Abnormality detection and fault diagnosis method and system for wind turbine

Abstract

The application provides an abnormality detection and fault diagnosis method and system for a wind turbine, relates to the technical field of wind turbine abnormality detection, and comprises the following steps: performing continuous wavelet transformation and synchronous compression on a vibration analysis signal to generate a time-frequency diagram of a current sampling time; inputting a normalized working condition vector into a pre-trained feature dynamic weight network to obtain a weight vector corresponding to the current sampling time, and dividing the weight vector into a feature weight vector and a vibration frequency band weight vector according to dimensions; further generating a dynamic feature diagram of the current sampling time; inputting the dynamic feature diagram into a pre-trained main diagnosis model to obtain a corresponding deep feature vector and a fault probability distribution; and performing deep fusion of weighted working condition information and the time-frequency diagram to construct a dynamic feature diagram, so that multi-modal fusion of operating state data and vibration signal features is realized, and the accuracy of fault recognition is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of wind turbine generator abnormality detection technology, specifically to methods and systems for abnormality detection and fault diagnosis of wind turbine generators. Background Technology

[0002] Wind turbine generators typically operate in areas with complex and variable environmental conditions, such as high altitudes, coastal areas, or offshore areas. They are subjected to unsteady operating conditions such as varying wind speeds and loads over long periods. Key components such as transmission chains, gearboxes, and main bearings are highly susceptible to progressive damage and sudden failures. Early fault characteristics often manifest as weak time-varying components in vibration signals, which are easily masked by strong background noise and other non-fault frequency components, posing a significant challenge to fault diagnosis. Existing diagnostic methods usually process operating parameters and vibration signals separately, failing to effectively integrate multi-source information and lacking a mechanism to dynamically adjust feature weights based on real-time operating conditions. This results in insufficient generalization ability and diagnostic accuracy of the model under different operating conditions. In the prior art, patent document CN118934489A proposes a method for anomaly detection and fault diagnosis of wind turbine generator sets based on a combination of data mining and failure mode analysis. This method first analyzes the failure modes and fault propagation mechanisms of the wind turbine generator set's structure through failure mode and effect analysis and fault tree analysis. Then, it uses data mining methods to preprocess SCADA / CMS measured data and wind turbine simulation data to achieve anomaly detection. However, in terms of time-frequency feature extraction, this method relies solely on data mining and statistical analysis for anomaly feature extraction, without employing high-resolution time-frequency transformation technology. This makes it difficult to effectively extract the time-varying features of early, weak faults under strong background noise. This method limits the sensitivity and accuracy of fault diagnosis. Regarding adaptability, while it utilizes the operating parameters of the SCADA system, it lacks a mechanism to dynamically adjust feature weights based on real-time operating conditions. This makes it unable to adapt to the drastic changes in signal characteristics of wind turbine generators under unsteady conditions such as varying wind speeds and loads, thus limiting the model's generalization ability under different operating conditions. In terms of multi-source information fusion, this method combines failure mode analysis with data mining results for inference, but it does not deeply integrate operating status data with vibration signals at the feature level. The two remain in relatively independent processing paths, failing to fully utilize the correlation and complementarity of multimodal information. Therefore, there is an urgent need for anomaly detection and fault diagnosis methods specifically for wind turbine generators.

[0003] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for anomaly detection and fault diagnosis of wind turbine generator sets, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for anomaly detection and fault diagnosis of wind turbine generator sets, comprising the following specific steps: S1: At each sampling time, acquire the operating status data of the wind turbine generator set at that time, and standardize the operating status data into a working condition vector; at the same time, set a local window with the sampling time as the center, and collect the vibration signal waveform data within the local window as the vibration analysis signal at that time. S2: Perform continuous wavelet transform and synchronous compression on the vibration analysis signal to generate a time-frequency diagram at the current sampling time; input the standardized working condition vector into a pre-trained feature dynamic weight network to obtain the weight vector corresponding to the current sampling time; divide the weight vector into feature weight vector and vibration frequency band weight vector according to the dimension. S3: Standardize the time-frequency graph; generate a dynamic feature map for the current sampling time based on the standardized time-frequency graph, feature weight vector, and vibration frequency band weight vector; input the dynamic feature map into the pre-trained main diagnostic model to obtain the corresponding depth feature vector and fault probability distribution; S4: Calculate the prediction confidence level based on the fault probability distribution, and simultaneously input the deep feature vector into the pre-constructed support vector description model to calculate the health baseline distance; determine the anomaly based on the confidence level and the health baseline distance, and trigger an alarm when the warning conditions are met.

[0006] Furthermore, the operating status data includes the wind speed at the hub height of the wind turbine, the active power output at the generator end of the wind turbine, and the rotational speed. The vibration analysis signal is a continuous vibration signal segment centered on the current sampling time and cut out with a preset window length.

[0007] Furthermore, the specific steps for generating the time-frequency diagram include: The vibration analysis signal is inner productd with a scalable and translational Morlet wavelet function, where the scale parameter controls the scaling of the wavelet function and the translation parameter controls the movement of the wavelet function along the time axis. By traversing all scales and all translation positions, the wavelet coefficients of the signal and the wavelet function at different scales and positions are calculated. For the wavelet coefficients at each time-frequency position obtained by continuous wavelet transform, first calculate the partial derivative of the wavelet coefficient with respect to the translation parameter. This partial derivative reflects the rate at which the phase of the wavelet coefficient changes with the translation parameter. Then divide the partial derivative by the wavelet coefficient itself, multiply by the negative imaginary unit, and divide by twice pi to calculate the instantaneous frequency at that time-frequency position. The wavelet coefficients are subjected to synchronous compression transformation as follows: The frequency axis is pre-divided into several discrete frequency bins, each with a center frequency and a preset bandwidth. For each translation position, all discrete scales are traversed to determine whether the estimated instantaneous frequency at that scale falls within the bandwidth of a certain frequency bin. Only the wavelet coefficients at those scales that meet the condition are accumulated. Each wavelet coefficient participating in the accumulation needs to be multiplied by the scale interval corresponding to that scale and the negative cube of that scale for amplitude and scale interval normalization correction. Finally, the accumulation result is divided by the frequency bin bandwidth to complete the energy density normalization, resulting in synchronous compression transformation coefficients. The final output time-frequency graph is a two-dimensional image. Its row direction corresponds to the frequency axis, and the resolution is determined by the number of preset frequency bins. Its column direction corresponds to the time axis, and the resolution is determined by the number of sampling points in the preset window. The value of each pixel in the time-frequency graph is the amplitude of the synchronous compression transform coefficient at the corresponding sampling time and frequency.

[0008] Furthermore, the feature dynamic weight network is based on a lightweight multilayer perceptron, including an input layer and an output layer. The input layer contains several neurons, and the output is processed by the ReLU activation function. The output layer contains multiple neurons and the output value of each neuron is compressed using the Sigmoid activation function. The feature dynamic weight network and the main diagnostic model are obtained through the following joint training method: acquire historical operating status data, vibration signal waveform data and corresponding fault category labels, execute steps S1 to S3 sequentially on the historical operating status data and vibration signal waveform data to generate historical dynamic feature maps corresponding to each historical sampling time, input the historical dynamic feature maps into the main diagnostic model, take the cross-entropy loss between the fault category prediction result output by the main diagnostic model and the fault category label as the optimization objective, and synchronously update all trainable parameters in the feature dynamic weight network and the main diagnostic model through the backpropagation algorithm until the cross-entropy loss converges to the preset value.

[0009] Furthermore, the weight vector is divided into a feature weight vector and a vibration frequency band weight vector according to its dimensions, specifically: The standardized operating condition vector is input into the feature dynamic weight network. First, it undergoes a first-layer linear transformation and is processed by the ReLU activation function to extract the deep latent features of the operating condition vector. Then, it undergoes a second-layer linear transformation and is processed by the Sigmoid activation function to compress the output value to between 0 and 1, generating a weight vector corresponding to the input operating condition vector. This weight vector is then divided by dimension, and the first three elements are used to form the feature weight vector, while the remaining elements form the vibration frequency band weight vector. The total number of elements in the vibration frequency band weight vector is equal to the preset number of frequency band divisions.

[0010] Furthermore, the specific steps for generating the dynamic feature map include: The feature weight vector is multiplied element-wise with the operating condition vector to obtain the weighted operating condition feature vector. The feature weight vector contains three elements, which correspond to the three operating condition parameters of wind speed, active power and speed in the standardized operating condition vector. Element-wise multiplication means multiplying the first element of the feature weight vector with the first element of the operating condition vector, the second element with the second element, and the third element with the third element. That is, wind speed is multiplied by the corresponding wind speed weight, active power is multiplied by the corresponding active power weight, and speed is multiplied by the corresponding speed weight. This ensures that the three operating condition parameters are multiplied by their respective weight values ​​at the same sampling time, thereby obtaining three weighted operating condition feature values. Finally, the weighted operating condition feature vector is output. The weighted condition feature vector is spatially broadcast and replicated, expanding along the frequency axis to the same dimension as the frequency resolution of the normalized time-frequency plot, and simultaneously expanding along the time axis to the same dimension as the time resolution of the normalized time-frequency plot. It maintains three channels, with each channel having identical values ​​at all spatial locations, thus constructing a three-dimensional tensor. The spatial dimensions of this three-dimensional tensor are the same as those of the normalized time-frequency plot. Figure 1 Therefore, the number of channels is three.

[0011] Furthermore, the step of generating the dynamic feature map also includes: The vibration frequency band weight vector is extracted and multiplied with the corresponding elements in the standardized time-frequency graph. Each element in the vibration frequency band weight vector corresponds to a frequency band in the time-frequency graph, and the value of this element is the weight value corresponding to that frequency band. The standardized time-frequency graph is a two-dimensional matrix with rows arranged according to frequency bins and columns arranged according to the time axis. Each row represents the energy amplitude at all time positions within a frequency band. During the weighted multiplication, the matrix composed of all pixels in the time-frequency graph belonging to the first frequency band is multiplied by the first element in the vibration frequency band weight vector, the matrix composed of all pixels in the second frequency band is multiplied by the second element, and so on. That is, all pixels in each frequency band are multiplied by the same weight value corresponding to that frequency band, so that the frequency band with a higher weight value is enhanced as a whole, and the frequency band with a lower weight value is suppressed as a whole. The final weighted time-frequency graph has the same spatial size as the original time-frequency graph. The weighted time-frequency plot is a two-dimensional matrix, considered as a single-channel feature map. The three-dimensional tensor has the same frequency and time resolution as the weighted time-frequency plot in space, but it has three channels. The three channels respectively carry the weighted wind speed, active power and rotational speed feature information. The weighted time-frequency plot and the three-dimensional tensor are concatenated by channels, so that the number of channels of the resulting dynamic feature map increases to four while keeping the spatial size unchanged. The first channel is the weighted time-frequency plot, and the second to fourth channels are the three weighted operating condition features respectively.

[0012] Furthermore, the specific structure of the master diagnostic model is as follows: The first to third layers consist of three convolutional blocks with the same structure. Each convolutional block contains, in sequence: a two-dimensional convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max pooling layer. The fourth layer is a channel attention module. This module first performs global average pooling on the input tensor to compress it into a channel description vector. Then, the channel description vector is passed sequentially through a fully connected layer with 1 / 16 the number of channels, a ReLU activation function layer, a fully connected layer with the same number of neurons as the original number of channels, and a Sigmoid activation function layer to generate the weight coefficients for each channel. Finally, the weight coefficients are multiplied by the input tensor channel by channel to output a three-dimensional convolutional feature tensor. The main diagnostic model has a feature extraction layer after the fourth layer. This feature extraction layer flattens the three-dimensional convolutional feature tensor into a one-dimensional vector to obtain a deep feature vector. The deep feature vector is input into the fifth layer for classification on the one hand, and used to calculate the health baseline distance for the constructed support vector description model on the other hand. The fifth layer is the classification output layer. This classification output layer inputs the deep feature vector output from the fourth layer into a fully connected layer. The number of neurons in this fully connected layer is equal to the total number of fault categories. Finally, the Softmax function is applied to the output of this fully connected layer to convert it into a vector representing the probability distribution of each fault category.

[0013] Furthermore, the confidence level and the distance to the health baseline are determined, specifically as follows: The confidence level is the maximum value of each element in the fault probability distribution vector. The support vector description model is constructed as follows: obtain historical operating status data and vibration signal waveform data of the wind turbine generator under normal conditions, execute steps S1 to S3 to generate dynamic feature maps, and then input the dynamic feature maps into the main diagnostic model to obtain multiple sets of deep feature vectors under normal conditions. The Gaussian radial basis kernel function is selected to map multiple sets of normal depth feature vectors to obtain the center of a sphere that is a weighted combination of all depth feature vectors. The square of the radius of the sphere is the critical distance value. The depth feature vector at the current sampling time is input into the support vector description model. The squared distance from the depth feature vector to the center of the sphere is calculated as the health baseline distance. This squared distance is obtained by subtracting the depth feature vector from the center of the sphere element by element and then summing the squares of each component. An alarm is triggered when the confidence level is less than the preset confidence threshold, or the health baseline distance is greater than the critical distance value, and the current sampling time is marked as abnormal.

[0014] The present invention also provides an anomaly detection and fault diagnosis system for wind turbine generator sets, wherein the anomaly detection and fault diagnosis system for wind turbine generator sets is used to execute the above-described anomaly detection and fault diagnosis method for wind turbine generator sets, including: The data acquisition module is used to acquire the operating status data of the wind turbine generator at each sampling time, and to standardize the operating status data into a working condition vector. At the same time, a local window is set with the sampling time as the center, and the vibration signal waveform data within the local window is acquired as the vibration analysis signal at that time. The data processing module is used to perform continuous wavelet transform and synchronous compression on the vibration analysis signal to generate a time-frequency diagram of the current sampling time; the standardized working condition vector is input into a pre-trained feature dynamic weight network to obtain the weight vector corresponding to the current sampling time, and the weight vector is divided into feature weight vector and vibration frequency band weight vector according to the dimension. The image generation module is used to standardize the time-frequency map; based on the standardized time-frequency map, feature weight vector, and vibration frequency band weight vector, it generates a dynamic feature map for the current sampling time; and inputs the dynamic feature map into the pre-trained main diagnostic model to obtain the corresponding depth feature vector and fault probability distribution. The anomaly detection module is used to calculate the prediction confidence level based on the fault probability distribution, and simultaneously input the deep feature vector into a pre-constructed support vector description model to calculate the health baseline distance; perform anomaly determination based on the confidence level and the health baseline distance, and trigger an alarm when the warning conditions are met.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This scheme employs a time-frequency processing method combining continuous wavelet transform and synchronous compression transform. By rearranging and compressing the wavelet coefficients along the frequency direction, the originally divergent time-frequency energy is concentrated onto the true instantaneous frequency line, effectively overcoming the shortcomings of traditional time-frequency analysis methods such as insufficient resolution, energy divergence, and mode mixing. It can output high-definition time-frequency maps, providing high-quality vibration feature input for subsequent deep learning diagnosis. A feature dynamic weight network based on operating condition vectors is constructed. This network can adaptively generate feature weight vectors and vibration frequency band weight vectors according to real-time operating conditions such as wind speed, power, and rotational speed at the current sampling time. This allows the model to dynamically adjust the operating state features and vibration under different varying operating conditions. The importance of frequency bands is effectively addressed by solving the shortcomings of traditional fixed-weight methods in generalizing under varying operating conditions. By deeply fusing weighted operating condition information with time-frequency maps to construct dynamic feature maps, multimodal fusion of operating status data and vibration signal features is achieved. This enables the main diagnostic model to make comprehensive judgments using both operating condition information and time-frequency features, significantly improving the accuracy of fault identification. A dual-threshold judgment mechanism based on deep feature vectors combined with confidence and health baseline distance is used to identify known fault categories through the confidence of the fault probability distribution and to detect unknown abnormal states through the health baseline distance calculated by the support vector description model, thereby enhancing the safety and reliability of wind turbine operation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a graph showing the relationship between the synchronous compression transform coefficients and the center frequency of the frequency bin in this invention. Figure 3 This is a schematic diagram of the overall system structure of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0019] Example: Please see Figures 1-2 The present invention provides a technical solution: A method for anomaly detection and fault diagnosis of wind turbine generator sets, comprising the following specific steps: S1: At each sampling time, acquire the operating status data of the wind turbine generator set at that time, and standardize the operating status data into a working condition vector; at the same time, set a local window with the sampling time as the center, and collect the vibration signal waveform data within the local window as the vibration analysis signal at that time. In this embodiment, the operating status data includes the wind speed at the hub height of the wind turbine generator set, the active power output from the generator end of the wind turbine generator set, and the rotational speed. The vibration analysis signal is a continuous vibration signal segment centered on the current sampling time and cut out with a preset window length.

[0020] During the above process, when wind turbine generators operate under unsteady conditions such as varying wind speed and load, their vibration signals exhibit significant time-varying non-stationary characteristics. The fault characteristic frequency drifts with changes in rotational speed. Setting a local window centered on each sampling moment allows a continuous waveform segment before and after that moment to be used as the vibration analysis signal for that moment. This enables the time-frequency analysis results to reflect the local time-frequency characteristics of that moment, avoiding frequency ambiguity due to an excessively large window span, and also avoiding insufficient frequency resolution due to an excessively small window span. Combining wind speed, active power, and rotational speed into a condition vector and standardizing it aims to provide a unified dimensional input for the subsequent feature dynamic weighting network. This allows the network to adaptively generate weight vectors based on the current operating conditions, thereby achieving dynamic adjustment of the importance of features and frequency bands under different operating conditions. The preset window length is determined based on the generator's rotational frequency at rated speed and the gear meshing frequency. It is generally set to 5 to 10 times the reciprocal of the rotational frequency at rated speed, or according to empirical values, it is set to an integer power of 2 such as 2048 points or 4096 points, in order to balance calculation efficiency and time-frequency resolution. After assembling the operational status data into a condition vector, standardization is performed. Specifically, the three operational status data acquired at the current sampling time—wind speed, active power, and engine speed—are arranged in a fixed order to form a three-dimensional condition vector. The first element of this vector is the wind speed value, the second element is the active power value, and the third element is the engine speed value. Subsequently, this condition vector is standardized using the Z-score standardization method. That is, for each element in the condition vector, the mean of that operational status parameter in the historical data is subtracted, and then divided by the standard deviation of that operational status parameter in the historical data. This ensures that the mean of each element in the standardized condition vector is zero and the standard deviation is one, eliminating the impact of different operational status parameters on subsequent network training due to differences in units and numerical ranges.

[0021] S2: Perform continuous wavelet transform and synchronous compression on the vibration analysis signal to generate a time-frequency diagram at the current sampling time; input the standardized working condition vector into a pre-trained feature dynamic weight network to obtain the weight vector corresponding to the current sampling time; divide the weight vector into feature weight vector and vibration frequency band weight vector according to the dimension. In this embodiment, the specific steps for generating the time-frequency diagram include: The vibration analysis signal is inner productd with a scalable and translational Morlet wavelet function, where the scale parameter controls the scaling of the wavelet function and the translation parameter controls the movement of the wavelet function along the time axis. By traversing all scales and all translation positions, the wavelet coefficients of the signal and the wavelet function at different scales and positions are calculated. For the wavelet coefficients at each time-frequency position obtained by continuous wavelet transform, first calculate the partial derivative of the wavelet coefficient with respect to the translation parameter. This partial derivative reflects the rate at which the phase of the wavelet coefficient changes with the translation parameter. Then divide the partial derivative by the wavelet coefficient itself, multiply by the negative imaginary unit, and divide by twice pi to calculate the instantaneous frequency at that time-frequency position. The wavelet coefficients are subjected to synchronous compression transformation as follows: The frequency axis is pre-divided into several discrete frequency bins, each with a center frequency and a preset bandwidth. For each translation position, all discrete scales are traversed to determine whether the estimated instantaneous frequency at that scale falls within the bandwidth of a certain frequency bin. Only the wavelet coefficients at those scales that meet the condition are accumulated. Each wavelet coefficient involved in the accumulation needs to be multiplied by the scale interval corresponding to that scale and the negative cube of that scale for amplitude and scale interval normalization correction. Finally, the accumulation result is divided by the frequency bin bandwidth to complete the energy density normalization, resulting in synchronous compression transformation coefficients. The formula upon which the above process is based is: in, Represents vibration analysis signals; Indicates the scale parameter; Indicates the translation parameter; This represents the complex conjugate of the Morlet wavelet function; Indicated in scale Translation parameters Wavelet coefficients at; Calculate the instantaneous frequency: in, Indicated in scale Translation parameters The estimated instantaneous frequency; Represents the imaginary unit; The wavelet coefficients are then subjected to synchronous compression transformation: in, Indicates the first The center frequency of each frequency compartment; Represented at discrete scale Translation parameters The instantaneous frequency at that point; Indicates the preset frequency compartment bandwidth; Represented at discrete scale Scale intervals at locations; Indicates the translation parameters Location, center frequency Synchronous compression transform coefficients at the location; Indicates the preset discrete scale number index; Indicates the frequency bin index; In the above process, although continuous wavelet transform can map one-dimensional vibration signals to a two-dimensional time-scale plane, its energy exhibits significant divergence in the scale direction. That is, the same frequency component will generate non-zero wavelet coefficients at multiple adjacent scales, causing the energy in the time-frequency graph to diffuse within a region rather than be concentrated on a line, resulting in low frequency resolution and difficulty in distinguishing adjacent frequency components. Introducing instantaneous frequency into the formula... This allows for the independent estimation of the true instantaneous frequency at each time-frequency location, thus determining which frequency bin the wavelet coefficients at that location should belong to on the frequency axis. Synchronous compression transform utilizes this mechanism to redistribute and concentrate the divergent time-frequency energy from the original wavelet transform onto its corresponding instantaneous frequency line, thereby significantly improving the frequency resolution and energy concentration of the time spectrum while maintaining the same time resolution. Simultaneously, due to the actual calculation of mesoscale parameters... Discretized into a finite number of discrete scales The contributions of wavelet coefficients differ at different discrete scales, as shown in the formula. The term is used to perform amplitude normalization correction on the wavelet coefficients at each discrete scale, eliminating the attenuation or amplification effect of the inherent amplitude variation with scale in wavelet transform; multiplied by The term is used to compensate for the weights of discrete intervals at different scales, ensuring that the contributions of each scale are correctly handled when the scale intervals are non-uniform; finally, divide by The term is used to normalize the bandwidth of the frequency compartment, so that the energy density in the compressed frequency compartment is independent of the bandwidth of the frequency compartment, ensuring that the amplitude of the synchronous compression transform coefficient is comparable between different frequency compartments, and the final output time-frequency diagram can truly reflect the energy distribution of each frequency component of the vibration signal. In engineering practice, the preset number of discrete scales is usually determined by uniformly sampling in logarithmic space based on the ratio of the minimum analysis frequency to the maximum analysis frequency. It is generally set to an integer value between 64 and 256. For applications that require fine resolution of gear meshing frequencies and their modulation sidebands, the number of discrete scales is usually set to 128 to ensure sufficient frequency resolution.

[0022] In the above embodiments, 20 sets of data on the center frequency of the frequency compartment and the corresponding synchronization compression transform coefficients are given to reflect the changes in the synchronization compression transform coefficients as the center frequency of the frequency compartment changes, as shown in Table 1: Table 1: Relationship between synchronous compression transform coefficients and the center frequency of the frequency bin In Table 1 above, the vibration analysis signal sampling frequency is set to 25600Hz, the preset window length is 4096 sampling points, and the number of discrete scales is... ,scale Uniformly distributed in logarithmic space, frequency bin bandwidth Hz, translation parameter Fixed at the center of the window, the signal contains the simulated gear meshing frequency component (approximately 800 Hz) and its sidebands; the synchronous compression transform coefficient reaches a peak of 0.512 near the gear meshing frequency of 800 Hz, reflecting that this frequency component dominates the energy in the vibration signal; as the frequency deviates from 800 Hz, the coefficient amplitude gradually decreases, reflecting the concentration characteristics of time-frequency energy around the real instantaneous frequency line, as well as the sharpening and enhancement effect of the synchronous compression transform on the main frequency component.

[0023] The final output time-frequency graph is a two-dimensional image. Its row direction corresponds to the frequency axis, and the resolution is determined by the number of preset frequency bins. Its column direction corresponds to the time axis, and the resolution is determined by the number of sampling points in the preset window. The value of each pixel in the time-frequency graph is the amplitude of the synchronous compression transform coefficient at the corresponding sampling time and frequency.

[0024] In the above process, when the wind turbine is running under variable speed conditions, the frequency components of its vibration signal drift with the change of speed. The fault characteristic frequency is not a fixed value. By mapping the one-dimensional vibration analysis signal to a two-dimensional time-scale plane through continuous wavelet transform, the multi-resolution characteristics of wavelet transform can be used to balance the frequency resolution of the low-frequency band and the time positioning accuracy of the high-frequency band. On this basis, since the energy of wavelet transform diverges in the scale direction, the time spectrum becomes blurred. By calculating the instantaneous frequency of each time frequency position and performing synchronous compression along the frequency direction, all wavelet coefficients falling near the same frequency bin are rearranged and clustered. This can squeeze the originally divergent time frequency energy to the center frequency line, effectively improving the frequency resolution and energy concentration of the time spectrum, thus providing a high-quality time frequency map with clear frequency band boundaries and concentrated energy for subsequent frequency band dynamic weighting. In engineering practice, the number of frequency bins is usually determined based on the ratio of the sampling frequency to the minimum frequency resolution, and is generally set to an integer value between 128 and 1024. For key transmission components such as gearboxes and main bearings, in order to ensure that the modulation sideband of the gear meshing frequency and the bearing fault characteristic frequency can be clearly distinguished, the number of frequency bins is usually set to 512 or 1024.

[0025] The feature dynamic weight network is based on a lightweight multilayer perceptron and includes an input layer and an output layer. The input layer contains several neurons, and the output is processed by the ReLU activation function. The output layer contains multiple neurons and the output value of each neuron is compressed using the Sigmoid activation function. The feature dynamic weight network and the main diagnostic model are obtained through the following joint training method: acquire historical operating status data, vibration signal waveform data and corresponding fault category labels, execute steps S1 to S3 sequentially on the historical operating status data and vibration signal waveform data to generate historical dynamic feature maps corresponding to each historical sampling time, input the historical dynamic feature maps into the main diagnostic model, take the cross-entropy loss between the fault category prediction result output by the main diagnostic model and the fault category label as the optimization objective, and synchronously update all trainable parameters in the feature dynamic weight network and the main diagnostic model through the backpropagation algorithm until the cross-entropy loss converges to the preset value.

[0026] In the above process, during training, after each training epoch, the current model is evaluated on the validation set, and the cross-entropy loss value and fault classification accuracy on the validation set are calculated. When the cross-entropy loss value on the validation set no longer shows a significant decreasing trend for several consecutive training epochs (usually set to 10 to 20 epochs), and the fault classification accuracy on the validation set tends to stabilize and reach an acceptable level (usually not lower than 95%), the cross-entropy loss value corresponding to that moment is used as the preset value, or the preset value is directly set to a sufficiently small constant (such as 0.01 or 0.001). When the training loss drops below this value, it is determined to be converged. The output of the dynamic feature weight network itself lacks direct supervisory labels and cannot be trained using an independent loss function. Only by training it end-to-end with the main diagnostic model as a whole, and using the cross-entropy loss between the fault category prediction results and the true fault category labels output by the main diagnostic model as the optimization objective, can the classification error be propagated back to the dynamic feature weight network layer by layer through the backpropagation algorithm. This drives the network to learn how to generate weight vectors that are beneficial to the final classification based on the input operating condition vector. Historical operating status data and vibration signal waveform data are processed sequentially through steps S1 to S3 to generate historical dynamic feature maps before being input into the main diagnostic model. This ensures that the data processing flow during the training phase is completely consistent with that during the inference phase, allowing the trained network to be used directly in actual deployment. Synchronously updating all trainable parameters in the dynamic feature weight network and the main diagnostic model enables the two networks to collaborate and optimize together during training. The dynamic feature weight network learns to dynamically adjust the importance of features and frequency bands according to the current operating conditions, while the main diagnostic model learns a more accurate classification decision boundary based on this optimized dynamic feature map.

[0027] The weight vector is divided into feature weight vector and vibration frequency band weight vector according to dimension, specifically: The standardized operating condition vector is input into the feature dynamic weight network. First, it undergoes a first-layer linear transformation and is processed by the ReLU activation function to extract the deep latent features of the operating condition vector. Then, it undergoes a second-layer linear transformation and is processed by the Sigmoid activation function to compress the output value to between 0 and 1, generating a weight vector corresponding to the input operating condition vector. This weight vector is then divided by dimension, and the first three elements are used to form the feature weight vector, while the remaining elements form the vibration frequency band weight vector. The total number of elements in the vibration frequency band weight vector is equal to the preset number of frequency band divisions.

[0028] In the above process, the operating condition vector contains three physical quantities: wind speed, active power, and rotational speed. These quantities collectively describe the current operating condition of the wind turbine generator. Through a first-layer linear transformation and processing with the ReLU activation function, the original three-dimensional operating condition vector can be mapped to a high-dimensional latent space, extracting deep latent features related to fault diagnosis. This allows the network to capture the influence of the nonlinear interaction between operating condition parameters on vibration characteristics. A second-layer linear transformation and a Sigmoid activation function compress the output value to between 0 and 1, generating a weight vector corresponding to the input operating condition vector. This weight vector... Each element value of the quantity is constrained between 0 and 1, so that the weight value has a clear physical meaning: a value close to 1 indicates that the feature or frequency band is highly important for fault diagnosis under the current working condition and should be enhanced, while a value close to 0 indicates that it should be suppressed. The weight vector is divided by dimension, and the first 3 elements are used to form a feature weight vector for element-wise weighting of the working condition vector. The remaining elements form a vibration frequency band weight vector for weighting each frequency band in the time-frequency diagram. This allows the same working condition vector to generate two different weights for different purposes, realizing both dynamic recalibration of the operating status data and adaptive selection of vibration frequency bands. In engineering practice, the number of frequency bands is usually consistent with the number of frequency blocks preset in synchronous compression conversion. That is, each frequency block is treated as an independent frequency band, or several adjacent frequency blocks are merged into one frequency band according to the frequency range of concern for diagnosis. The number of frequency bands is generally set to an integer value between 64 and 256. For scenarios such as gearboxes and main bearings that require fine differentiation of fault characteristic frequencies and their modulation sidebands, the number of frequency bands is usually set to 128 or 256.

[0029] S3: Standardize the time-frequency graph; generate a dynamic feature map for the current sampling time based on the standardized time-frequency graph, feature weight vector, and vibration frequency band weight vector; input the dynamic feature map into the pre-trained main diagnostic model to obtain the corresponding depth feature vector and fault probability distribution; In this embodiment, the specific steps for generating the dynamic feature map include: The feature weight vector is multiplied element-wise with the operating condition vector to obtain the weighted operating condition feature vector. The feature weight vector contains three elements, which correspond to the three operating condition parameters of wind speed, active power and speed in the standardized operating condition vector. Element-wise multiplication means multiplying the first element of the feature weight vector with the first element of the operating condition vector, the second element with the second element, and the third element with the third element. That is, wind speed is multiplied by the corresponding wind speed weight, active power is multiplied by the corresponding active power weight, and speed is multiplied by the corresponding speed weight. This ensures that the three operating condition parameters are multiplied by their respective weight values ​​at the same sampling time, thereby obtaining three weighted operating condition feature values. Finally, the weighted operating condition feature vector is output. The formula used to obtain the weighted working condition feature vector is: in, Indicates the first Feature weight vector at each sampling time; This represents element-wise multiplication; Indicates the first Standardized operating condition vector at each sampling time; Indicates the first The weighted feature vector of the working condition at each sampling time; Indicates the index of the current sampling time; The weighted condition feature vector is spatially broadcast and replicated, expanding along the frequency axis to the same dimension as the frequency resolution of the normalized time-frequency plot, and simultaneously expanding along the time axis to the same dimension as the time resolution of the normalized time-frequency plot. It maintains three channels, with each channel having identical values ​​at all spatial locations, thus constructing a three-dimensional tensor. The spatial dimensions of this three-dimensional tensor are the same as those of the normalized time-frequency plot. Figure 1 Therefore, the number of channels is three.

[0030] In the above process, there is a complex coupling relationship between wind speed, active power, and rotational speed in the operating condition vector and fault features in the vibration signal. Multiplying the feature weight vector element-wise with the operating condition vector allows for the application of different weights to the three operating condition parameters based on the current operating conditions. This strengthens the operating condition parameters that contribute more to the current fault diagnosis and suppresses those that contribute less or are irrelevant, thereby improving the model's adaptability to changes in operating conditions. Since the weighted operating condition information needs to be concatenated with the time-frequency graph, and the time-frequency graph is two-dimensional spatial data (frequency × time), while the weighted operating condition feature vector is only one-dimensional, it cannot be directly aligned with the time-frequency graph in terms of spatial dimension. Therefore, it is necessary to perform spatial broadcast replication, extending along the frequency axis to the same dimension as the frequency resolution of the time-frequency graph and along the time axis to the same dimension as the time resolution of the time-frequency graph. This ensures that the weighted operating condition information appears at every time-frequency position, thus constructing a spatial dimension and time-frequency... Figure 1 The three-dimensional tensor has three channels. The reason for keeping the three channels and ensuring that each channel has the same value is that the three operating parameters each represent an independent physical quantity and carry different operating information. In subsequent convolution operations, the three channels can be independently assigned different convolution kernel weights, so that the network can flexibly learn and utilize the different contributions of the three operating parameters to fault diagnosis.

[0031] The step of generating the dynamic feature map further includes: The vibration frequency band weight vector is extracted and multiplied with the corresponding elements in the standardized time-frequency graph. Each element in the vibration frequency band weight vector corresponds to a frequency band in the time-frequency graph, and the value of this element is the weight value corresponding to that frequency band. The standardized time-frequency graph is a two-dimensional matrix with rows arranged according to frequency bins and columns arranged according to the time axis. Each row represents the energy amplitude at all time positions within a frequency band. During the weighted multiplication, the matrix composed of all pixels in the time-frequency graph belonging to the first frequency band is multiplied by the first element in the vibration frequency band weight vector, the matrix composed of all pixels in the second frequency band is multiplied by the second element, and so on. That is, all pixels in each frequency band are multiplied by the same weight value corresponding to that frequency band, so that the frequency band with a higher weight value is enhanced as a whole, and the frequency band with a lower weight value is suppressed as a whole. The final weighted time-frequency graph has the same spatial size as the original time-frequency graph. The weighted time-frequency plot is a two-dimensional matrix, considered as a single-channel feature map. The three-dimensional tensor has the same frequency and time resolution as the weighted time-frequency plot in space, but it has three channels. The three channels respectively carry the weighted wind speed, active power and rotational speed feature information. The weighted time-frequency plot and the three-dimensional tensor are concatenated by channels, so that the number of channels of the resulting dynamic feature map increases to four while keeping the spatial size unchanged. The first channel is the weighted time-frequency plot, and the second to fourth channels are the three weighted operating condition features respectively.

[0032] The formula upon which the above process is based is: in, This represents the first element in the vibration frequency band weight vector at the current sampling time. The weight value corresponding to each frequency band; This indicates that the time-frequency graph after standardization at the current sampling time belongs to the [number]th [sample]. A matrix composed of all pixels in each frequency band; Indicates the frequency band index; This represents the weighted time-frequency plot at the current sampling time. Amplitude matrix of each frequency band; By concatenating the weighted time-frequency graph with the three-dimensional tensor, the dynamic feature map at the current sampling time is obtained: in, A dynamic feature map representing the current sampling time; This represents the weighted time-frequency graph at the current sampling time. Represents the three-dimensional tensor at the current sampling time; This indicates a splicing operation along the channel dimension.

[0033] In the above process, different frequency bands in the standardized time-frequency graph contribute differently to fault diagnosis. Some frequency bands, such as gear meshing frequency and its sidebands, and the frequency band containing bearing fault characteristic frequencies, carry key fault information, while other frequency bands, such as power frequency and its harmonics, and the frequency band dominated by environmental noise, contribute less to fault diagnosis or even have an interfering effect. By weighting the vibration frequency band weight vector with the corresponding frequency band in the time-frequency graph, the fault diagnosis can be improved. It can independently apply a weight value between 0 and 1 to each frequency band according to the current operating conditions, thereby enhancing fault-sensitive frequency bands and suppressing irrelevant or interfering frequency bands, achieving adaptive screening and dynamic recalibration of vibration frequency bands; the weighted time-frequency graph The three-dimensional tensor obtained by copying the working condition features via broadcast. splicing along the channel dimension This increases the number of channels in the dynamic feature map to four, with the first channel being the weighted time-frequency map and the second to fourth channels being the three weighted operating condition features. This achieves multimodal fusion of operating status data and vibration signal features across channels, enabling the two-dimensional convolution operation in the subsequent main diagnostic model to simultaneously extract joint features between operating condition information and time-frequency features across channels. This allows the use of operating condition information to assist in the interpretation of vibration features and the use of vibration features to verify the rationality of operating condition information, significantly improving the accuracy and robustness of fault diagnosis in variable operating condition scenarios.

[0034] The specific structure of the master diagnostic model is as follows: The first to third layers consist of three convolutional blocks with the same structure. Each convolutional block contains, in sequence: a two-dimensional convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max pooling layer. The fourth layer is a channel attention module. This module first performs global average pooling on the input tensor to compress it into a channel description vector. Then, the channel description vector is passed sequentially through a fully connected layer with 1 / 16 the number of channels, a ReLU activation function layer, a fully connected layer with the same number of neurons as the original number of channels, and a Sigmoid activation function layer to generate the weight coefficients for each channel. Finally, the weight coefficients are multiplied by the input tensor channel by channel to output a three-dimensional convolutional feature tensor. The main diagnostic model has a feature extraction layer after the fourth layer. This feature extraction layer flattens the three-dimensional convolutional feature tensor into a one-dimensional vector to obtain a deep feature vector. The deep feature vector is input into the fifth layer for classification on the one hand, and used to calculate the health baseline distance for the constructed support vector description model on the other hand. The fifth layer is the classification output layer. This classification output layer inputs the deep feature vector output from the fourth layer into a fully connected layer. The number of neurons in this fully connected layer is equal to the total number of fault categories. Finally, the Softmax function is applied to the output of this fully connected layer to convert it into a vector representing the probability distribution of each fault category.

[0035] In the above process, the dynamic feature map contains both vibration time-frequency information and operating condition feature information. Spatially, it carries the correspondence between frequency bands and time; channel-wise, it carries the correspondence between vibration energy and operating condition parameters. Through three structurally identical convolutional blocks (layers one through three), each containing a two-dimensional convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max-pooling layer, the two-dimensional convolutional operation can simultaneously extract the local correlation between adjacent time-frequency positions in the spatial dimension and the cross-channel correlation between vibration features and operating condition features in the channel dimension. The batch normalization layer accelerates training convergence and suppresses internal covariate shifts. The ReLU activation function... To enhance nonlinear expressive power, the max pooling layer progressively compresses the spatial size and strengthens the translation invariance of features. The fourth layer employs a channel attention module, which compresses the 3D convolutional feature tensor into a channel description vector through global average pooling. This vector is then sequentially passed through a fully connected layer with 1 / 16 of the number of channels for dimensionality reduction, ReLU activation, a fully connected layer with the original number of channels to restore the dimension, and a Sigmoid activation function to generate channel weight coefficients. This achieves adaptive evaluation of the importance of each channel in the 3D convolutional feature tensor. The weight coefficients are multiplied with the input tensor channel by channel to enhance important channels and suppress irrelevant channels, thereby focusing on the feature channels that contribute the most to the current fault diagnosis. After the fourth layer, a feature extraction layer is set up to flatten the three-dimensional convolutional feature tensor into a one-dimensional vector to obtain a deep feature vector. This deep feature vector serves as the input to the fifth fully connected layer and the Softmax function to output the probability distribution of the fault category. On the other hand, it is used by the support vector description model to calculate the health baseline distance, realizing the sharing of the same deep feature in classification tasks and anomaly detection tasks. This enables the main diagnostic model to identify known fault categories and provide discriminative deep features for the detection of unknown anomalies.

[0036] S4: Calculate the prediction confidence level based on the fault probability distribution, and simultaneously input the deep feature vector into the pre-constructed support vector description model to calculate the health baseline distance; determine the anomaly based on the confidence level and the health baseline distance, and trigger an alarm when the warning conditions are met.

[0037] In this embodiment, the confidence level and the distance to the health baseline are determined as follows: The confidence level is the maximum value of each element in the failure probability distribution vector, based on the following formula: in, This represents the fault probability distribution vector at the current sampling moment; This represents the prediction confidence level at the current sampling time; The support vector description model is constructed as follows: historical operating status data and vibration signal waveform data of wind turbine generator under normal conditions are obtained, steps S1 to S3 are executed to generate dynamic feature maps, and then the dynamic feature maps are input into the main diagnostic model to obtain multiple sets of deep feature vectors under normal conditions. The Gaussian radial basis kernel function is selected to map multiple sets of normal depth feature vectors to obtain the center of a sphere that is a weighted combination of all depth feature vectors. The square of the radius of the sphere is the critical distance value. The depth feature vector at the current sampling time is input into the support vector description model. The squared distance from the depth feature vector to the center of the sphere is calculated as the health baseline distance. This squared distance is obtained by subtracting the depth feature vector from the center of the sphere element by element and then summing the squares of each component. The formula used to calculate the health baseline distance is: in, Indicates the center of the ball; This represents the depth feature vector at the current sampling time. Indicates the health baseline distance at the current sampling time; An alarm is triggered when the confidence level is less than the preset confidence threshold, or the health baseline distance is greater than the critical distance value, and the current sampling time is marked as abnormal.

[0038] In the above process, the fault probability distribution vector output by the fifth layer of the main diagnostic model It includes the probability values ​​of the current sampling time belonging to various types of faults, among which the maximum probability value This confidence level represents the model's level of certainty regarding the current prediction result. A high confidence level indicates that the model's classification decision has high reliability, while a low confidence level indicates that the model is hesitant between categories, possibly due to encountering unknown abnormal patterns not seen in the training set or being in the boundary region of the fault category. Simultaneously, by acquiring historical data under normal conditions, performing steps S1 to S3 to generate a dynamic feature map, and inputting it into the main diagnostic model to extract depth feature vectors, a hyperspherical description model of the depth feature distribution under normal operating conditions can be established. A Gaussian radial basis kernel function is selected to map the normal depth features to a high-dimensional feature space, resulting in a sphere that can enclose all normal samples with the smallest radius. The square of the radius of the sphere is the critical distance value, when the depth feature vector of the sample to be detected... To the center of the ball Square distance When the distance exceeds this critical distance, it indicates that the current sample deviates from the depth feature distribution range under normal operating conditions, which may correspond to an unknown type of abnormal state or early failure. The system uses confidence level and health baseline distance together for judgment. When the confidence level is lower than the preset confidence level threshold or the health baseline distance is greater than the critical distance value, an alarm is triggered. This allows the main diagnostic model to provide a confidence assessment for known fault categories and to effectively perceive unknown abnormal states. Thus, it can simultaneously take into account the dual functions of fault classification and abnormal warning within the same framework. The preset reliability threshold needs to be determined based on the confidence distribution on the validation set and the trade-off between the false negative rate and the false positive rate. Specifically, after the model training is completed, the validation set data is input into the model to obtain the confidence value at each sampling time. The confidence distribution of normal samples and various types of fault samples are statistically analyzed. Normal samples usually have a high confidence, while fault samples often have a low confidence because they deviate from the training distribution. The receiver operating characteristic curve is plotted based on the false negative rate and false positive rate corresponding to different confidence thresholds on the validation set. The confidence value that balances the false negative rate and the false positive rate is selected as the preset reliability threshold. Alternatively, in engineering practice, a fixed value is set according to the specific safety requirements of the wind farm, usually between 0.6 and 0.9. The construction process of a support vector description model is essentially solving an optimization problem that can enclose all normal depth feature vectors with the smallest hypersphere. Specifically, let the set of depth feature vectors in the normal state be... The goal of the support vector model is to find a center of a sphere in a high-dimensional feature space. and radius This allows the hypersphere to enclose all or most of the normal samples with a minimum volume, and its optimization objective is to minimize... ,in For penalty parameters, As relaxation variables, some outliers are allowed to lie outside the hypersphere to enhance the robustness of the model. Since normal samples are not simply spherically distributed in the high-dimensional feature space, directly solving for the center and radius of the sphere in the high-dimensional space is computationally difficult. Therefore, the Gaussian radial basis function is chosen. The normal depth feature vectors are implicitly mapped to a higher-dimensional feature space, in which a set of weight coefficients are obtained by solving a dual optimization problem. center of the ball That is, the weighted combination of all normal samples after mapping, where The samples are the support vectors, which determine the boundary of the hypersphere; the critical distance value is the square of the sphere's radius. The distance to the center of the sphere is determined by calculating the squared distance from any support vector to the center of the sphere for a new sample to be detected. Its health benchmark distance The kernel function can be used to directly calculate the distance without explicitly knowing the specific coordinates of the sphere's center. When this distance is greater than the critical distance value, it indicates that the sample being tested has deviated from the characteristic distribution range of the normal state and is judged as abnormal.

[0039] Please see Figure 3 The present invention also provides an anomaly detection and fault diagnosis system for wind turbine generator sets, wherein the anomaly detection and fault diagnosis system for wind turbine generator sets is used to execute the above-described anomaly detection and fault diagnosis method for wind turbine generator sets, including: The data acquisition module is used to acquire the operating status data of the wind turbine generator at each sampling time, and to standardize the operating status data into a working condition vector. At the same time, a local window is set with the sampling time as the center, and the vibration signal waveform data within the local window is acquired as the vibration analysis signal at that time. The data processing module is used to perform continuous wavelet transform and synchronous compression on the vibration analysis signal to generate a time-frequency diagram of the current sampling time; the standardized working condition vector is input into a pre-trained feature dynamic weight network to obtain the weight vector corresponding to the current sampling time, and the weight vector is divided into feature weight vector and vibration frequency band weight vector according to the dimension. The image generation module is used to standardize the time-frequency map; based on the standardized time-frequency map, feature weight vector, and vibration frequency band weight vector, it generates a dynamic feature map for the current sampling time; and inputs the dynamic feature map into the pre-trained main diagnostic model to obtain the corresponding depth feature vector and fault probability distribution. The anomaly detection module is used to calculate the prediction confidence level based on the fault probability distribution, and simultaneously input the deep feature vector into a pre-constructed support vector description model to calculate the health baseline distance; perform anomaly determination based on the confidence level and the health baseline distance, and trigger an alarm when the warning conditions are met.

[0040] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by 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.

[0041] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0042] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for anomaly detection and fault diagnosis of wind turbine generator sets, characterized in that, include: S1: At each sampling time, acquire the operating status data of the wind turbine generator set at that time, and standardize the operating status data into a working condition vector; at the same time, set a local window with the sampling time as the center, and collect the vibration signal waveform data within the local window as the vibration analysis signal at that time. S2: Perform continuous wavelet transform and synchronous compression on the vibration analysis signal to generate a time-frequency diagram at the current sampling time; input the standardized working condition vector into a pre-trained feature dynamic weight network to obtain the weight vector corresponding to the current sampling time; divide the weight vector into feature weight vector and vibration frequency band weight vector according to the dimension. S3: Standardize the time-frequency graph; generate a dynamic feature map for the current sampling time based on the standardized time-frequency graph, feature weight vector, and vibration frequency band weight vector; input the dynamic feature map into the pre-trained main diagnostic model to obtain the corresponding depth feature vector and fault probability distribution; S4: Calculate the prediction confidence level based on the fault probability distribution, and simultaneously input the deep feature vector into the pre-constructed support vector description model to calculate the health baseline distance; determine the anomaly based on the confidence level and the health baseline distance, and trigger an alarm when the warning conditions are met.

2. The method for anomaly detection and fault diagnosis of wind turbine generator sets according to claim 1, characterized in that, The operating status data includes the wind speed at the hub height of the wind turbine generator set, the active power output from the generator end of the wind turbine generator set, and the rotational speed. The vibration analysis signal is a continuous vibration signal segment centered on the current sampling time and cut out with a preset window length.

3. The anomaly detection and fault diagnosis method for wind turbine generator sets according to claim 1, characterized in that, The specific steps for generating the time-frequency diagram include: The vibration analysis signal is inner productd with a scalable and translational Morlet wavelet function, where the scale parameter controls the scaling of the wavelet function and the translation parameter controls the movement of the wavelet function along the time axis. By traversing all scales and all translation positions, the wavelet coefficients of the signal and the wavelet function at different scales and positions are calculated. For the wavelet coefficients at each time-frequency position obtained by continuous wavelet transform, first calculate the partial derivative of the wavelet coefficient with respect to the translation parameter. This partial derivative reflects the rate at which the phase of the wavelet coefficient changes with the translation parameter. Then divide the partial derivative by the wavelet coefficient itself, multiply by the negative imaginary unit, and divide by twice pi to calculate the instantaneous frequency at that time-frequency position. The wavelet coefficients are subjected to synchronous compression transformation as follows: The frequency axis is pre-divided into several discrete frequency bins, each with a center frequency and a preset bandwidth. For each translation position, all discrete scales are traversed to determine whether the estimated instantaneous frequency at that scale falls within the bandwidth of a certain frequency bin. Only the wavelet coefficients at those scales that meet the condition are accumulated. Each wavelet coefficient participating in the accumulation needs to be multiplied by the scale interval corresponding to that scale and the negative cube of that scale for amplitude and scale interval normalization correction. Finally, the accumulation result is divided by the frequency bin bandwidth to complete the energy density normalization, resulting in synchronous compression transformation coefficients. The final output time-frequency graph is a two-dimensional image. Its row direction corresponds to the frequency axis, and the resolution is determined by the number of preset frequency bins. Its column direction corresponds to the time axis, and the resolution is determined by the number of sampling points in the preset window. The value of each pixel in the time-frequency graph is the amplitude of the synchronous compression transform coefficient at the corresponding sampling time and frequency.

4. The anomaly detection and fault diagnosis method for wind turbine generator sets according to claim 1, characterized in that, The feature dynamic weight network is based on a lightweight multilayer perceptron and includes an input layer and an output layer. The input layer contains several neurons, and the output is processed by the ReLU activation function. The output layer contains multiple neurons and the output value of each neuron is compressed using the Sigmoid activation function. The feature dynamic weight network and the main diagnostic model are obtained through the following joint training method: acquire historical operating status data, vibration signal waveform data and corresponding fault category labels, execute steps S1 to S3 sequentially on the historical operating status data and vibration signal waveform data to generate historical dynamic feature maps corresponding to each historical sampling time, input the historical dynamic feature maps into the main diagnostic model, take the cross-entropy loss between the fault category prediction result output by the main diagnostic model and the fault category label as the optimization objective, and synchronously update all trainable parameters in the feature dynamic weight network and the main diagnostic model through the backpropagation algorithm until the cross-entropy loss converges to the preset value.

5. The anomaly detection and fault diagnosis method for wind turbine generator sets according to claim 4, characterized in that, The weight vector is divided into feature weight vector and vibration frequency band weight vector according to dimension, specifically: The standardized working condition vector is input into the feature dynamic weight network. First, it undergoes a first-layer linear transformation and is processed by the ReLU activation function to extract the deep hidden features of the working condition vector. After a second linear transformation and a Sigmoid activation function, the output value is compressed to between 0 and 1, generating a weight vector corresponding to the input working condition vector. This weight vector is then divided by dimension, with the first three elements forming the feature weight vector and the remaining elements forming the vibration frequency band weight vector. The total number of elements in the vibration frequency band weight vector is equal to the preset number of frequency band divisions.

6. The anomaly detection and fault diagnosis method for wind turbine generator sets according to claim 1, characterized in that, The specific steps for generating the dynamic feature map include: The feature weight vector is multiplied element-wise with the operating condition vector to obtain the weighted operating condition feature vector. The feature weight vector contains three elements, which correspond to the three operating condition parameters of wind speed, active power and speed in the standardized operating condition vector. Element-wise multiplication means multiplying the first element of the feature weight vector with the first element of the operating condition vector, the second element with the second element, and the third element with the third element. That is, wind speed is multiplied by the corresponding wind speed weight, active power is multiplied by the corresponding active power weight, and speed is multiplied by the corresponding speed weight. This ensures that the three operating condition parameters are multiplied by their respective weight values ​​at the same sampling time, thereby obtaining three weighted operating condition feature values. Finally, the weighted operating condition feature vector is output. The weighted operating condition feature vector is broadcast and copied spatially, extended along the frequency axis to the same dimension as the frequency resolution of the normalized time-frequency plot, and simultaneously extended along the time axis to the same dimension as the time resolution of the normalized time-frequency plot. It also maintains three channels in the channel dimension, with each channel having the same value at each spatial location, thus constructing a three-dimensional tensor. The spatial size of this three-dimensional tensor is consistent with that of the normalized time-frequency plot, and the number of channels is three.

7. The anomaly detection and fault diagnosis method for wind turbine generator sets according to claim 6, characterized in that, The step of generating the dynamic feature map further includes: The vibration frequency band weight vector is extracted and multiplied with the corresponding elements in the standardized time-frequency graph. Each element in the vibration frequency band weight vector corresponds to a frequency band in the time-frequency graph, and the value of this element is the weight value corresponding to that frequency band. The standardized time-frequency graph is a two-dimensional matrix with rows arranged according to frequency bins and columns arranged according to the time axis. Each row represents the energy amplitude at all time positions within a frequency band. During the weighted multiplication, the matrix composed of all pixels in the time-frequency graph belonging to the first frequency band is multiplied by the first element in the vibration frequency band weight vector, the matrix composed of all pixels in the second frequency band is multiplied by the second element, and so on. That is, all pixels in each frequency band are multiplied by the same weight value corresponding to that frequency band, so that the frequency band with a higher weight value is enhanced as a whole, and the frequency band with a lower weight value is suppressed as a whole. The final weighted time-frequency graph has the same spatial size as the original time-frequency graph. The weighted time-frequency plot is a two-dimensional matrix, considered as a single-channel feature map. The three-dimensional tensor has the same frequency and time resolution as the weighted time-frequency plot in space, but it has three channels. The three channels respectively carry the weighted wind speed, active power and rotational speed feature information. The weighted time-frequency plot and the three-dimensional tensor are concatenated by channels, so that the number of channels of the resulting dynamic feature map increases to four while keeping the spatial size unchanged. The first channel is the weighted time-frequency plot, and the second to fourth channels are the three weighted operating condition features respectively.

8. The anomaly detection and fault diagnosis method for wind turbine generator sets according to claim 4, characterized in that, The specific structure of the master diagnostic model is as follows: The first to third layers consist of three convolutional blocks with the same structure. Each convolutional block contains, in sequence: a two-dimensional convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max pooling layer. The fourth layer is a channel attention module. This module first performs global average pooling on the input tensor to compress it into a channel description vector. Then, the channel description vector is passed sequentially through a fully connected layer with 1 / 16 the number of channels, a ReLU activation function layer, a fully connected layer with the same number of neurons as the original number of channels, and a Sigmoid activation function layer to generate the weight coefficients for each channel. Finally, the weight coefficients are multiplied by the input tensor channel by channel to output a three-dimensional convolutional feature tensor. The main diagnostic model has a feature extraction layer after the fourth layer. This feature extraction layer flattens the three-dimensional convolutional feature tensor into a one-dimensional vector to obtain a deep feature vector. The deep feature vector is input into the fifth layer for classification on the one hand, and used to calculate the health baseline distance for the constructed support vector description model on the other hand. The fifth layer is the classification output layer. This classification output layer inputs the deep feature vector output from the fourth layer into a fully connected layer. The number of neurons in this fully connected layer is equal to the total number of fault categories. Finally, the Softmax function is applied to the output of this fully connected layer to convert it into a vector representing the probability distribution of each fault category.

9. The anomaly detection and fault diagnosis method for wind turbine generator sets according to claim 8, characterized in that, The confidence level and the distance to the health baseline are determined as follows: The confidence level is the maximum value of each element in the failure probability distribution vector; The support vector description model is constructed as follows: historical operating status data and vibration signal waveform data of wind turbine generator under normal conditions are obtained, steps S1 to S3 are executed to generate dynamic feature maps, and then the dynamic feature maps are input into the main diagnostic model to obtain multiple sets of deep feature vectors under normal conditions. The Gaussian radial basis kernel function is selected to map multiple sets of normal depth feature vectors to obtain the center of a sphere that is a weighted combination of all depth feature vectors. The square of the radius of the sphere is the critical distance value. The depth feature vector at the current sampling time is input into the support vector description model. The squared distance from the depth feature vector to the center of the sphere is calculated as the health baseline distance. This squared distance is obtained by subtracting the depth feature vector from the center of the sphere element by element and then summing the squares of each component. An alarm is triggered when the confidence level is less than the preset confidence threshold, or the health baseline distance is greater than the critical distance value, and the current sampling time is marked as abnormal.

10. An anomaly detection and fault diagnosis system for wind turbine generator sets, wherein the anomaly detection and fault diagnosis system for wind turbine generator sets is used to implement the anomaly detection and fault diagnosis method for wind turbine generator sets according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire the operating status data of the wind turbine generator at each sampling time, and to standardize the operating status data into a working condition vector. At the same time, a local window is set with the sampling time as the center, and the vibration signal waveform data within the local window is acquired as the vibration analysis signal at that time. The data processing module is used to perform continuous wavelet transform and synchronous compression on the vibration analysis signal to generate a time-frequency diagram of the current sampling time; the standardized working condition vector is input into a pre-trained feature dynamic weight network to obtain the weight vector corresponding to the current sampling time, and the weight vector is divided into feature weight vector and vibration frequency band weight vector according to the dimension. The image generation module is used to standardize the time-frequency map; based on the standardized time-frequency map, feature weight vector, and vibration frequency band weight vector, it generates a dynamic feature map for the current sampling time; and inputs the dynamic feature map into the pre-trained main diagnostic model to obtain the corresponding depth feature vector and fault probability distribution. The anomaly detection module is used to calculate the prediction confidence level based on the fault probability distribution, and simultaneously input the deep feature vector into a pre-constructed support vector description model to calculate the health baseline distance; perform anomaly determination based on the confidence level and the health baseline distance, and trigger an alarm when the warning conditions are met.

Images