;The correlation between the excitation signal and the state signal is obtained, and the state signal that meets the correlation threshold requirement is screened;Extract the state information characteristics that meet the correlation threshold requirement and the state signal characteristics dimension match;State information characteristics and excitation signal characteristics size matching attention weight matrix is generated;After normalizing attention weight matrix, it is multiplied with excitation signal characteristics, and attention result is obtained;The attention result is added with excitation signal characteristics, and fusion characteristics are obtained;Fault diagnosis result is obtained by mapping fusion characteristics.The method dynamically adjusts the state characteristic weight under different excitation, so that the diagnosis model can better adapt to the working condition fluctuation, and improves the stability and reliability of fault diagnosis under non-stationary working condition.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and more specifically to a method and system for diagnosing faults in wind turbine equipment. Background Technology

[0002] Fault diagnosis of wind turbine equipment is a key technology in the industrial age. It not only concerns the reliability and economy of the equipment itself, but also directly affects production safety, energy efficiency, and environmental protection. With the development of sensor technology, artificial intelligence, and edge computing, fault diagnosis is upgrading from "passive response" to "proactive prevention," becoming an important support for enterprises to achieve intelligent and green transformation.

[0003] Currently, mainstream wind turbine equipment fault diagnosis methods primarily rely on the application of deep learning in various modules of the diagnostic process, such as feature extraction, diagnostic classification, or end-to-end methods. Within deep learning networks, experts have leveraged the strengths of SAE, DBN, RNN, and CNN, or integrated them, to achieve excellent results, tailored to the specific characteristics of real-world applications. While the collected and used diagnostic signals have yielded promising results, they invariably represent the response signals of the turbine equipment or components.

[0004] In real-world industrial scenarios, especially in various power generation devices, the operating conditions change constantly due to the dynamic and continuous nature of the excitation signal. However, most existing research on wind turbine equipment fault diagnosis focuses only on excitation signals under one or a few specific operating conditions, resulting in poor fault diagnosis performance under dynamic and continuous excitation. Summary of the Invention

[0005] To address the problems existing in the above-mentioned fields, this invention proposes a method and system for fault diagnosis of wind turbine equipment. By dynamically adjusting the weights of state features under different excitations, the diagnostic model can better adapt to fluctuations in operating conditions, thereby improving the stability and reliability of fault diagnosis under non-stationary operating conditions.

[0006] To address the aforementioned technical problems, this invention discloses a method for diagnosing faults in wind turbine equipment, comprising the following steps: Acquire state signals reflecting the internal physical state information of the wind turbine equipment at different time series and excitation signals reflecting the power source information of the equipment from different channels; Extract the excitation signal features; Obtain the correlation between the excitation signal and the state signal, and filter the state signals that meet the correlation threshold requirements; extract the state information features that match the feature dimensions of the excitation signal and the state signals that meet the correlation threshold requirements. Obtain an attention weight matrix whose size matches the size of the state information features and the stimulus signal features; normalize the attention weight matrix and multiply it with the stimulus signal features to obtain the attention result; sum the attention result with the stimulus signal features to obtain the fused feature. By fusing feature mapping, fault diagnosis results are obtained.

[0007] Preferably, the excitation signal features for extracting the excitation signal are obtained by using a ResNet18 network to extract the excitation signal features corresponding to the excitation signals of different channels.

[0008] Preferably, the step of acquiring the correlation between the excitation signal and the state signal, and filtering the state signals that meet the correlation threshold requirement, specifically includes: Using the Pearson correlation coefficient method, the correlation coefficient between each state signal and the excitation signal of each channel is obtained, and the maximum value is taken as the degree of correlation between the state signal and the excitation signal. The state signal with the highest correlation is selected as the state signal that meets the correlation threshold requirement.

[0009] Preferably, the extraction of state information features that match the feature dimensions of the state signal and the excitation signal, satisfying the correlation threshold requirement, specifically includes: The state signals that meet the correlation threshold requirements are extracted by convolution operation, and cross-channel information fusion and dimensional transformation are performed to extract state information features that match the feature dimensions of the excitation signal. The state information features are then transposed to obtain transposed features, which are used as the extracted state information features that match the feature dimensions of the excitation signal and the state signals that meet the correlation threshold requirements.

[0010] Preferably, the acquisition of the attention weight matrix that matches the size of the state information features with the size of the excitation signal features specifically includes: By performing convolution operations to fuse and transform the transposed features, an attention weight matrix is ​​generated that matches the size of the state information features with that of the excitation signal features.

[0011] Preferably, the step of normalizing the attention weight matrix and multiplying it with the excitation signal features to obtain the attention result specifically includes: The attention weight matrix is ​​normalized using the softmax function to obtain the normalized weight matrix; The normalized weight matrix is ​​unified into a probabilistic form at each time step, and the sum of the weight values ​​at all time steps is 1, thus obtaining the attention matrix of the excitation signal. The attention matrix is ​​multiplied by the features of the excitation signal to obtain the attention result.

[0012] Preferably, the step of fusing feature mapping to obtain fault diagnosis results specifically includes: The fused features are mapped using a classifier; the classifier includes an average pooling layer, a Dropout layer, a fully connected layer, and a softmax function. The average pooling layer downsamples the fused features to obtain downsampled features. The Dropout layer regularizes the downsampled features to obtain regularized features; The fully connected layer transforms the regularized features to obtain the weight of each fault category; The softmax function converts the weight of each fault category into a corresponding probability value, and uses the probability value as the fault diagnosis result.

[0013] Preferably, the acquisition of state signals reflecting the internal physical state information of the wind turbine equipment at different time series and excitation signals reflecting the power source information of the equipment from different channels specifically includes: The acquired status signals and excitation signals are subjected to data filtering processing, including filtering out excitation signals that are insufficient to drive the equipment operating conditions, equipment operating conditions that have not reached normal operating conditions, and status signals that never change. The acquired status signals include temperature, current, and voltage signals of the wind turbine equipment; the acquired excitation signals include wind speed and wind direction, which describe the power source information of the wind turbine equipment.

[0014] Preferably, it further includes a wind turbine equipment fault diagnosis system, comprising: The signal acquisition module is used to acquire state signals reflecting the internal physical state information of the wind turbine equipment at different time series and excitation signals reflecting the power source information of the equipment from different channels. The excitation signal feature extraction module is used to extract the excitation signal features of the excitation signal; The state information feature extraction module is used to obtain the correlation between the excitation signal and the state signal, filter the state signals that meet the correlation threshold requirements, and extract the state information features that match the feature dimensions of the excitation signal and the state signals that meet the correlation threshold requirements. An adaptive adjustment module is used to obtain an attention weight matrix that matches the size of the state information features and the excitation signal features; after normalizing the attention weight matrix, it is multiplied with the excitation signal features to obtain the attention result; the attention result is summed with the excitation signal features to obtain the fused features. The fault diagnosis module is used to map the fused features and obtain fault diagnosis results.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The wind turbine equipment fault diagnosis method proposed in this invention recognizes that the operating state of a wind turbine is often the result of the combined effect of input excitation and internal response. This method not only acquires state signals characterizing the internal physical state of the equipment but also introduces excitation signals characterizing power source information. This allows for a more complete depiction of the equipment's overall operation, avoiding misdiagnosis or missed diagnosis due to the indistinct features of a single signal, thus improving the comprehensiveness and accuracy of the diagnosis. Correlation screening effectively eliminates redundant data or noise signals unrelated to power source changes, retaining key state signals highly correlated with the current operating conditions. This reduces the dimensionality of data processing and improves the targeting and signal-to-noise ratio of subsequent feature extraction. The attention mechanism adaptively assigns higher weights to important regions in the excitation signal based on the feature distribution of state information, enabling the model to focus on the feature frequency bands or time periods most sensitive to fault diagnosis. The additive residual connection method integrates deep features of attention modulation while retaining the basic information of the original excitation signal, preventing information loss during feature extraction and enhancing the robustness of the model. This method dynamically adjusts the weights of state features under different stimuli through dimension matching and attention mechanisms, enabling the diagnostic model to better adapt to fluctuations in operating conditions and improving the stability and reliability of fault diagnosis under non-stationary operating conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart of the wind turbine equipment fault diagnosis method proposed in this invention; Figure 2 This is the network architecture of the network model for fault diagnosis of wind turbine equipment provided in the embodiments of the present invention; Figure 3 The network architecture of the state attention module in the network model provided in the embodiments of the present invention. Detailed Implementation

[0017] The following will refer to the appendices in the embodiments of the present invention. Figures 1-3 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0018] Example like Figure 1 As shown, this invention proposes a method for diagnosing faults in wind turbine equipment, comprising the following steps: S1: Acquire state signals reflecting the internal physical state information of the wind turbine equipment at different time series and excitation signals reflecting the power source information of the equipment from different channels; S2: Extract the excitation signal features; S3: Obtain the correlation between the excitation signal and the state signal, and filter the state signals that meet the correlation threshold requirements; extract the state information features that match the feature dimensions of the excitation signal and the state signal that meet the correlation threshold requirements. S4: Obtain the attention weight matrix that matches the size of the state information features and the stimulus signal features; normalize the attention weight matrix and multiply it with the stimulus signal features to obtain the attention result; sum the attention result with the stimulus signal features to obtain the fusion feature; S5: Obtain fault diagnosis results by fusing feature mapping.

[0019] Specifically, in step S1, according to the state-space model of modern control theory, the state-space representation of the wind turbine equipment under the action of dynamic excitation signal is as shown in equation (1).

[0020] (1) In the formula, These are the state variables of the wind turbine equipment. This is the external input for wind turbine equipment. The system output of the wind turbine equipment is the state variable, which is a quantitative reflection of the state of the wind turbine equipment itself and changes continuously under the influence of external inputs (i.e., excitation signals). The system output of the wind turbine equipment is determined by the current state variables and the external inputs.

[0021] In fault diagnosis scenarios, the power source of wind turbine equipment is its excitation signal, such as the wind force acting on the turbine, the power supply driving the machine tool, and the water flow velocity driving the turbine. The state variables (i.e., state signals) of the wind turbine equipment are relatively abundant, such as temperature and voltage signals at various points. Under the action of dynamic excitation signals, the excitation changes frequently, leading to continuous changes in the system's state and operating conditions. To accurately predict the system output of the wind turbine equipment, i.e., to investigate whether a fault has occurred, this invention uses both the excitation signal and state signals as inputs to the fault diagnosis network model, and uses different modules to extract corresponding features.

[0022] Therefore, this invention acquires state signals from different time series and excitation signals from different channels of wind turbine equipment. Since the acquired data from the wind turbine equipment may contain invalid or default values, data filtering is required to extract valid data suitable for training and testing, which serves as preprocessed data. The data to be filtered includes, but is not limited to, excitation signals insufficient to drive the equipment, equipment operating conditions not reaching normal operating conditions, and the exclusion of state signals that remain unchanged.

[0023] The preprocessed data is normalized for each channel to avoid the influence of different units on the diagnostic results, and then the normalized data is divided.

[0024] First, the normalized data is divided into multiple training data sets using a sliding window with a fixed window size and step size, and each data set is labeled. These are then randomly divided into training and testing sets according to a certain ratio for subsequent model training and testing.

[0025] The operating environment of this embodiment is the server of the wind farm monitoring center, which receives the sensor data of the wind turbine units collected by the SCADA system, and then performs fault diagnosis on the wind turbine equipment.

[0026] Data from wind turbines is collected using the existing SCADA system, including wind speed, temperature signals at various points, and voltage signals at various points, from a total of 35 channels. The data acquisition frequency is determined by the sampling frequency of the SCADA system, which is usually low in actual use, set to once every 10 minutes, continuously collecting wind turbine operating data for one year. Fault categories are defined as binary categories, i.e., whether a fault has occurred. The dataset is labeled according to the time interval of fault occurrence to establish a fault sample library.

[0027] Valid data for wind turbines is collected when the turbines are driven by sufficient wind and are in normal operating condition. During data filtering, conditions for removal include default values, invalid values, and values ​​where the hub speed is below normal operating range. Additionally, there are instances of turbine shutdown for maintenance, which are filtered based on generator power. After data preprocessing, valid data is segmented using a sliding window. A window length of L = 256 data points and a sliding step size of S = 128 data points are set, resulting in 11968 data entries, which are randomly divided into training and testing sets at a 4:1 ratio. Wind direction, wind speed, and ambient temperature are selected as excitation signals, corresponding to three excitation signal channels. A state signal selection module processes the state signals, selecting the four state signals with the strongest correlation to the excitation signals: nacelle temperature, main bearing temperature, hub temperature, and generator stator temperature, corresponding to four state signal channels. At this point, the dataset for training and testing the equipment fault diagnosis model is complete.

[0028] The method proposed in this invention optimizes computational efficiency by improving the processing flow of input data (such as sliding window segmentation and normalization), enabling it to quickly process real-time data streams collected by SCADA systems and meet the timeliness requirements of online fault early warning in industrial sites.

[0029] In step S2, the present invention uses a ResNet18 network to extract the excitation signal features corresponding to the excitation signals of different channels.

[0030] Because the state signals of a single object can have many types and sources, not every type is closely related to changes in the excitation signal or the occurrence of faults. Many state signals change little or even remain constant during equipment operation. Only by finding the state signals that are strongly correlated with the excitation signal can effective information be extracted from the input data, thereby achieving accurate fault diagnosis and analysis. The Pearson correlation coefficient is one of the most direct methods for measuring the overall linear correlation between two continuous signals. This index can eliminate the influence of the dimensions of different variables on the correlation judgment, resulting in a unified measurement standard.

[0031] This invention uses the Pearson correlation coefficient method to determine the correlation coefficient between each state signal in the training set and the excitation signal of each channel; takes the maximum value of the correlation coefficient corresponding to each channel as the degree of correlation between the state signal and the excitation signal in the training set, and selects the state signal corresponding to the highest degree of correlation as the input of the first convolutional layer in the state attention module.

[0032] The formula for calculating the Pearson correlation coefficient is shown in equation (2).

[0033] (2) In the formula, The correlation coefficient, , For the first of two variables One sample point, , This represents the average of the two variables across all sample points.

[0034] The method proposed in this invention uses the Pearson correlation coefficient to determine the correlation between the excitation signal and the state signal, and then selects the most correlated signals for subsequent feature extraction and fault diagnosis processes. Since the excitation signal has multiple channels, the correlation coefficient for each state signal is calculated for each channel of the excitation signal, and the maximum value is taken as the degree of correlation between the state signal and the excitation signal. Thus, the state signal with the highest correlation is selected.

[0035] This invention uses the Pearson correlation coefficient method to obtain the correlation coefficient of each state signal with the excitation signal of each channel, and takes the maximum value as the degree of correlation between the state signal and the excitation signal. The state signal with the highest correlation is selected as the state signal that meets the correlation threshold requirement.

[0036] In step S3, state signals that meet the correlation threshold requirements are extracted through convolution operations, and cross-channel information fusion and dimensional transformation are performed to extract state information features that match the feature dimensions of the excitation signal. The state information features are then transposed to obtain transposed features, which are used as the extracted state information features that match the feature dimensions of the excitation signal and the state signals that meet the correlation threshold requirements.

[0037] In step S4, the transposed features are fused and transformed again through convolution to generate an attention weight matrix that matches the size of the state information features and the excitation signal features.

[0038] To further improve the fault diagnosis accuracy of the method proposed in this invention, such as Figure 2 The diagram shows the network model constructed in this invention, which includes a feature extraction module, a state attention module, and a classifier, wherein: ResNet is one of the most widely used network models in convolutional neural networks. Its unique residual connection structure connects the beginning and end of each bottleneck block, which can solve the degradation problem during model training and improve classification accuracy. In fault diagnosis scenarios, the number of data points for each excitation signal sample is small, usually ranging from hundreds to thousands. Using a network model with too many layers is likely to lead to overfitting and reduce model performance.

[0039] Therefore, this invention uses ResNet18 as the backbone network for feature extraction modules to extract features of excitation signals, and initially extracts effective information from state signals.

[0040] In the network model constructed in this invention, the feature extraction module uses a ResNet18 network as the backbone. The ResNet18 network employs one-dimensional convolutional layers and uses 1×1 convolutional kernels in the residual connections to perform convolution operations to extract excitation signal features. The size of these excitation signal features is (B... Cu,Lu).

[0041] This invention adds convolutional computation to the residual connections of the ResNet network, further improving the feature extraction capability of the model. In the improved ResNet network, the input and output relationship of each bottleneck layer is shown in Equation (3).

[0042] (3) In the formula, As input features, The output features after passing through the bottleneck layer, The residual learned from the bottleneck layer serves as the feature of the output excitation signal. This is the convolution kernel. It is processed before the input is directly connected to the output. Convolution calculations are performed to obtain the weights of the residual connections through training.

[0043] While external attention can acquire attention to features through training, reducing its dependence on the features themselves, it remains fixed during application and does not change with device state. This means that external attention cannot keenly detect changes in device state and thus adjust the weighting of features.

[0044] The state attention module proposed in this invention integrates the feature information of the state signal on the basis of external attention, realizes dynamic weight allocation of the excitation signal features, and effectively combines the excitation signal and the state signal.

[0045] like Figure 3 The state attention module shown includes a first convolutional layer, a second convolutional layer, and a softmax function connected in sequence; The number of input channels of the first convolutional layer is determined to be The number of output channels is The kernel size is 1×1; the size is... The state signal undergoes cross-channel information fusion and dimensionality transformation through the first convolutional layer, extracting state information features from the input signal whose feature dimensions match those of the excitation signal. The size of this state information feature is (B). L u ,L x ); For the output size (B) L u ,L x The state information feature is transposed to obtain the transposed feature, and the size becomes (B). L x ,L u ).

[0046] The number of input channels for the second convolutional layer is determined to be The number of output channels is The convolutional kernel size is 1×1; the transposed features are fused and transformed through a second convolutional layer to generate an attention weight matrix whose size matches the size of the state information features and the excitation signal features. The size of this attention weight matrix is ​​(B L u ,L u ).

[0047] Step 4 also includes placing the dimension (B) L u ,L uThe attention weight matrix is ​​normalized using the softmax function to obtain a normalized weight matrix. The weight values ​​at each time step are then standardized to probabilistic form, with the sum of the weight values ​​at all time steps being 1. This yields the attention matrix for the excitation signal, with a size of (B). Lu,Lu).

[0048] The state attention module proposed in this invention is implemented based on the attention mechanism theory, and its calculation formula is shown in equation (4): (4) In the formula, The result is the calculation result of the state attention module. This is the original state signal. , This corresponds to the first and second convolutional layers, both of which are one-dimensional; These are the characteristics of the excitation signal.

[0049] The obtained attention matrix is ​​multiplied by the excitation signal features to obtain the attention result. Since the attention result has the same feature size as the feature extracted by the backbone network, the two can be directly summed for processing and analysis in the next module. Therefore, the attention result is summed with the excitation signal features to obtain the fused feature.

[0050] The classifier in the network model of this invention includes an average pooling layer, a Dropout layer, a fully connected layer, and a softmax function: The fused features are input into the average pooling layer for downsampling to obtain downsampled features.

[0051] The downsampled features are regularized using a Dropout layer, and the regularized features are then input into a fully connected layer to obtain the weight of each fault category.

[0052] The softmax function is used to convert the weight of each fault category into a corresponding probability value, and the probability value is used as the fault diagnosis result.

[0053] The network model constructed in this invention uses ResNet18 as the backbone network, avoiding the overfitting problem that may be caused by deep networks, while reducing the number of model parameters. Combined with the lightweight design of the state attention module, the model reduces the computational resource requirements while maintaining high performance.

[0054] The method proposed in this invention introduces joint analysis of excitation signals and state signals, enabling the model to adapt to fault diagnosis needs under different working conditions and avoiding the dependence of traditional methods on specific working conditions.

[0055] The fault diagnosis method proposed in this invention belongs to the classification task in machine learning, and it is necessary to use classification-related performance indicators to measure the training effect of the model.

[0056] This invention uses accuracy, precision, recall and F1 score as performance evaluation indicators, and their calculation formulas are shown in equations (5)-(8).

[0057] (5) (6) (7) (8) In the formula, TP, FP, TN, and FN represent the number of true positives, false positives, true negatives, and false negatives, respectively.

[0058] Classification cross-entropy is a widely used loss function in fault diagnosis and even deep learning classification tasks. It can effectively calculate the loss between the predicted result and the label. This invention will use classification cross-entropy as the loss function for the fault diagnosis model. Simultaneously, Adam is used as the optimizer, which can adaptively update the learning rate based on training progress and improve training efficiency through bias correction, demonstrating good performance across different scenarios.

[0059] The model parameters of the backbone network of this invention are input as excitation signals, with dimensions of [missing information]. Its design is inspired by the ResNet18 network. In the first layer, a 7x7 convolutional kernel is used to process the raw signal, and the number of channels is converted to 64 to facilitate further feature extraction. After each convolution calculation in the backbone network, batch normalization is performed, and then the signal is processed by the ReLU activation function.

[0060] The parameters in the state attention module are set according to the actual feature size. In this embodiment, the data size is set based on the backbone network and the dataset, and the excitation signal features input to the module are... , The status signals are 256 and 16 respectively. , The output sizes are 4 and 256 respectively, and the output size is consistent with the backbone network. In the classifier module, the output size of the average pooling layer is set to 1, the Dropout size is initially set to 0.5, and the number of output channels of the fully connected layer is consistent with the number of fault categories, set to 2.

[0061] After building the network model, it was trained using a dataset. The batch size was set to 64, the number of training epochs to 100, and the initial learning rate to 0.001. The loss function and optimizer remained consistent with the core approach. After each training epoch, the model performance was evaluated using a test set to visually understand how the model changed during training. By adjusting the training parameters, multiple training runs were performed to achieve optimal model performance.

[0062] After training the network model using the training set, the performance of the trained network model is tested using classification evaluation metrics.

[0063] Table 1 Classification performance of each model As shown in Table 1, the classification performance of each model is superior to that of mainstream equipment fault diagnosis models in all classification metrics. Furthermore, the performance is better than that of ResNet18 using only the backbone network, demonstrating the effectiveness and accuracy of the proposed method.

[0064] This invention improves equipment fault diagnosis methods by introducing modern control theory into deep learning methods, significantly enhancing the accuracy of equipment fault diagnosis under dynamic excitation, and further revealing the mechanism of equipment fault occurrence and the law of excitation effect on equipment.

[0065] The network model for fault diagnosis constructed in this invention has a small number of parameters and requires less sample data. For real-time SCADA data streams, it can quickly obtain diagnostic results, meet the timeliness requirements of online real-time early warning of equipment faults in industrial sites, help to discover potential faults in a timely manner, reduce maintenance and monitoring costs, and avoid catastrophic downtime accidents.

[0066] This invention also proposes a fault diagnosis system for wind turbine equipment, comprising: The signal acquisition module is used to acquire state signals reflecting the internal physical state information of the wind turbine equipment at different time series and excitation signals reflecting the power source information of the equipment from different channels. The excitation signal feature extraction module is used to extract the excitation signal features of the excitation signal; The state information feature extraction module is used to obtain the correlation between the excitation signal and the state signal, filter the state signals that meet the correlation threshold requirements, and extract the state information features that match the feature dimensions of the excitation signal and the state signals that meet the correlation threshold requirements. An adaptive adjustment module is used to obtain an attention weight matrix that matches the size of the state information features and the excitation signal features; after normalizing the attention weight matrix, it is multiplied with the excitation signal features to obtain the attention result; the attention result is summed with the excitation signal features to obtain the fused features. The fault diagnosis module is used to map the fused features and obtain fault diagnosis results.

[0067] This method acquires state signals (such as temperature and voltage) from different time series of wind turbine equipment and excitation signals (such as wind speed and power) from different channels simultaneously. The model can comprehensively capture the operating status of the equipment and reduce misjudgments or omissions caused by a single signal source.

[0068] The constructed network model dynamically generates an attention weight matrix that matches the feature size of the excitation signal through two sequentially connected convolutional layers and a softmax function in the state attention module. This allows the model to adaptively adjust the contribution of different state signals to fault diagnosis. This dynamic weight allocation mechanism enables the model to maintain stable performance under complex operating conditions (such as sudden wind speed changes and load fluctuations). This mechanism allows the model to focus on state information features strongly correlated with the fault, suppressing irrelevant noise and thus improving the accuracy of fault classification. For example, under the action of dynamic excitation signals, the model can accurately capture the causal relationship between equipment state changes and fault occurrence. The collaborative design of the feature extraction module (such as ResNet18) and the state attention module allows the model to extract both deep features of the excitation signal and dynamic information of the state signal. This structure effectively solves the problem of insufficient feature extraction and state correlation in traditional methods.

[0069] In summary, this method significantly improves the accuracy, robustness, and real-time performance of wind turbine equipment fault diagnosis through multi-source signal fusion, dynamic attention mechanism, and lightweight network model design, and has broad application prospects.

[0070] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0071] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for diagnosing faults in wind turbine equipment, characterized in that, Includes the following steps: Acquire state signals reflecting the internal physical state information of the wind turbine equipment at different time series and excitation signals reflecting the power source information of the equipment from different channels; Extract the excitation signal features; Obtain the correlation between the excitation signal and the state signal, and filter the state signals that meet the correlation threshold requirements; extract the state information features that match the feature dimensions of the excitation signal and the state signals that meet the correlation threshold requirements. Obtain an attention weight matrix whose size matches the size of the state information features and the stimulus signal features; normalize the attention weight matrix and multiply it with the stimulus signal features to obtain the attention result; sum the attention result with the stimulus signal features to obtain the fused feature. By fusing feature mapping, fault diagnosis results are obtained.

2. The wind turbine equipment fault diagnosis method according to claim 1, characterized in that, The excitation signal features extracted are obtained by using a ResNet18 network to extract the excitation signal features corresponding to the excitation signals of different channels.

3. The wind turbine equipment fault diagnosis method according to claim 1, characterized in that, The process of acquiring the correlation between the excitation signal and the state signal, and filtering the state signals that meet the correlation threshold requirements, specifically includes: Using the Pearson correlation coefficient method, the correlation coefficient between each state signal and the excitation signal of each channel is obtained, and the maximum value is taken as the degree of correlation between the state signal and the excitation signal. The state signal with the highest correlation is selected as the state signal that meets the correlation threshold requirement.

4. The wind turbine equipment fault diagnosis method according to claim 1, characterized in that, The extraction of state information features that match the feature dimensions of the excitation signal and satisfy the correlation threshold requirement specifically includes: The state signals that meet the correlation threshold requirements are extracted by convolution operation, and cross-channel information fusion and dimensional transformation are performed to extract state information features that match the feature dimensions of the excitation signal. The state information features are then transposed to obtain transposed features, which are used as the extracted state information features that match the feature dimensions of the excitation signal and the state signals that meet the correlation threshold requirements.

5. The wind turbine equipment fault diagnosis method according to claim 4, characterized in that, The attention weight matrix that matches the size of the acquired state information features with the size of the excitation signal features specifically includes: By performing convolution operations to fuse and transform the transposed features, an attention weight matrix is ​​generated that matches the size of the state information features with that of the excitation signal features.

6. The wind turbine equipment fault diagnosis method according to claim 1, characterized in that, The normalization of the attention weight matrix, followed by multiplication with the activation signal features, yields the attention result, specifically including: The attention weight matrix is ​​normalized using the softmax function to obtain the normalized weight matrix; The normalized weight matrix is ​​unified into a probabilistic form at each time step, and the sum of the weight values ​​at all time steps is 1, thus obtaining the attention matrix of the excitation signal. The attention matrix is ​​multiplied by the features of the excitation signal to obtain the attention result.

7. The wind turbine equipment fault diagnosis method according to claim 1, characterized in that, The process of obtaining fault diagnosis results by fusing feature mapping specifically includes: The fused features are mapped using a classifier; the classifier includes an average pooling layer, a Dropout layer, a fully connected layer, and a softmax function. The average pooling layer downsamples the fused features to obtain downsampled features. The Dropout layer regularizes the downsampled features to obtain regularized features; The fully connected layer transforms the regularized features to obtain the weight of each fault category; The softmax function converts the weight of each fault category into a corresponding probability value, and uses the probability value as the fault diagnosis result.

8. The wind turbine equipment fault diagnosis method according to claim 1, characterized in that, The acquisition of state signals reflecting the internal physical state information of the wind turbine equipment at different time series and excitation signals reflecting the power source information of the equipment at different channels specifically includes: The acquired status signals and excitation signals are subjected to data filtering processing, including filtering out excitation signals that are insufficient to drive the equipment operating conditions, equipment operating conditions that have not reached normal operating conditions, and status signals that never change. The acquired status signals include temperature, current, and voltage signals of the wind turbine equipment; the acquired excitation signals include wind speed and wind direction, which describe the power source information of the wind turbine equipment.

9. A fault diagnosis system for wind turbine equipment, characterized in that, include: The signal acquisition module is used to acquire state signals reflecting the internal physical state information of the wind turbine equipment at different time series and excitation signals reflecting the power source information of the equipment from different channels. The excitation signal feature extraction module is used to extract the excitation signal features of the excitation signal; The state information feature extraction module is used to obtain the correlation between the excitation signal and the state signal, filter the state signals that meet the correlation threshold requirements, and extract the state information features that match the feature dimensions of the excitation signal and the state signals that meet the correlation threshold requirements. An adaptive adjustment module is used to obtain an attention weight matrix that matches the size of the state information features and the excitation signal features; after normalizing the attention weight matrix, it is multiplied with the excitation signal features to obtain the attention result; the attention result is summed with the excitation signal features to obtain the fused features. The fault diagnosis module is used to map the fused features and obtain fault diagnosis results.

Images