Fault detection method and system for blade root bolts of wind turbine blades

Abstract

The present invention provides a fault detection method and system for a blade root bolt of a wind turbine blade, which relates to the technical field of fault detection, and comprises: collecting multi-dimensional operation data of the blade root bolt of the wind turbine blade, executing multi-layer wavelet transform to decompose the multi-dimensional operation data, performing noise reduction processing, dividing the data into multiple time windows, and constructing a time-frequency feature matrix; constructing a hybrid deep learning model, extracting forward time series features and backward time series features of the time-frequency feature matrix and combining them to generate a time series feature graph, extracting a fault feature probability vector and generating a preliminary fault diagnosis result, executing fuzzy rule reasoning, performing correction according to a Markov chain, obtaining a fault development trend prediction result and determining the fault level, and generating early warning information.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to a method and system for detecting faults of blade root bolts of a wind turbine blade. Background Art

[0002] Wind turbine blades are key components of wind turbines. Their reliability directly affects the power generation efficiency and service life of wind turbines. Blade root bolts, as important connectors between wind turbine blades and hubs, are exposed to complex alternating loads for a long time and are prone to fatigue damage, loosening and other failures.

[0003] Traditional wind turbine blade root bolt fault detection mainly relies on manual inspection and regular maintenance, which has problems such as low detection efficiency, high labor cost, and long detection cycle. However, most existing fault detection methods are based on a single signal processing or machine learning method, with limited diagnostic and prediction accuracy, which is difficult to meet the actual operation and maintenance needs of wind farms.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for detecting faults of blade root bolts of a wind turbine blade, which can at least solve some of the problems existing in the prior art.

[0006] According to a first aspect of an embodiment of the present invention, a method for detecting a fault of a blade root bolt of a wind turbine blade is provided, comprising:

[0007] Collect multidimensional operation data of blade root bolts of wind turbine blades, calculate signal variance ratios of the multidimensional operation data at different scales, determine the optimal decomposition layer number of wavelet transform according to the signal variance ratio, perform multi-layer wavelet transform to decompose the multidimensional operation data to obtain low-frequency coefficients and high-frequency coefficients, perform denoising on the high-frequency coefficients through a soft threshold function, calculate the optimal threshold parameters in combination with the Bayesian criterion and reconstruct the high-frequency coefficients to obtain denoised operation data, divide the denoised operation data into multiple time windows according to preset time intervals, perform Fourier transform on the operation data in each time window, obtain frequency domain features, and construct a time-frequency feature matrix by combining the denoised operation data;

[0008] Construct a hybrid deep learning model of a bidirectional neural network and an attention convolutional network, input the time-frequency feature matrix into the bidirectional neural network, extract the forward time series features and backward time series features of the time-frequency feature matrix and combine them to generate a time series feature graph, set an attention weight coefficient for the time series feature graph, generate a weighted time series feature graph and add it to the attention convolutional network, extract local features through convolution operations and generate a convolutional feature graph, process the convolutional feature graph through a fully connected layer to obtain a fault feature probability vector, construct a fault diagnosis decision tree and update the threshold parameters corresponding to each node in the fault diagnosis decision tree in combination with a particle swarm optimization algorithm, add the fault feature probability vector to the fault diagnosis decision tree and generate a preliminary fault diagnosis result;

[0009] A fuzzy neural network verification model is constructed, the preliminary fault diagnosis result and the denoised operating data are input into the fuzzy neural network verification model, the membership function parameters of the fuzzy neural network verification model are optimized by a genetic algorithm, fuzzy rule reasoning is performed based on the optimized membership function parameters, the preliminary fault diagnosis result is verified, and a verified fault diagnosis result is obtained, an improved grey prediction model is constructed, the verified fault diagnosis result is input into the improved grey prediction model to generate a first prediction result, the first prediction result is corrected according to a Markov chain, a fault development trend prediction result is obtained and a fault level is determined, warning information is generated, and the warning information and the time-frequency feature matrix are stored in a fault feature database.

[0010] In an optional embodiment,

[0011] Collect multidimensional operation data of the root bolts of the wind turbine blades, calculate the signal variance ratio of the multidimensional operation data at different scales, determine the optimal decomposition layer number of wavelet transform according to the signal variance ratio, perform multi-layer wavelet transform to decompose the multidimensional operation data, obtain low-frequency coefficients and high-frequency coefficients, perform denoising on the high-frequency coefficients through a soft threshold function, calculate the optimal threshold parameter in combination with the Bayesian criterion and reconstruct the high-frequency coefficients to obtain denoised operation data, divide the denoised operation data into multiple time windows according to a preset time interval, perform Fourier transform on the operation data in each time window, obtain frequency domain features and combine them with the denoised operation data to construct a time-frequency feature matrix, including:

[0012] Collecting multi-dimensional operation data of a blade root bolt of a wind turbine blade, wherein the multi-dimensional operation data includes vibration acceleration data, strain data and temperature data, installing a first acceleration sensor, a second acceleration sensor, a first strain sensor, a second strain sensor and a temperature sensor on the blade root bolt, cross-validating the vibration acceleration data collected by the first acceleration sensor and the second acceleration sensor, cross-validating the strain data collected by the first strain sensor and the second strain sensor, eliminating abnormal data, and obtaining verified multi-dimensional operation data;

[0013] Calculating the autocorrelation coefficient and the cross-correlation coefficient of the verified multidimensional operation data, determining the periodicity characteristics of the vibration acceleration data according to the autocorrelation coefficient, determining the correlation characteristics between the vibration acceleration data and the rotational speed parameter and the wind speed parameter according to the cross-correlation coefficient, determining a time scale combination based on the periodicity characteristics and the correlation characteristics, calculating the variance ratio of the verified multidimensional operation data at different time scales according to the time scale combination, and determining the optimal decomposition layer number of the wavelet transform by analyzing the variation trend of the variance ratio;

[0014] Perform wavelet transform decomposition on the verified multidimensional operation data according to the optimal decomposition layer number to obtain low-frequency coefficients and high-frequency coefficients, perform noise reduction processing on the high-frequency coefficients using a soft threshold function, calculate the optimal threshold parameter by minimizing the Stein unbiased risk estimation criterion, repeatedly perform wavelet transform and threshold processing on the verified multidimensional operation data after time shift and average the processing results, reconstruct the high-frequency coefficients based on the Bayesian criterion to obtain the noise-reduced multidimensional operation data;

[0015] The denoised multidimensional operation data is divided into a plurality of time windows according to a preset time interval, a window function is applied to the denoised multidimensional operation data in each time window, a short-time Fourier transform is performed on the data after the window function is applied, a frequency domain feature is obtained, and the frequency domain feature is resampled in the angle domain to obtain a resampled frequency domain feature;

[0016] The wavelet energy features and singular value features of the denoised multidimensional operating data are extracted, and the statistical features of the denoised multidimensional operating data are extracted. The resampled frequency domain features, the wavelet energy features, the singular value features and the statistical features are combined to construct a time-frequency feature matrix. The time-frequency feature matrix is ​​normalized and dimensionally reduced to obtain a dimensionally reduced time-frequency feature matrix.

[0017] In an optional embodiment,

[0018] The optimal threshold parameter is calculated by minimizing the Stein unbiased risk estimation criterion as shown in the following formula:

[0019] ;

[0020] Among them, λ opt represents the optimal threshold parameter, represents the estimated value of the wavelet coefficient after denoising, represents the wavelet coefficient of the original signal, d represents the data dimension, N represents the data length, Represents the median absolute deviation of the wavelet coefficients, which is used to estimate the noise variance.

[0021] In an optional embodiment,

[0022] Constructing a hybrid deep learning model of a bidirectional neural network and an attention convolutional network, inputting the time-frequency feature matrix into the bidirectional neural network, extracting the forward time series features and the backward time series features of the time-frequency feature matrix and combining them to generate a time series feature graph, setting an attention weight coefficient for the time series feature graph, generating a weighted time series feature graph and adding it to the attention convolutional network, extracting local features through convolution operations and generating a convolutional feature graph, processing the convolutional feature graph through a fully connected layer, obtaining a fault feature probability vector, constructing a fault diagnosis decision tree and updating the threshold parameters corresponding to each node in the fault diagnosis decision tree in combination with a particle swarm optimization algorithm, adding the fault feature probability vector to the fault diagnosis decision tree and generating a preliminary fault diagnosis result, including:

[0023] Constructing a hybrid deep learning model, the hybrid deep learning model includes a bidirectional neural network and an attention convolutional neural network, wherein the attention convolutional neural network includes a multi-head attention calculation unit, a first convolution block, a second convolution block, a third convolution block, an attention pooling layer and a multi-layer perceptron;

[0024] Input the time-frequency feature matrix into the bidirectional neural network, the forward neural network of the bidirectional neural network takes the time-frequency feature matrix as input, the backward neural network of the bidirectional neural network takes the matrix of the time-frequency feature matrix arranged in reverse order of time steps as input, the forward neural network outputs the forward final hidden state, the backward neural network outputs the backward final hidden state, the forward final hidden state and the backward final hidden state are concatenated in the feature dimension to obtain a bidirectional hidden state matrix, and the bidirectional hidden state matrix is ​​mapped into a time series feature map through a fully connected layer;

[0025] Perform multi-head attention calculation on the temporal feature graph, calculate the correlation weights of different time steps through multiple parallel scaled dot product attention heads, each scaled dot product attention head performs query matrix transformation, key matrix transformation and value matrix transformation on the temporal feature graph, splice the outputs of multiple scaled dot product attention heads in the feature dimension and transform them through a fully connected layer to obtain a self-attention temporal feature graph;

[0026] Inputting the self-attention temporal feature map into the first convolution block, the second convolution block and the third convolution block in sequence, wherein the first convolution block, the second convolution block and the third convolution block each include a one-dimensional causal convolution layer, a batch normalization layer and a ReLU activation layer, to obtain a multi-scale local feature map;

[0027] The multi-scale local feature map is input into the attention pooling layer, the attention pooling layer performs a convolution operation on the multi-scale local feature map through a first convolution kernel and a second convolution kernel, and obtains a first significance weight vector and a second significance weight vector through a nonlinear transformation of the first fully connected layer and the second fully connected layer, the first significance weight vector and the second significance weight vector are added and normalized to obtain an attention weight vector, and the local feature vectors in the multi-scale local feature map are weighted and summed according to the attention weight vector to obtain a global key feature vector;

[0028] The global key feature vector is input into the multilayer perceptron, the multilayer perceptron comprises a first hidden layer, a second hidden layer and an output layer, the output layer outputs a fault feature probability vector through a Softmax activation function, a fault diagnosis decision tree is constructed based on the prior knowledge of a fault maintenance expert, the fault feature probability vector is input into the fault diagnosis decision tree, and a path search is performed according to the threshold parameters and branch probability parameters of each non-leaf node to obtain a preliminary diagnosis result;

[0029] A joint optimization model including the threshold parameters and the branch probability parameters is constructed, and the joint optimization model is solved by using a particle swarm optimization algorithm to obtain an optimal parameter configuration of the fault diagnosis decision tree.

[0030] In an optional embodiment,

[0031] Constructing a joint optimization model including the threshold parameter and the branch probability parameter, solving the joint optimization model using a particle swarm optimization algorithm, and obtaining the optimal parameter configuration of the fault diagnosis decision tree includes:

[0032] Collecting equipment operation data, performing feature extraction and standardization processing on the equipment operation data, and obtaining a training sample set including fault features and fault labels, wherein each training sample includes a multi-dimensional fault feature vector and a corresponding fault type label;

[0033] Constructing a joint optimization model, wherein the joint optimization model includes a decision tree threshold parameter and a branch probability parameter, takes the cross entropy loss as the optimization target, adds a regularization constraint on the complexity of the decision tree, and sets a confidence lower bound constraint on the search path probability of each training sample in the decision tree;

[0034] Initialize the global optimal ant lion position corresponding to the ant population and the current ant population, each of the ants contains a threshold variable and a branch probability variable, the threshold variable is encoded by real numbers and has a value range of 0 to 1, the branch probability variable is encoded by discrete numbers and meets the normalization requirement, perform a decoding operation on each ant in the ant population, decode the threshold variable and the branch probability variable into the threshold and branch probability of a decision tree, infer the decision tree according to the training sample to obtain a diagnosis result and a search path, calculate the average cross entropy loss, confidence violation and complexity regularization term, and perform weighted summation to obtain the fitness value of the ant;

[0035] The historical optimal position of the ant is updated according to the fitness value, the historical optimal position is compared with the current global optimal antlion position and the global optimal antlion position is updated, the threshold variable is updated, and the threshold vector of the previous iteration is added to the difference product of the antlion vector and the ant vector corresponding to the global optimal antlion position to obtain the threshold vector of the current iteration, where the coefficient of the difference product is a random number ranging from 0 to 2;

[0036] The branch probability variable is discretely updated, a fitness ratio strategy is used for selection operation, arithmetic mean is used for crossover operation, and uniform random perturbation is used for mutation operation. The operation is repeated until the change in the fitness value is less than a preset threshold or the average distance between the ant lion and the ant is less than a preset threshold. The global optimal ant lion position obtained by the last execution is decoded into the optimal threshold parameter and the optimal branch probability parameter of the decision tree.

[0037] In an optional embodiment,

[0038] Constructing a fuzzy neural network verification model, inputting the preliminary fault diagnosis result and the noise-reduced operation data into the fuzzy neural network verification model, and optimizing the membership function parameters of the fuzzy neural network verification model by genetic algorithm includes:

[0039] Obtaining a 16-dimensional preliminary diagnosis result vector and a 16-dimensional noise-reduced operating data feature vector of the rotating machinery, and inputting them into a pre-built fuzzy neural network verification model, wherein the fuzzy neural network verification model comprises an input layer, a fuzzification layer, a rule layer and an output layer, and the input layer comprises 32 neurons;

[0040] The neurons in the input layer transmit the 16-dimensional preliminary diagnosis result vector and the 16-dimensional noise-reduced operation data feature vector to the fuzzification layer. The fuzzification layer uses a Gaussian membership function to fuzzify the input data. Each input variable corresponds to three language values: low, medium, and high. The fuzzification layer transmits the fuzzified data to the rule layer. The rule layer sets a number of fuzzy rules according to a pre-acquired expert knowledge base. Each fuzzy rule corresponds to a neuron. The neurons in the rule layer transmit the reasoning results to the output layer. The output layer includes 16 neurons, which respectively represent the membership of 16 fault modes.

[0041] Genetic algorithm is used to optimize the mean and variance parameters of Gaussian membership function. The membership function parameters are encoded as a real number vector as the individuals of genetic algorithm. A population of 50 individuals is initialized, each of which contains 192 genes corresponding to 96 membership functions. The diagnostic accuracy of fuzzy neural network on the training set is used as the fitness function. Roulette wheel selection, scatter crossover and Gaussian mutation operators are used for population evolution. The crossover probability is 0.8 and the mutation probability is 0.1. After 300 generations of evolution, the optimal membership function parameters are obtained.

[0042] In an optional embodiment,

[0043] Performing fuzzy rule reasoning based on the optimized membership function parameters, verifying the preliminary fault diagnosis result, obtaining a verified fault diagnosis result, constructing an improved grey prediction model, inputting the verified fault diagnosis result into the improved grey prediction model to generate a first prediction result, correcting the first prediction result according to the Markov chain, obtaining a fault development trend prediction result and determining a fault level, generating early warning information, and storing the early warning information and the time-frequency feature matrix in a fault feature database includes:

[0044] Based on the optimal membership function parameters obtained in advance, the fuzzy neural network is reconfigured, and the preliminary diagnosis results of the samples to be tested and the denoised operating data feature vectors are input into the trained fuzzy neural network. Through membership synthesis and fuzzy reasoning, the output vectors representing the possibility of 16 fault modes are obtained, and the one with the greatest possibility is taken as the verified diagnosis result. For each fault mode, an improved grey prediction model with adaptive background value and dynamic time lag coefficient is introduced, and the fault time series is cumulatively generated to obtain a monotonically increasing first-order cumulative generation sequence, and the background value sequence is calculated, wherein the background value is determined by the weighted average of the current value of the first-order cumulative generation sequence and the previous several step values. The grey differential equation whitening value and grey action are estimated by the least squares method, and the prediction model is restored and the cumulative subtraction operation is performed to obtain the prediction value of the original sequence as the first prediction result;

[0045] Based on the first prediction result, a discrete-time Markov chain model is introduced to correct the prediction deviation, and the fault level is divided into four states: normal, mild, moderate, and severe, which are numbered 1, 2, 3, and 4 respectively. The state transfer probability matrix is ​​obtained according to historical data statistics, and the state distribution vector of the first prediction result is right-multiplied by the transfer probability matrix to obtain the corrected state distribution. The state with the maximum probability is taken as the final predicted fault level, and the weighted average is calculated as the fault degree indicator to generate corresponding warning information. The warning information and the time-frequency feature matrix are stored in the fault feature database.

[0046] A second aspect of an embodiment of the present invention provides a fault detection system for a blade root bolt of a wind turbine blade, comprising:

[0047] The first unit is used to collect multi-dimensional operation data of blade root bolts of wind turbine blades, calculate signal variance ratios of the multi-dimensional operation data at different scales, determine the optimal decomposition layer number of wavelet transform according to the signal variance ratio, perform multi-layer wavelet transform to decompose the multi-dimensional operation data to obtain low-frequency coefficients and high-frequency coefficients, perform noise reduction processing on the high-frequency coefficients through a soft threshold function, calculate the optimal threshold parameters in combination with the Bayesian criterion and reconstruct the high-frequency coefficients to obtain noise-reduced operation data, divide the noise-reduced operation data into multiple time windows according to preset time intervals, perform Fourier transform on the operation data in each time window, obtain frequency domain features, and combine them with the noise-reduced operation data to construct a time-frequency feature matrix;

[0048] The second unit is used to construct a hybrid deep learning model of a bidirectional neural network and an attention convolutional network, input the time-frequency feature matrix into the bidirectional neural network, extract the forward time series features and backward time series features of the time-frequency feature matrix and combine them to generate a time series feature graph, set the attention weight coefficient for the time series feature graph, generate a weighted time series feature graph and add it to the attention convolutional network, extract local features through convolution operation and generate a convolution feature graph, process the convolution feature graph through a fully connected layer to obtain a fault feature probability vector, construct a fault diagnosis decision tree and update the threshold parameters corresponding to each node in the fault diagnosis decision tree in combination with a particle swarm optimization algorithm, add the fault feature probability vector to the fault diagnosis decision tree and generate a preliminary fault diagnosis result;

[0049] The third unit is used to construct a fuzzy neural network verification model, input the preliminary fault diagnosis result and the denoised operating data into the fuzzy neural network verification model, optimize the membership function parameters of the fuzzy neural network verification model through a genetic algorithm, perform fuzzy rule reasoning based on the optimized membership function parameters, verify the preliminary fault diagnosis result, obtain a verified fault diagnosis result, construct an improved grey prediction model, input the verified fault diagnosis result into the improved grey prediction model to generate a first prediction result, correct the first prediction result according to the Markov chain, obtain a fault development trend prediction result and determine the fault level, generate warning information, and store the warning information and the time-frequency feature matrix in a fault feature database.

[0050] According to a third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] A fourth aspect of the embodiments of the present invention is:

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0057] In the present invention, the collected multi-dimensional operating data is denoised by combining wavelet transform and Bayesian criterion, which effectively removes redundant information and random interference in the data and improves data quality. The advantages of bidirectional neural network in extracting time series features and attention convolution network in extracting local features are fully utilized to improve the accuracy and comprehensiveness of fault feature extraction. A fault diagnosis decision tree is introduced and the decision tree parameters are optimized in combination with particle swarm optimization algorithm, which improves the accuracy and efficiency of fault diagnosis. An improved grey prediction model is constructed and combined with Markov chain to predict the development trend of faults, so that potential faults can be discovered early and preventive measures can be taken to reduce the probability of faults. In summary, the present invention realizes accurate fault diagnosis and trend prediction of wind turbine blade root bolts through the organic combination of hybrid deep learning model and improved grey prediction model, thereby improving the reliability and economy of wind turbines. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic flow chart of a method for detecting a fault of a blade root bolt of a wind turbine blade according to an embodiment of the present invention;

[0059] Figure 2 The figure is a schematic diagram of the structure of a fault detection system for a blade root bolt of a wind turbine blade according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 FIG. 1 is a flow chart of a method for detecting a fault of a blade root bolt of a wind turbine blade according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] S1. Collect multi-dimensional operation data of blade root bolts of wind turbine blades, calculate signal variance ratios of the multi-dimensional operation data at different scales, determine the optimal decomposition layer number of wavelet transform according to the signal variance ratio, perform multi-layer wavelet transform to decompose the multi-dimensional operation data, obtain low-frequency coefficients and high-frequency coefficients, perform noise reduction processing on the high-frequency coefficients through a soft threshold function, calculate the optimal threshold parameters in combination with the Bayesian criterion and reconstruct the high-frequency coefficients to obtain noise-reduced operation data, divide the noise-reduced operation data into multiple time windows according to preset time intervals, perform Fourier transform on the operation data in each time window, obtain frequency domain features, and combine them with the noise-reduced operation data to construct a time-frequency feature matrix;

[0064] The signal variance ratio refers to the variance ratio between different frequency bands or different components in the signal, which is usually used to measure the distribution characteristics of the signal or analyze the relative strength of noise and signal. The optimal number of decomposition layers refers to the number of decomposition layers selected in multi-scale signal analysis so that the representation of the signal can retain sufficient information while avoiding redundancy caused by excessive decomposition. The soft threshold function is a function commonly used in signal denoising, and its characteristic is to compress the signal value to zero below the threshold and maintain the original amplitude above the threshold.

[0065] In an optional embodiment,

[0066] Collect multidimensional operation data of the root bolts of the wind turbine blades, calculate the signal variance ratio of the multidimensional operation data at different scales, determine the optimal decomposition layer number of wavelet transform according to the signal variance ratio, perform multi-layer wavelet transform to decompose the multidimensional operation data, obtain low-frequency coefficients and high-frequency coefficients, perform denoising on the high-frequency coefficients through a soft threshold function, calculate the optimal threshold parameter in combination with the Bayesian criterion and reconstruct the high-frequency coefficients to obtain denoised operation data, divide the denoised operation data into multiple time windows according to a preset time interval, perform Fourier transform on the operation data in each time window, obtain frequency domain features and combine them with the denoised operation data to construct a time-frequency feature matrix, including:

[0067] Collecting multi-dimensional operation data of a blade root bolt of a wind turbine blade, wherein the multi-dimensional operation data includes vibration acceleration data, strain data and temperature data, installing a first acceleration sensor, a second acceleration sensor, a first strain sensor, a second strain sensor and a temperature sensor on the blade root bolt, cross-validating the vibration acceleration data collected by the first acceleration sensor and the second acceleration sensor, cross-validating the strain data collected by the first strain sensor and the second strain sensor, eliminating abnormal data, and obtaining verified multi-dimensional operation data;

[0068] Calculating the autocorrelation coefficient and the cross-correlation coefficient of the verified multidimensional operation data, determining the periodicity characteristics of the vibration acceleration data according to the autocorrelation coefficient, determining the correlation characteristics between the vibration acceleration data and the rotational speed parameter and the wind speed parameter according to the cross-correlation coefficient, determining a time scale combination based on the periodicity characteristics and the correlation characteristics, calculating the variance ratio of the verified multidimensional operation data at different time scales according to the time scale combination, and determining the optimal decomposition layer number of the wavelet transform by analyzing the variation trend of the variance ratio;

[0069] Perform wavelet transform decomposition on the verified multidimensional operation data according to the optimal decomposition layer number to obtain low-frequency coefficients and high-frequency coefficients, perform noise reduction processing on the high-frequency coefficients using a soft threshold function, calculate the optimal threshold parameter by minimizing the Stein unbiased risk estimation criterion, repeatedly perform wavelet transform and threshold processing on the verified multidimensional operation data after time shift and average the processing results, reconstruct the high-frequency coefficients based on the Bayesian criterion to obtain the noise-reduced multidimensional operation data;

[0070] The denoised multidimensional operation data is divided into a plurality of time windows according to a preset time interval, a window function is applied to the denoised multidimensional operation data in each time window, a short-time Fourier transform is performed on the data after the window function is applied, a frequency domain feature is obtained, and the frequency domain feature is resampled in the angle domain to obtain a resampled frequency domain feature;

[0071] The wavelet energy features and singular value features of the denoised multidimensional operating data are extracted, and the statistical features of the denoised multidimensional operating data are extracted. The resampled frequency domain features, the wavelet energy features, the singular value features and the statistical features are combined to construct a time-frequency feature matrix. The time-frequency feature matrix is ​​normalized and dimensionally reduced to obtain a dimensionally reduced time-frequency feature matrix.

[0072] The cross-correlation coefficient is a statistic that measures the similarity between two signals, and its value range is from -1 to 1, which indicates the strength of the linear relationship between the two signals. The Stan unbiased risk estimation criterion is a standard for model selection, which aims to minimize the risk of the estimator while ensuring that the estimation result does not deviate from the true value. The window function is a function used in signal analysis, which is usually applied to the signal before Fourier transform to reduce boundary effects and spectral leakage by windowing. The angular domain resampling refers to improving the resolution or reducing the sampling error by changing the angular sampling method in signal processing or imaging. The singular value feature refers to the eigenvalue of the matrix obtained by singular value decomposition, which has important applications in dimensionality reduction, data compression and signal decomposition.

[0073] Collect multi-dimensional operation data of the root bolt of the wind turbine blade, including vibration acceleration data, strain data and temperature data. Install the first acceleration sensor, the second acceleration sensor, the first strain sensor, the second strain sensor and the temperature sensor on the root bolt. Use the first and second acceleration sensors to collect vibration acceleration data, and perform cross-validation to eliminate abnormal data. Use the first and second strain sensors to collect strain data, and perform cross-validation to eliminate abnormal data. Finally, obtain the verified multi-dimensional operation data, including acceleration, strain and temperature data.

[0074] For example, the acceleration data sequence collected by the first acceleration sensor is [1.2, 1.3, 1.4, 10, 1.5, 1.6], and the acceleration data sequence collected by the second acceleration sensor is [1.1, 1.2, 1.3, 1.4, 1.5, 1.7]. By comparison, it is found that the data 10 collected by the first acceleration sensor at the fourth time point is obviously abnormal, so the data is discarded.

[0075] Calculate the autocorrelation coefficient and cross-correlation coefficient of the verified multidimensional operating data. Determine the periodic characteristics of the vibration acceleration data through the autocorrelation coefficient. For example, if the autocorrelation coefficient peaks after a specific time interval, it indicates that the time interval is the period of the vibration acceleration data. Determine the correlation characteristics between the vibration acceleration data and the rotational speed parameters and wind speed parameters through the cross-correlation coefficient. For example, if the cross-correlation coefficient between the vibration acceleration data and the rotational speed parameters is high, it indicates that there is a strong correlation between the two. Determine the time scale combination based on the periodic characteristics and correlation characteristics. For example, the period of the vibration acceleration data can be used as one of the time scales, and the change period of the rotational speed parameter can be used as another time scale.

[0076] According to the determined time scale combination, calculate the variance ratio of the verified multidimensional operating data at different time scales. For example, calculate the variance of the data at periods T1 and T2, and then calculate the ratio of the two variances. Determine the optimal number of decomposition layers of the wavelet transform by analyzing the changing trend of the variance ratio. For example, if the variance ratio tends to be stable after a certain number of decomposition layers, then this number of layers can be used as the optimal number of decomposition layers. Assuming that the calculated trend of the variance ratio as the number of decomposition layers increases is [10, 5, 2, 1.5, 1.2, 1.1], it can be determined that the optimal number of decomposition layers is 5.

[0077] According to the determined optimal number of decomposition layers, the verified multidimensional operation data is decomposed by multi-layer wavelet transform to obtain low-frequency coefficients and high-frequency coefficients, and the high-frequency coefficients are denoised using a soft threshold function. The optimal threshold parameter is calculated by minimizing the Stein unbiased risk estimation criterion. In order to further improve the denoising effect, the wavelet transform and threshold processing are repeated on the multidimensional operation data after time shift, and the processing results are averaged. For example, the data sequence is shifted backward by one time unit, and the wavelet transform and threshold processing are performed again, and then the two processing results are averaged. The high-frequency coefficients are reconstructed based on the Bayesian criterion to obtain the denoised multidimensional operation data.

[0078] The denoised multidimensional operating data is divided into multiple time windows according to preset time intervals. For example, the data is divided into a time window every 10 minutes. A window function, such as a Hamming window, is applied to the denoised multidimensional operating data in each time window, and a short-time Fourier transform is performed on the data after the window function is applied to obtain frequency domain features. The frequency domain features are resampled in the angle domain to obtain the resampled frequency domain features. For example, the frequency domain features are converted from the original 360-degree equally spaced sampling to 720-degree equally spaced sampling.

[0079] Extract the wavelet energy features and singular value features of the denoised multidimensional running data. Extract the statistical features of the denoised multidimensional running data, such as mean, variance, kurtosis and skewness. Combine the resampled frequency domain features, wavelet energy features, singular value features and statistical features to construct a time-frequency feature matrix. Perform normalization and dimensionality reduction processing on the time-frequency feature matrix, such as principal component analysis, to obtain a time-frequency feature matrix after dimensionality reduction.

[0080] In this embodiment, by installing multiple sensors of the same type on the blade root bolt for cross-validation, abnormal data can be effectively identified and eliminated, and data reliability can be improved. The collection of multi-dimensional data can comprehensively reflect the operating status of the blade root bolt, providing a complete data basis for subsequent analysis. The inherent periodicity of the vibration data is revealed through autocorrelation coefficient analysis, and the time scale combination is determined based on the correlation characteristics, which provides a scientific scale selection basis for wavelet decomposition. In summary, this embodiment significantly improves the data quality and the accuracy of feature extraction through multi-sensor cross-validation, intelligent noise reduction processing and multi-feature fusion analysis, provides a reliable data basis for status monitoring and fault diagnosis of wind turbine blade root bolts, promotes the development of predictive maintenance technology for wind power equipment, and is of great significance to improving the operating reliability of wind power equipment and reducing maintenance costs.

[0081] In an optional embodiment,

[0082] The optimal threshold parameter is calculated by minimizing the Stein unbiased risk estimation criterion as shown in the following formula:

[0083] ;

[0084] Among them, λ opt represents the optimal threshold parameter, represents the estimated value of the wavelet coefficient after denoising, represents the wavelet coefficient of the original signal, d represents the data dimension, N represents the data length, Represents the median absolute deviation of the wavelet coefficients, which is used to estimate the noise variance.

[0085] In this embodiment, the adaptive optimization of threshold parameters is achieved by minimizing the risk estimation criterion, avoiding the subjectivity brought by artificially setting thresholds. By calculating the optimal threshold, noise interference is suppressed to the greatest extent while retaining useful signals. The difference between the wavelet coefficients of the original signal and the wavelet coefficients after denoising is considered to ensure that the essential characteristics of the signal are not excessively filtered out during the denoising process. The threshold calculation method based on mathematical optimization criteria has a strict theoretical basis, ensures the reliability of the results, and reduces the risk of overfitting or underfitting in the denoising process. In summary, this embodiment realizes the adaptive optimization of the signal denoising process, effectively suppresses noise interference while ensuring signal integrity, provides a high-quality data basis for subsequent feature extraction and fault diagnosis, and improves the reliability and accuracy of the entire fault diagnosis system.

[0086] S2. Construct a hybrid deep learning model of a bidirectional neural network and an attention convolutional network, input the time-frequency feature matrix into the bidirectional neural network, extract the forward time series features and backward time series features of the time-frequency feature matrix and combine them to generate a time series feature graph, set an attention weight coefficient for the time series feature graph, generate a weighted time series feature graph and add it to the attention convolutional network, extract local features through convolution operations and generate a convolutional feature graph, process the convolutional feature graph through a fully connected layer to obtain a fault feature probability vector, construct a fault diagnosis decision tree and update the threshold parameters corresponding to each node in the fault diagnosis decision tree in combination with a particle swarm optimization algorithm, add the fault feature probability vector to the fault diagnosis decision tree and generate a preliminary fault diagnosis result;

[0087] The forward time series feature refers to the feature extracted from the starting point of the time series to the future time point, which is usually used to predict future change trends or dynamic patterns. The backward time series feature refers to the feature extracted from the end point of the time series to the past time point, which is usually used to analyze the impact of historical data or restore past trends. The threshold parameter refers to the critical value used to determine whether an event or state occurs in signal processing or machine learning. It plays the role of a dividing standard in the decision-making process.

[0088] In an optional embodiment,

[0089] Constructing a hybrid deep learning model of a bidirectional neural network and an attention convolutional network, inputting the time-frequency feature matrix into the bidirectional neural network, extracting the forward time series features and the backward time series features of the time-frequency feature matrix and combining them to generate a time series feature graph, setting an attention weight coefficient for the time series feature graph, generating a weighted time series feature graph and adding it to the attention convolutional network, extracting local features through convolution operations and generating a convolutional feature graph, processing the convolutional feature graph through a fully connected layer, obtaining a fault feature probability vector, constructing a fault diagnosis decision tree and updating the threshold parameters corresponding to each node in the fault diagnosis decision tree in combination with a particle swarm optimization algorithm, adding the fault feature probability vector to the fault diagnosis decision tree and generating a preliminary fault diagnosis result, including:

[0090] Constructing a hybrid deep learning model, the hybrid deep learning model includes a bidirectional neural network and an attention convolutional neural network, wherein the attention convolutional neural network includes a multi-head attention calculation unit, a first convolution block, a second convolution block, a third convolution block, an attention pooling layer and a multi-layer perceptron;

[0091] Input the time-frequency feature matrix into the bidirectional neural network, the forward neural network of the bidirectional neural network takes the time-frequency feature matrix as input, the backward neural network of the bidirectional neural network takes the matrix of the time-frequency feature matrix arranged in reverse order of time steps as input, the forward neural network outputs the forward final hidden state, the backward neural network outputs the backward final hidden state, the forward final hidden state and the backward final hidden state are concatenated in the feature dimension to obtain a bidirectional hidden state matrix, and the bidirectional hidden state matrix is ​​mapped into a time series feature map through a fully connected layer;

[0092] Perform multi-head attention calculation on the temporal feature graph, calculate the correlation weights of different time steps through multiple parallel scaled dot product attention heads, each scaled dot product attention head performs query matrix transformation, key matrix transformation and value matrix transformation on the temporal feature graph, splice the outputs of multiple scaled dot product attention heads in the feature dimension and transform them through a fully connected layer to obtain a self-attention temporal feature graph;

[0093] Inputting the self-attention temporal feature map into the first convolution block, the second convolution block and the third convolution block in sequence, wherein the first convolution block, the second convolution block and the third convolution block each include a one-dimensional causal convolution layer, a batch normalization layer and a ReLU activation layer, to obtain a multi-scale local feature map;

[0094] The multi-scale local feature map is input into the attention pooling layer, the attention pooling layer performs a convolution operation on the multi-scale local feature map through a first convolution kernel and a second convolution kernel, and obtains a first significance weight vector and a second significance weight vector through a nonlinear transformation of the first fully connected layer and the second fully connected layer, the first significance weight vector and the second significance weight vector are added and normalized to obtain an attention weight vector, and the local feature vectors in the multi-scale local feature map are weighted and summed according to the attention weight vector to obtain a global key feature vector;

[0095] The global key feature vector is input into the multilayer perceptron, the multilayer perceptron comprises a first hidden layer, a second hidden layer and an output layer, the output layer outputs a fault feature probability vector through a Softmax activation function, a fault diagnosis decision tree is constructed based on the prior knowledge of a fault maintenance expert, the fault feature probability vector is input into the fault diagnosis decision tree, and a path search is performed according to the threshold parameters and branch probability parameters of each non-leaf node to obtain a preliminary diagnosis result;

[0096] A joint optimization model including the threshold parameters and the branch probability parameters is constructed, and the joint optimization model is solved by using a particle swarm optimization algorithm to obtain an optimal parameter configuration of the fault diagnosis decision tree.

[0097] The time-frequency feature matrix is ​​a matrix formed by combining the information of the signal in the time domain and the frequency domain, and is often used to describe the energy distribution characteristics of the signal at different frequencies and times. The bidirectional hidden state matrix refers to a matrix obtained by a bidirectional recurrent neural network or a bidirectional long short-term memory network when processing sequence data, which contains the forward and backward information of the sequence. The scaled dot product attention head refers to the weight coefficient calculated by scaling the dot product in the attention mechanism, which is used to emphasize the part of the input sequence that is most relevant to the current task. The non-leaf node refers to an internal node in the tree structure, which does not directly contain the final output result and is usually used to transmit information or further decompose the problem. The branch probability parameter is the branch decision probability of each node in the tree structure, which is used to determine the selection probability of different paths or evaluate the possibility of different branches.

[0098] The collected equipment vibration signal is preprocessed, including denoising, averaging and other operations, and the preprocessed signal is subjected to time-frequency analysis, such as using short-time Fourier transform, to obtain a time-frequency feature matrix. This matrix reflects the energy distribution of the signal in different time periods and different frequency components. For example, a vibration signal with a sampling frequency of 1000 Hz and a sampling time of 1 second can obtain a 513×100 time-frequency feature matrix after short-time Fourier transform, where 513 represents the number of frequency points and 100 represents the number of time points.

[0099] The obtained time-frequency feature matrix is ​​input into the bidirectional neural network, which consists of a forward neural network and a backward neural network. The forward neural network processes the time-frequency feature matrix in normal time order and extracts the forward time series features; the backward neural network processes the time-frequency feature matrix in reverse time order and extracts the backward time series features. Assuming that the number of hidden layer units of the bidirectional neural network is 64, the forward neural network and the backward neural network output a 64×100 final hidden state matrix respectively. The two matrices are concatenated in the feature dimension to obtain a 128×100 bidirectional hidden state matrix. Then, through a fully connected layer, the bidirectional hidden state matrix is ​​mapped to a 32×100 time series feature map.

[0100] Perform multi-head attention calculation on the temporal feature map. The multi-head attention mechanism can capture the correlation between different time steps. Assume that 8 attention heads are used, and each attention head transforms the 32×100 temporal feature map into a query matrix, a key matrix, and a value matrix, each with a dimension of 32×100. Each attention head calculates a 32×100 attention output, and the outputs of the 8 attention heads are concatenated in the feature dimension to obtain a 256×100 matrix. Then, through a fully connected layer, it is transformed into a 32×100 self-attention temporal feature map.

[0101] The self-attention temporal feature map is input into the attention convolutional network. The attention convolutional network consists of three convolutional blocks, an attention pooling layer, and a multi-layer perceptron. Each convolutional block contains a one-dimensional causal convolutional layer, a batch normalization layer, and a ReLU activation layer. For example, the convolution kernel size of the first convolutional block is 3, the number of convolution kernels is 64, and the stride is 1. Then, after the input 32×100 self-attention temporal feature map passes through the first convolutional block, the output is a 64×100 feature map. The three convolutional blocks are processed in sequence to obtain multi-scale local feature maps.

[0102] The attention pooling layer performs weighted summation on the multi-scale local feature maps to extract global key features. Assuming that the attention pooling layer uses two convolution kernels of size 3 and 5, two saliency weight vectors are obtained after convolution, fully connected layer and nonlinear transformation. These two vectors are added and normalized to obtain the final attention weight vector. Based on this vector, the multi-scale local feature maps are weighted summed to obtain a global key feature vector of fixed dimension, for example, dimension 128.

[0103] The global key feature vector is input into the multilayer perceptron. The multilayer perceptron contains two hidden layers and one output layer. Assume that the number of units in the two hidden layers is 64 and 32 respectively, and the number of units in the output layer is 5, representing 5 different fault types. The output layer uses the Softmax activation function to output a 5-dimensional fault feature probability vector, indicating the probability of each fault type. For example, if the output vector is [0.1, 0.05, 0.7, 0.1, 0.05], it means that the probability of the third fault type is the highest.

[0104] A fault diagnosis decision tree is constructed based on the prior knowledge of fault maintenance experts. The fault feature probability vector is input into the decision tree, and a path search is performed based on the threshold parameters and branch probability parameters of each non-leaf node to obtain a preliminary diagnosis result. For example, the root node of the decision tree branches based on whether the probability of the first fault type is greater than 0.2. If it is greater than 0.2, it enters the left subtree, otherwise it enters the right subtree.

[0105] The particle swarm optimization algorithm is used to optimize the threshold parameters and branch probability parameters of the decision tree to improve the diagnostic accuracy. For example, a group of particles is initialized, each particle represents a set of parameter configurations. By iteratively updating the position and speed of the particles, the optimal parameter configuration is found to maximize the diagnostic accuracy.

[0106] In this embodiment, the advantages of the bidirectional neural network and the attention convolutional neural network are combined to improve the feature extraction ability of the model. The bidirectional processing mechanism captures the forward and backward dependencies of the time series data. The feature splicing and transformation operations enhance the model's ability to capture long-range dependencies. The two-way convolution and nonlinear transformation improve the recognition ability of significant features. In summary, this embodiment significantly improves the accuracy and reliability of fault diagnosis through multi-level feature extraction and multi-angle feature fusion, realizes the unity of intelligent fault diagnosis and explainability, and provides a new technical path for the intelligent operation and maintenance of wind power equipment. It has important practical value for improving equipment operation reliability and reducing maintenance costs.

[0107] In an optional embodiment,

[0108] Constructing a joint optimization model including the threshold parameter and the branch probability parameter, solving the joint optimization model using a particle swarm optimization algorithm, and obtaining the optimal parameter configuration of the fault diagnosis decision tree includes:

[0109] Collecting equipment operation data, performing feature extraction and standardization processing on the equipment operation data, and obtaining a training sample set including fault features and fault labels, wherein each training sample includes a multi-dimensional fault feature vector and a corresponding fault type label;

[0110] Constructing a joint optimization model, wherein the joint optimization model includes a decision tree threshold parameter and a branch probability parameter, takes the cross entropy loss as the optimization target, adds a regularization constraint on the complexity of the decision tree, and sets a confidence lower bound constraint on the search path probability of each training sample in the decision tree;

[0111] Initialize the global optimal ant lion position corresponding to the ant population and the current ant population, each of the ants contains a threshold variable and a branch probability variable, the threshold variable is encoded by real numbers and has a value range of 0 to 1, the branch probability variable is encoded by discrete numbers and meets the normalization requirement, perform a decoding operation on each ant in the ant population, decode the threshold variable and the branch probability variable into the threshold and branch probability of a decision tree, infer the decision tree according to the training sample to obtain a diagnosis result and a search path, calculate the average cross entropy loss, confidence violation and complexity regularization term, and perform weighted summation to obtain the fitness value of the ant;

[0112] The historical optimal position of the ant is updated according to the fitness value, the historical optimal position is compared with the current global optimal antlion position and the global optimal antlion position is updated, the threshold variable is updated, and the threshold vector of the previous iteration is added to the difference product of the antlion vector and the ant vector corresponding to the global optimal antlion position to obtain the threshold vector of the current iteration, where the coefficient of the difference product is a random number ranging from 0 to 2;

[0113] The branch probability variable is discretely updated, a fitness ratio strategy is used for selection operation, arithmetic mean is used for crossover operation, and uniform random perturbation is used for mutation operation. The operation is repeated until the change in the fitness value is less than a preset threshold or the average distance between the ant lion and the ant is less than a preset threshold. The global optimal ant lion position obtained by the last execution is decoded into the optimal threshold parameter and the optimal branch probability parameter of the decision tree.

[0114] The confidence lower bound constraint refers to the minimum confidence requirement set in machine learning or optimization problems to ensure the stability or credibility of model predictions. The confidence violation refers to the degree to which the actual confidence is lower than the preset confidence threshold in model decision-making, which is usually used to measure model uncertainty or the risk of wrong decisions. The fitness proportional strategy refers to adjusting the decision or selection strategy according to the proportional relationship between the individual's fitness and the target in evolutionary algorithms or reinforcement learning.

[0115] Collect the operating data of the target device. For example, collect the operating data of a certain type of wind turbine, including wind speed, rotation speed, temperature, vibration and other sensor data.

[0116] The collected equipment operation data is feature extracted and standardized. For example, time domain features (such as mean, variance, peak, etc.) and frequency domain features (such as spectrum energy, characteristic frequency, etc.) are extracted from the operation data of the wind turbine. Then, the extracted features are standardized, for example, using the Z-score standardization method to convert the feature values ​​into a standard normal distribution with a mean of 0 and a standard deviation of 1. Assume that after feature extraction and standardization, a training sample set containing 1,000 samples is obtained, each sample contains a 10-dimensional fault feature vector and a corresponding fault type label, and the fault types are divided into three types: normal, bearing fault, and gearbox fault.

[0117] Construct a joint optimization model. The goal of this model is to minimize the cross entropy loss, add regularization constraints on the complexity of the decision tree, and set a confidence lower bound constraint on the search path probability of each training sample in the decision tree. For example, set the confidence lower bound to 0.8.

[0118] Initialize the antlion population. Assume that the population size is 50, and each antlion contains 10 threshold variables (corresponding to 10-dimensional features) and branch probability variables for each node. The threshold variables are encoded with real numbers, ranging from 0 to 1. The branch probability variables are encoded with discrete numbers, such as one-hot encoding, and meet the normalization requirements, that is, the sum of all branch probabilities for each node is 1.

[0119] Perform a decoding operation on each antlion in the antlion population. Decode the threshold variable and branch probability variable into the threshold and branch probability of the decision tree. Infer the decision tree based on the training sample to obtain the diagnosis result and search path. For example, the search path of a sample is "feature 1>0.5, feature 3<0.2, fault type: bearing fault". Calculate the average cross entropy loss, confidence violation, and complexity regularization term, for example, using the number of nodes in the decision tree as a complexity indicator. The weighted sum of these three items is used to obtain the fitness value of the antlion. The weight coefficient can be adjusted according to the actual situation.

[0120] Update the historical optimal position of the ant lion according to the fitness value, compare the historical optimal position with the current global optimal ant lion position and update the global optimal ant lion position. Update the threshold variable, add the threshold vector of the previous iteration to the difference product of the ant lion vector and the global optimal ant lion vector to get the threshold vector of the current iteration. The coefficient of the difference product is a random number ranging from 0 to 2.

[0121] The branch probability variables are discretely updated, the fitness ratio strategy is used for selection, the arithmetic mean is used for crossover, and the uniform random perturbation is used for mutation. For example, the branch probability vectors of two parent ant lions are selected, and their arithmetic mean is calculated as the branch probability vector of the offspring ant lion. Then, the branch probability vector of the offspring ant lion is uniformly randomly perturbed, for example, the branch probability value is randomly changed with a probability of 0.1.

[0122] Repeat the above steps until the change in fitness value is less than a preset threshold (e.g., 0.001) or the average distance between the ant lion and the global optimal ant lion is less than a preset threshold (e.g., 0.01). The global optimal ant lion position obtained from the last execution is decoded into the optimal threshold parameter and optimal branch probability parameter of the decision tree.

[0123] In this embodiment, by jointly optimizing the threshold parameters and branch probability parameters of the decision tree, the training data can be better fitted, thereby improving the accuracy of fault diagnosis. By adding a lower bound constraint on the confidence level, the diagnosis results of the decision tree can be guaranteed to have a high confidence level, thereby enhancing the robustness of fault diagnosis. The optimal threshold parameters and branch probability parameters are automatically searched, avoiding the tedious process of manual parameter adjustment and simplifying the process of building the fault diagnosis model. In summary, this embodiment provides an intelligent solution for the parameter configuration of the fault diagnosis decision tree, which has important practical significance for improving the performance of the fault diagnosis system. It also provides a technical idea that can be used as a reference for decision tree optimization in other fields.

[0124] S3. Construct a fuzzy neural network verification model, input the preliminary fault diagnosis result and the denoised operating data into the fuzzy neural network verification model, optimize the membership function parameters of the fuzzy neural network verification model through a genetic algorithm, perform fuzzy rule reasoning based on the optimized membership function parameters, verify the preliminary fault diagnosis result, obtain the verified fault diagnosis result, construct an improved grey prediction model, input the verified fault diagnosis result into the improved grey prediction model to generate a first prediction result, correct the first prediction result according to the Markov chain, obtain the fault development trend prediction result and determine the fault level, generate warning information, and store the warning information and the time-frequency feature matrix in a fault feature database.

[0125] The membership function parameter is a function parameter used to describe the degree of membership of a fuzzy set in a fuzzy logic system, and determines the membership of an input value in a fuzzy set. The fuzzy rule reasoning refers to the process of reasoning about input and generating output according to fuzzy rules (such as "if-then" rules) in a fuzzy system, and is usually inferred based on fuzzy logic. The gray prediction model is a model for prediction based on incomplete or uncertain data, and the future trend of the system is estimated by constructing a gray model using gray system theory. The Markov chain is a mathematical model that describes the transfer of system states at discrete time points, wherein the transfer of each state depends only on the current state and is independent of the past state.

[0126] In an optional embodiment,

[0127] Constructing a fuzzy neural network verification model, inputting the preliminary fault diagnosis result and the noise-reduced operation data into the fuzzy neural network verification model, and optimizing the membership function parameters of the fuzzy neural network verification model by genetic algorithm includes:

[0128] Obtaining a 16-dimensional preliminary diagnosis result vector and a 16-dimensional noise-reduced operating data feature vector of the rotating machinery, and inputting them into a pre-built fuzzy neural network verification model, wherein the fuzzy neural network verification model comprises an input layer, a fuzzification layer, a rule layer and an output layer, and the input layer comprises 32 neurons;

[0129] The neurons in the input layer transmit the 16-dimensional preliminary diagnosis result vector and the 16-dimensional noise-reduced operation data feature vector to the fuzzification layer. The fuzzification layer uses a Gaussian membership function to fuzzify the input data. Each input variable corresponds to three language values: low, medium, and high. The fuzzification layer transmits the fuzzified data to the rule layer. The rule layer sets a number of fuzzy rules according to a pre-acquired expert knowledge base. Each fuzzy rule corresponds to a neuron. The neurons in the rule layer transmit the reasoning results to the output layer. The output layer includes 16 neurons, which respectively represent the membership of 16 fault modes.

[0130] Genetic algorithm is used to optimize the mean and variance parameters of Gaussian membership function. The membership function parameters are encoded as a real number vector as the individuals of genetic algorithm. A population of 50 individuals is initialized, each of which contains 192 genes corresponding to 96 membership functions. The diagnostic accuracy of fuzzy neural network on the training set is used as the fitness function. Roulette wheel selection, scatter crossover and Gaussian mutation operators are used for population evolution. The crossover probability is 0.8 and the mutation probability is 0.1. After 300 generations of evolution, the optimal membership function parameters are obtained.

[0131] The fuzzification layer is the process of converting precise numerical values ​​into fuzzy sets in a fuzzy system. Inputs are usually mapped to different fuzzy language values ​​through membership functions. The language values ​​are words or terms used to represent fuzzy variables in a fuzzy system, usually such as "high", "low", "medium", etc., which are used to describe imprecise concepts. The Gaussian membership function is a common membership function, which takes the form of a Gaussian distribution function and is used to describe the law of how the membership of elements in a fuzzy set changes with input. The roulette wheel selection is a commonly used selection strategy. In a genetic algorithm, individuals are randomly selected according to their fitness (or probability). The higher the fitness, the greater the probability of being selected.

[0132] Acquire multi-source sensor data such as vibration signals of rotating machinery and perform preprocessing, such as removing outliers and eliminating trend items. Use signal processing methods such as fast Fourier transform and wavelet packet decomposition to extract 16 characteristic indicators to form a 16-dimensional operating data characteristic vector, including time domain indicators (such as root mean square value, peak factor), frequency domain indicators (such as specific frequency band energy) and time-frequency domain indicators. For example, the extracted characteristic vector can be [0.12, 0.05, ..., 1.58], where each value represents the value of a characteristic indicator. Use traditional diagnostic methods, such as support vector machines and neural networks, to perform preliminary fault diagnosis on rotating machinery, obtain preliminary diagnostic results of 16 fault modes, and form a 16-dimensional preliminary diagnostic result vector. For example, the preliminary diagnostic result vector can be [0.9, 0.1, ..., 0.05], where each value represents the probability of the corresponding fault mode.

[0133] The 16-dimensional preliminary diagnosis result vector obtained above and the 16-dimensional noise-reduced operation data feature vector are combined to form a 32-dimensional input vector. A fuzzy neural network verification model is constructed, which includes an input layer, a fuzzification layer, a rule layer, and an output layer. The input layer contains 32 neurons for receiving a 32-dimensional input vector. The input layer passes the received data to the fuzzification layer.

[0134] The fuzzification layer uses Gaussian membership functions to fuzzify the input data. Each input variable corresponds to three linguistic values: low, medium, and high, so each input variable requires three Gaussian membership functions to describe it. Each Gaussian membership function is determined by two parameters: mean and variance. For example, for the input variable "root mean square value", assuming that its range is 0 to 1, the membership function mean of "low" can be set to 0.2 and variance to 0.1, the membership function mean of "medium" can be set to 0.5 and variance to 0.15, and the membership function mean of "high" can be set to 0.8 and variance to 0.1. The fuzzification layer converts the numerical value of each input variable into three membership values ​​and passes these membership values ​​to the rule layer.

[0135] The rule layer sets several fuzzy rules based on the pre-acquired expert knowledge base. Each fuzzy rule corresponds to a neuron. For example, a rule can be "If the RMS value is high and the energy of a specific frequency band is medium, the possibility of fault mode A is high". The neurons in the rule layer perform fuzzy reasoning based on the output of the fuzzification layer and the preset rules, and pass the reasoning results to the output layer.

[0136] The output layer contains 16 neurons, which represent the membership of 16 fault modes. The output layer summarizes the reasoning results of the rule layer to obtain the final membership of each fault mode. For example, the output of the output layer can be [0.85, 0.1, …, 0.05], where each value represents the membership of the corresponding fault mode.

[0137] Genetic algorithm is used to optimize the mean and variance parameters of Gaussian membership function. All the parameters of membership function are encoded as a real vector as individuals of genetic algorithm. Since there are 32 input variables, each variable corresponds to 3 membership functions, and each membership function has two parameters, mean and variance, each individual contains 32*3*2=192 genes. A population of 50 individuals is initialized, each of which represents a set of membership function parameters. The diagnostic accuracy of fuzzy neural network on the training set is used as fitness function. Roulette selection, scatter crossover and Gaussian mutation operators are used for population evolution. The crossover probability is set to 0.8 and the mutation probability is set to 0.1. The population evolves for 300 generations to obtain the optimal membership function parameters. For example, after genetic algorithm optimization, the mean of the "low" membership function corresponding to the "root mean square value" may become 0.25 and the variance may become 0.12.

[0138] The optimized membership function parameters are applied to the fuzzy neural network verification model, and the fault diagnosis of the rotating machinery is performed again to obtain the final fault diagnosis results.

[0139] In this embodiment, the membership function parameters of the fuzzy neural network are optimized by genetic algorithms, so as to improve the diagnostic accuracy of the fuzzy neural network, thereby more accurately identifying the fault mode of the rotating machinery. Combining the preliminary diagnostic results and the denoised operating data, the model can consider various factors more comprehensively, enhance the robustness of the model, reduce the impact of noise and abnormal data on the diagnostic results, integrate expert knowledge into the fuzzy rules, and use genetic algorithms to automatically optimize the model parameters, thereby realizing the intelligent diagnosis of rotating machinery faults and reducing the reliance on manual experience. In summary, this embodiment not only improves the credibility of the diagnostic results, but also provides richer decision support information through membership output, which has important practical value for improving the scientific nature of equipment maintenance decisions.

[0140] In an optional embodiment,

[0141] Performing fuzzy rule reasoning based on the optimized membership function parameters, verifying the preliminary fault diagnosis result, obtaining a verified fault diagnosis result, constructing an improved grey prediction model, inputting the verified fault diagnosis result into the improved grey prediction model to generate a first prediction result, correcting the first prediction result according to the Markov chain, obtaining a fault development trend prediction result and determining a fault level, generating early warning information, and storing the early warning information and the time-frequency feature matrix in a fault feature database includes:

[0142] Based on the optimal membership function parameters obtained in advance, the fuzzy neural network is reconfigured, and the preliminary diagnosis results of the samples to be tested and the denoised operating data feature vectors are input into the trained fuzzy neural network. Through membership synthesis and fuzzy reasoning, the output vectors representing the possibility of 16 fault modes are obtained, and the one with the greatest possibility is taken as the verified diagnosis result. For each fault mode, an improved grey prediction model with adaptive background value and dynamic time lag coefficient is introduced, and the fault time series is cumulatively generated to obtain a monotonically increasing first-order cumulative generation sequence, and the background value sequence is calculated, wherein the background value is determined by the weighted average of the current value of the first-order cumulative generation sequence and the previous several step values. The grey differential equation whitening value and grey action are estimated by the least squares method, and the prediction model is restored and the cumulative subtraction operation is performed to obtain the prediction value of the original sequence as the first prediction result;

[0143] Based on the first prediction result, a discrete-time Markov chain model is introduced to correct the prediction deviation, and the fault level is divided into four states: normal, mild, moderate, and severe, which are numbered 1, 2, 3, and 4 respectively. The state transfer probability matrix is ​​obtained according to historical data statistics, and the state distribution vector of the first prediction result is right-multiplied by the transfer probability matrix to obtain the corrected state distribution. The state with the maximum probability is taken as the final predicted fault level, and the weighted average is calculated as the fault degree indicator to generate corresponding warning information. The warning information and the time-frequency feature matrix are stored in the fault feature database.

[0144] The adaptive background value refers to the background parameter value adjusted according to the environment or task changes in the model or algorithm, and is often used as an evaluation criterion in the optimization process or reinforcement learning. The dynamic lag coefficient refers to the lag parameter adjusted according to the system state or environmental changes in time series prediction or dynamic system, and is used to process the delay effect in time series data. The first-order cumulative generation sequence is a sequence obtained by performing a single cumulative calculation on the input data, and is usually used for time series analysis, signal processing or modeling of the generation process.

[0145] According to historical fault data and expert experience, 16 fault modes are pre-set, and the fuzzy C-means clustering algorithm is used to cluster the historical data to determine the cluster center of each fault mode. The membership function parameters are optimized using the gradient descent method to minimize the clustering error and obtain the optimal membership function parameters. For example, for the Gaussian membership function, the optimization parameters include the mean and standard deviation.

[0146] The time-frequency feature vector of the sample to be tested is input into a pre-trained fuzzy neural network for preliminary fault diagnosis. The structure of the fuzzy neural network includes an input layer, a fuzzification layer, a rule layer, a defuzzification layer, and an output layer. The number of nodes in the input layer is equal to the dimension of the feature vector, and the number of nodes in the output layer is 16, representing 16 fault modes. The input features are mapped to fuzzy sets through membership functions, and fuzzy rules are used for reasoning to finally obtain an output vector representing the possibility of 16 fault modes. For example, the output vector is [0.1, 0.2, 0.8, ..., 0.05], indicating that the sample is most likely to belong to the third fault mode, and the preliminary diagnosis result is the third fault mode.

[0147] Reconfigure the fuzzy neural network based on the optimal membership function parameters obtained in advance. Input the preliminary diagnosis results of the sample to be tested and the denoised operating data feature vector into the trained fuzzy neural network again. Through membership synthesis and fuzzy reasoning, an output vector representing the possibility of 16 fault modes is obtained. The one with the greatest possibility is taken as the verified diagnosis result. For example, if the output vector is [0.05, 0.1, 0.9, ..., 0.02], the verified diagnosis result is the third fault mode.

[0148] For each fault mode, an improved grey prediction model is established. Taking the third fault mode as an example, the historical data of the fault mode is collected to form a time series. The fault time series is cumulatively generated to obtain a monotonically increasing first-order cumulative generation sequence. For example, the original sequence is [10, 12, 15, 18, 20], and the first-order cumulative generation sequence is [10, 22, 37, 55, 75]. The background value sequence is calculated, where the background value is determined by the weighted average of the current value and the previous step values ​​of the first-order cumulative generation sequence. For example, the current value is 37, the previous two step values ​​are 10 and 22, and the weights are 0.2 and 0.8 respectively, then the background value is 0.2*10+0.8*22=19.6. The whitening value and grey action of the grey differential equation are estimated by the least squares method. The prediction model is restored and the cumulative subtraction operation is performed to obtain the predicted value of the original sequence as the first prediction result. For example, the predicted values ​​for the next three moments are [23, 26, 29].

[0149] Based on the first prediction result, the discrete time Markov chain model is introduced to correct the prediction deviation. The fault level is divided into four states: normal, mild, moderate, and severe, numbered 1, 2, 3, and 4 respectively. The state transition probability matrix is ​​obtained based on historical data statistics. For example, the state transition probability matrix is ​​[[0.8, 0.1, 0.05, 0.05], [0.2, 0.7, 0.08, 0.02], [0.1, 0.1, 0.7, 0.1], [0.05, 0.05, 0.1, 0.8]]. The state distribution vector of the first prediction result is right-multiplied by the transfer probability matrix to obtain the corrected state distribution. Assuming that the state distribution vector corresponding to the first prediction result is [0, 0.1, 0.8, 0.1], the corrected state distribution vector is [0.115, 0.168, 0.592, 0.125]. The state with the maximum probability is taken as the final predicted fault level. In this example, the final predicted fault level is moderate (state 3). Calculate the weighted average as the fault severity index. For example, the fault severity index is 1*0.115+2*0.168+3*0.592+4*0.125=2.811. Generate corresponding warning information, for example, "It is expected that the equipment will have a moderate fault in the next three moments, and the fault severity index is 2.811".

[0150] The warning information and time-frequency feature matrix are stored in the fault feature database for subsequent analysis and fault diagnosis.

[0151] In this embodiment, through the verification of the optimized membership function parameters and the fuzzy neural network, the misdiagnosis rate is effectively reduced and the accuracy of fault diagnosis is improved. The improved grey prediction model and Markov chain correction can be used to predict the development trend and level of faults, providing a basis for taking preventive measures in advance. The early warning information generated by combining the time-frequency feature matrix and the prediction results is more comprehensive and reliable, which helps to timely discover and deal with potential faults and avoid major losses. In summary, this embodiment provides a scientific basis for equipment maintenance decisions, has important practical significance for improving equipment operation reliability and reducing maintenance costs, and its scalability and adaptability make it have broad application prospects.

[0152] Figure 2 FIG. 1 is a schematic diagram of a structure of a fault detection system for a blade root bolt of a wind turbine blade according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0153] The first unit is used to collect multi-dimensional operation data of blade root bolts of wind turbine blades, calculate signal variance ratios of the multi-dimensional operation data at different scales, determine the optimal decomposition layer number of wavelet transform according to the signal variance ratio, perform multi-layer wavelet transform to decompose the multi-dimensional operation data to obtain low-frequency coefficients and high-frequency coefficients, perform noise reduction processing on the high-frequency coefficients through a soft threshold function, calculate the optimal threshold parameters in combination with the Bayesian criterion and reconstruct the high-frequency coefficients to obtain noise-reduced operation data, divide the noise-reduced operation data into multiple time windows according to preset time intervals, perform Fourier transform on the operation data in each time window, obtain frequency domain features, and combine them with the noise-reduced operation data to construct a time-frequency feature matrix;

[0154] The second unit is used to construct a hybrid deep learning model of a bidirectional neural network and an attention convolutional network, input the time-frequency feature matrix into the bidirectional neural network, extract the forward time series features and backward time series features of the time-frequency feature matrix and combine them to generate a time series feature graph, set the attention weight coefficient for the time series feature graph, generate a weighted time series feature graph and add it to the attention convolutional network, extract local features through convolution operation and generate a convolution feature graph, process the convolution feature graph through a fully connected layer to obtain a fault feature probability vector, construct a fault diagnosis decision tree and update the threshold parameters corresponding to each node in the fault diagnosis decision tree in combination with a particle swarm optimization algorithm, add the fault feature probability vector to the fault diagnosis decision tree and generate a preliminary fault diagnosis result;

[0155] The third unit is used to construct a fuzzy neural network verification model, input the preliminary fault diagnosis result and the denoised operating data into the fuzzy neural network verification model, optimize the membership function parameters of the fuzzy neural network verification model through a genetic algorithm, perform fuzzy rule reasoning based on the optimized membership function parameters, verify the preliminary fault diagnosis result, obtain a verified fault diagnosis result, construct an improved grey prediction model, input the verified fault diagnosis result into the improved grey prediction model to generate a first prediction result, correct the first prediction result according to the Markov chain, obtain a fault development trend prediction result and determine the fault level, generate warning information, and store the warning information and the time-frequency feature matrix in a fault feature database.

[0156] According to a third aspect of the embodiments of the present invention,

[0157] An electronic device is provided, comprising:

[0158] processor;

[0159] a memory for storing processor-executable instructions;

[0160] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0161] A fourth aspect of the embodiments of the present invention is:

[0162] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0163] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting a fault of a blade root bolt of a wind turbine blade, characterized in that: include: Collect multidimensional operation data of blade root bolts of wind turbine blades, calculate signal variance ratios of the multidimensional operation data at different scales, determine the optimal decomposition layer number of wavelet transform according to the signal variance ratio, perform multi-layer wavelet transform to decompose the multidimensional operation data to obtain low-frequency coefficients and high-frequency coefficients, perform denoising on the high-frequency coefficients through a soft threshold function, calculate the optimal threshold parameters in combination with the Bayesian criterion and reconstruct the high-frequency coefficients to obtain denoised operation data, divide the denoised operation data into multiple time windows according to preset time intervals, perform Fourier transform on the operation data in each time window, obtain frequency domain features, and construct a time-frequency feature matrix by combining the denoised operation data; Construct a hybrid deep learning model of a bidirectional neural network and an attention convolutional network, input the time-frequency feature matrix into the bidirectional neural network, extract the forward time series features and backward time series features of the time-frequency feature matrix and combine them to generate a time series feature graph, set an attention weight coefficient for the time series feature graph, generate a weighted time series feature graph and add it to the attention convolutional network, extract local features through convolution operations and generate a convolutional feature graph, process the convolutional feature graph through a fully connected layer to obtain a fault feature probability vector, construct a fault diagnosis decision tree and update the threshold parameters corresponding to each node in the fault diagnosis decision tree in combination with a particle swarm optimization algorithm, add the fault feature probability vector to the fault diagnosis decision tree and generate a preliminary fault diagnosis result; A fuzzy neural network verification model is constructed, the preliminary fault diagnosis result and the denoised operating data are input into the fuzzy neural network verification model, the membership function parameters of the fuzzy neural network verification model are optimized by a genetic algorithm, fuzzy rule reasoning is performed based on the optimized membership function parameters, the preliminary fault diagnosis result is verified, and a verified fault diagnosis result is obtained, an improved grey prediction model is constructed, the verified fault diagnosis result is input into the improved grey prediction model to generate a first prediction result, the first prediction result is corrected according to a Markov chain, a fault development trend prediction result is obtained and a fault level is determined, warning information is generated, and the warning information and the time-frequency feature matrix are stored in a fault feature database.

2. The method according to claim 1, characterized in that Collect multidimensional operation data of the root bolts of the wind turbine blades, calculate the signal variance ratio of the multidimensional operation data at different scales, determine the optimal decomposition layer number of wavelet transform according to the signal variance ratio, perform multi-layer wavelet transform to decompose the multidimensional operation data, obtain low-frequency coefficients and high-frequency coefficients, perform denoising on the high-frequency coefficients through a soft threshold function, calculate the optimal threshold parameter in combination with the Bayesian criterion and reconstruct the high-frequency coefficients to obtain denoised operation data, divide the denoised operation data into multiple time windows according to a preset time interval, perform Fourier transform on the operation data in each time window, obtain frequency domain features and combine them with the denoised operation data to construct a time-frequency feature matrix, including: Collecting multi-dimensional operation data of a blade root bolt of a wind turbine blade, wherein the multi-dimensional operation data includes vibration acceleration data, strain data and temperature data, installing a first acceleration sensor, a second acceleration sensor, a first strain sensor, a second strain sensor and a temperature sensor on the blade root bolt, cross-validating the vibration acceleration data collected by the first acceleration sensor and the second acceleration sensor, cross-validating the strain data collected by the first strain sensor and the second strain sensor, eliminating abnormal data, and obtaining verified multi-dimensional operation data; Calculating the autocorrelation coefficient and the cross-correlation coefficient of the verified multidimensional operation data, determining the periodicity characteristics of the vibration acceleration data according to the autocorrelation coefficient, determining the correlation characteristics between the vibration acceleration data and the rotational speed parameter and the wind speed parameter according to the cross-correlation coefficient, determining a time scale combination based on the periodicity characteristics and the correlation characteristics, calculating the variance ratio of the verified multidimensional operation data at different time scales according to the time scale combination, and determining the optimal decomposition layer number of the wavelet transform by analyzing the variation trend of the variance ratio; Perform wavelet transform decomposition on the verified multidimensional operation data according to the optimal decomposition layer number to obtain low-frequency coefficients and high-frequency coefficients, perform noise reduction processing on the high-frequency coefficients using a soft threshold function, calculate the optimal threshold parameter by minimizing the Stein unbiased risk estimation criterion, repeatedly perform wavelet transform and threshold processing on the verified multidimensional operation data after time shift and average the processing results, reconstruct the high-frequency coefficients based on the Bayesian criterion to obtain the noise-reduced multidimensional operation data; The denoised multidimensional operation data is divided into a plurality of time windows according to a preset time interval, a window function is applied to the denoised multidimensional operation data in each time window, a short-time Fourier transform is performed on the data after the window function is applied, a frequency domain feature is obtained, and the frequency domain feature is resampled in the angle domain to obtain a resampled frequency domain feature; The wavelet energy features and singular value features of the denoised multidimensional operating data are extracted, and the statistical features of the denoised multidimensional operating data are extracted. The resampled frequency domain features, the wavelet energy features, the singular value features and the statistical features are combined to construct a time-frequency feature matrix. The time-frequency feature matrix is ​​normalized and dimensionally reduced to obtain a dimensionally reduced time-frequency feature matrix.

3. The method according to claim 2, characterized in that The optimal threshold parameter is calculated by minimizing the Stein unbiased risk estimation criterion as shown in the following formula: ； Among them, λ opt represents the optimal threshold parameter, represents the estimated value of the wavelet coefficient after denoising, represents the wavelet coefficient of the original signal, d represents the data dimension, N represents the data length, Represents the median absolute deviation of the wavelet coefficients, which is used to estimate the noise variance.

4. The method according to claim 1, characterized in that: Constructing a hybrid deep learning model of a bidirectional neural network and an attention convolutional network, inputting the time-frequency feature matrix into the bidirectional neural network, extracting the forward time series features and the backward time series features of the time-frequency feature matrix and combining them to generate a time series feature graph, setting an attention weight coefficient for the time series feature graph, generating a weighted time series feature graph and adding it to the attention convolutional network, extracting local features through convolution operations and generating a convolutional feature graph, processing the convolutional feature graph through a fully connected layer, obtaining a fault feature probability vector, constructing a fault diagnosis decision tree and updating the threshold parameters corresponding to each node in the fault diagnosis decision tree in combination with a particle swarm optimization algorithm, adding the fault feature probability vector to the fault diagnosis decision tree and generating a preliminary fault diagnosis result, including: Constructing a hybrid deep learning model, the hybrid deep learning model includes a bidirectional neural network and an attention convolutional neural network, wherein the attention convolutional neural network includes a multi-head attention calculation unit, a first convolution block, a second convolution block, a third convolution block, an attention pooling layer and a multi-layer perceptron; Input the time-frequency feature matrix into the bidirectional neural network, the forward neural network of the bidirectional neural network takes the time-frequency feature matrix as input, the backward neural network of the bidirectional neural network takes the matrix of the time-frequency feature matrix arranged in reverse order of time steps as input, the forward neural network outputs the forward final hidden state, the backward neural network outputs the backward final hidden state, the forward final hidden state and the backward final hidden state are concatenated in the feature dimension to obtain a bidirectional hidden state matrix, and the bidirectional hidden state matrix is ​​mapped into a time series feature map through a fully connected layer; Perform multi-head attention calculation on the temporal feature graph, calculate the correlation weights of different time steps through multiple parallel scaled dot product attention heads, each scaled dot product attention head performs query matrix transformation, key matrix transformation and value matrix transformation on the temporal feature graph, splice the outputs of multiple scaled dot product attention heads in the feature dimension and transform them through a fully connected layer to obtain a self-attention temporal feature graph; Inputting the self-attention temporal feature map into the first convolution block, the second convolution block and the third convolution block in sequence, wherein the first convolution block, the second convolution block and the third convolution block each include a one-dimensional causal convolution layer, a batch normalization layer and a ReLU activation layer, to obtain a multi-scale local feature map; The multi-scale local feature map is input into the attention pooling layer, the attention pooling layer performs a convolution operation on the multi-scale local feature map through a first convolution kernel and a second convolution kernel, and obtains a first significance weight vector and a second significance weight vector through a nonlinear transformation of the first fully connected layer and the second fully connected layer, the first significance weight vector and the second significance weight vector are added and normalized to obtain an attention weight vector, and the local feature vectors in the multi-scale local feature map are weighted and summed according to the attention weight vector to obtain a global key feature vector; The global key feature vector is input into the multilayer perceptron, the multilayer perceptron comprises a first hidden layer, a second hidden layer and an output layer, the output layer outputs a fault feature probability vector through a Softmax activation function, a fault diagnosis decision tree is constructed based on the prior knowledge of a fault maintenance expert, the fault feature probability vector is input into the fault diagnosis decision tree, and a path search is performed according to the threshold parameters and branch probability parameters of each non-leaf node to obtain a preliminary diagnosis result; A joint optimization model including the threshold parameters and the branch probability parameters is constructed, and the joint optimization model is solved by using a particle swarm optimization algorithm to obtain an optimal parameter configuration of the fault diagnosis decision tree.

5. The method according to claim 4, characterized in that Constructing a joint optimization model including the threshold parameter and the branch probability parameter, solving the joint optimization model using a particle swarm optimization algorithm, and obtaining the optimal parameter configuration of the fault diagnosis decision tree includes: Collecting equipment operation data, performing feature extraction and standardization processing on the equipment operation data, and obtaining a training sample set including fault features and fault labels, wherein each training sample includes a multi-dimensional fault feature vector and a corresponding fault type label; Constructing a joint optimization model, wherein the joint optimization model includes a decision tree threshold parameter and a branch probability parameter, takes the cross entropy loss as the optimization target, adds a regularization constraint on the complexity of the decision tree, and sets a confidence lower bound constraint on the search path probability of each training sample in the decision tree; Initialize the global optimal ant lion position corresponding to the ant population and the current ant population, each of the ants contains a threshold variable and a branch probability variable, the threshold variable is encoded by real numbers and has a value range of 0 to 1, the branch probability variable is encoded by discrete numbers and meets the normalization requirement, perform a decoding operation on each ant in the ant population, decode the threshold variable and the branch probability variable into the threshold and branch probability of a decision tree, infer the decision tree according to the training sample to obtain a diagnosis result and a search path, calculate the average cross entropy loss, confidence violation and complexity regularization term, and perform weighted summation to obtain the fitness value of the ant; The historical optimal position of the ant is updated according to the fitness value, the historical optimal position is compared with the current global optimal antlion position and the global optimal antlion position is updated, the threshold variable is updated, and the threshold vector of the previous iteration is added to the difference product of the antlion vector and the ant vector corresponding to the global optimal antlion position to obtain the threshold vector of the current iteration, where the coefficient of the difference product is a random number ranging from 0 to 2; The branch probability variable is discretely updated, a fitness ratio strategy is used for selection operation, arithmetic mean is used for crossover operation, and uniform random perturbation is used for mutation operation. The operation is repeated until the change in the fitness value is less than a preset threshold or the average distance between the ant lion and the ant is less than a preset threshold. The global optimal ant lion position obtained by the last execution is decoded into the optimal threshold parameter and the optimal branch probability parameter of the decision tree.

6. The method according to claim 1, characterized in that Constructing a fuzzy neural network verification model, inputting the preliminary fault diagnosis result and the noise-reduced operation data into the fuzzy neural network verification model, and optimizing the membership function parameters of the fuzzy neural network verification model by genetic algorithm includes: Obtaining a 16-dimensional preliminary diagnosis result vector and a 16-dimensional noise-reduced operating data feature vector of the rotating machinery, and inputting them into a pre-built fuzzy neural network verification model, wherein the fuzzy neural network verification model comprises an input layer, a fuzzification layer, a rule layer and an output layer, and the input layer comprises 32 neurons; The neurons in the input layer transmit the 16-dimensional preliminary diagnosis result vector and the 16-dimensional noise-reduced operation data feature vector to the fuzzification layer. The fuzzification layer uses a Gaussian membership function to fuzzify the input data. Each input variable corresponds to three language values: low, medium, and high. The fuzzification layer transmits the fuzzified data to the rule layer. The rule layer sets a number of fuzzy rules according to a pre-acquired expert knowledge base. Each fuzzy rule corresponds to a neuron. The neurons in the rule layer transmit the reasoning results to the output layer. The output layer includes 16 neurons, which respectively represent the membership of 16 fault modes. Genetic algorithm is used to optimize the mean and variance parameters of Gaussian membership function. The membership function parameters are encoded as a real number vector as the individuals of genetic algorithm. A population of 50 individuals is initialized, each of which contains 192 genes corresponding to 96 membership functions. The diagnostic accuracy of fuzzy neural network on the training set is used as the fitness function. Roulette wheel selection, scatter crossover and Gaussian mutation operators are used for population evolution. The crossover probability is 0.8 and the mutation probability is 0.

1. After 300 generations of evolution, the optimal membership function parameters are obtained.

7. The method according to claim 1, characterized in that Performing fuzzy rule reasoning based on the optimized membership function parameters, verifying the preliminary fault diagnosis result, obtaining a verified fault diagnosis result, constructing an improved grey prediction model, inputting the verified fault diagnosis result into the improved grey prediction model to generate a first prediction result, correcting the first prediction result according to the Markov chain, obtaining a fault development trend prediction result and determining a fault level, generating early warning information, and storing the early warning information and the time-frequency feature matrix in a fault feature database includes: Based on the optimal membership function parameters obtained in advance, the fuzzy neural network is reconfigured, and the preliminary diagnosis results of the samples to be tested and the denoised operating data feature vectors are input into the trained fuzzy neural network. Through membership synthesis and fuzzy reasoning, the output vectors representing the possibility of 16 fault modes are obtained, and the one with the greatest possibility is taken as the verified diagnosis result. For each fault mode, an improved grey prediction model with adaptive background value and dynamic time lag coefficient is introduced, and the fault time series is cumulatively generated to obtain a monotonically increasing first-order cumulative generation sequence, and the background value sequence is calculated, wherein the background value is determined by the weighted average of the current value of the first-order cumulative generation sequence and the previous several step values. The grey differential equation whitening value and grey action are estimated by the least squares method, and the prediction model is restored and the cumulative subtraction operation is performed to obtain the prediction value of the original sequence as the first prediction result; Based on the first prediction result, a discrete-time Markov chain model is introduced to correct the prediction deviation, and the fault level is divided into four states: normal, mild, moderate, and severe, which are numbered 1, 2, 3, and 4 respectively. The state transfer probability matrix is ​​obtained according to historical data statistics, and the state distribution vector of the first prediction result is right-multiplied by the transfer probability matrix to obtain the corrected state distribution. The state with the maximum probability is taken as the final predicted fault level, and the weighted average is calculated as the fault degree indicator to generate corresponding warning information. The warning information and the time-frequency feature matrix are stored in the fault feature database.

8. A fault detection system for blade root bolts of wind turbine blades, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect multi-dimensional operation data of blade root bolts of wind turbine blades, calculate signal variance ratios of the multi-dimensional operation data at different scales, determine the optimal decomposition layer number of wavelet transform according to the signal variance ratio, perform multi-layer wavelet transform to decompose the multi-dimensional operation data to obtain low-frequency coefficients and high-frequency coefficients, perform noise reduction processing on the high-frequency coefficients through a soft threshold function, calculate the optimal threshold parameters in combination with the Bayesian criterion and reconstruct the high-frequency coefficients to obtain noise-reduced operation data, divide the noise-reduced operation data into multiple time windows according to preset time intervals, perform Fourier transform on the operation data in each time window, obtain frequency domain features, and combine them with the noise-reduced operation data to construct a time-frequency feature matrix; The second unit is used to construct a hybrid deep learning model of a bidirectional neural network and an attention convolutional network, input the time-frequency feature matrix into the bidirectional neural network, extract the forward time series features and backward time series features of the time-frequency feature matrix and combine them to generate a time series feature graph, set the attention weight coefficient for the time series feature graph, generate a weighted time series feature graph and add it to the attention convolutional network, extract local features through convolution operation and generate a convolution feature graph, process the convolution feature graph through a fully connected layer to obtain a fault feature probability vector, construct a fault diagnosis decision tree and update the threshold parameters corresponding to each node in the fault diagnosis decision tree in combination with a particle swarm optimization algorithm, add the fault feature probability vector to the fault diagnosis decision tree and generate a preliminary fault diagnosis result; The third unit is used to construct a fuzzy neural network verification model, input the preliminary fault diagnosis result and the denoised operating data into the fuzzy neural network verification model, optimize the membership function parameters of the fuzzy neural network verification model through a genetic algorithm, perform fuzzy rule reasoning based on the optimized membership function parameters, verify the preliminary fault diagnosis result, obtain a verified fault diagnosis result, construct an improved grey prediction model, input the verified fault diagnosis result into the improved grey prediction model to generate a first prediction result, correct the first prediction result according to the Markov chain, obtain a fault development trend prediction result and determine the fault level, generate warning information, and store the warning information and the time-frequency feature matrix in a fault feature database.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images