Variable working condition blade root bolt degradation diagnosis method and system based on data amplification

Abstract

The invention discloses a variable working condition blade root bolt degradation diagnosis method and system based on data amplification. According to the method, the fault information is extracted from the degraded bolt sample data and fused into the healthy bolt sample data to generate diversified degraded bolt data, so that the problem of data sample shortage is solved; domain expert knowledge is concrete into a diagnosis strategy in a deep learning model, and the interference of fan variable working condition operation on bolt data degradation characteristics is solved; and the difference between the generated data and the real data is eliminated through a decision-making mode of double Loss and double classifiers. Compared with a traditional diagnosis model, the problems that the number of samples is unbalanced and variable working conditions interfere with data features are solved from the two aspects of data and the model. And a new solution is provided for the degradation diagnosis of the blade root bolt.

Description

Technical Field

[0001] The present invention belongs to the field of wind power generation, and specifically relates to a variable operating condition blade root bolt degradation diagnosis method and system based on data amplification. Technical Background

[0002] Against the backdrop of global emissions and carbon reduction, wind power has become one of the fastest growing and most promising renewable energy sources in the world.

[0003] As wind turbines develop towards larger sizes, the stability and reliability of wind turbines have become one of the key factors in ensuring power generation efficiency and safe operation. During wind turbine operation, due to changes in wind speed, the force exerted by the blades on the blade roots is unstable when the blades rotate, which affects the stress of the blade root bolts. This unstable force change will interfere with the data during bolt monitoring, obliterating the real degradation signal of the bolts, making it difficult for data analysts to identify and diagnose whether the bolts are deteriorating from the monitoring data.

[0004] At present, deep learning technology has shown great potential in the field of fault diagnosis. It can achieve efficient fault diagnosis by separating data features from complex signals. This data-driven fault diagnosis method provides a new idea for blade root bolt degradation monitoring. However, the actual data samples of blade root bolt degradation are relatively small, and deep learning technology usually relies on a large amount of sample data for model training to achieve good diagnostic results. Secondly, in the actual operation of the wind turbine, the operating conditions fluctuate greatly and there is a lot of interference noise, which further increases the difficulty of diagnosis. Therefore, how to use deep learning technology to achieve efficient and accurate blade root bolt degradation diagnosis under the conditions of unstable operating conditions, strong noise interference and limited number of samples has become a technical problem that needs to be solved in the current field of wind turbine operation and maintenance. Summary of the invention

[0005] The present invention proposes a variable operating condition blade root bolt degradation diagnosis method and system based on data amplification. Aiming at the current situation of the scarcity of degradation data on fan blade root bolts, the invention extracts the degradation signals existing in a small number of samples, superimposes them with the data of healthy samples, and obtains degradation simulation data of different bolts. This is used as sample expansion to provide a certain scale of data for deep learning methods to learn. Secondly, in view of the common phenomenon that the operating conditions change with wind speed during the operation of fans, a deep learning model combined with expert knowledge is designed to alleviate the problem that the degradation characteristics in the data change with the changes in operating conditions. Finally, considering the difference between the generated signal and the real signal, the model iteration process and the final result are optimized by introducing a dual classifier and a dual loss method. The overall process of the present invention is as follows: Figure 1As shown, it includes a data collection module, a data screening module, a data generation module, a data set construction module, a model training module, and a model testing module.

[0006] First, in the data collection module, the preload data of the deteriorated bolts and healthy bolts at the root of the wind turbine blades and the load data of the corresponding blades are collected respectively.

[0007] Secondly, in the data screening module, the data is screened according to the following requirements: for the degraded blade root bolt data, it is required to include the preload monitoring data during the period of the bolt's change from healthy to degraded, that is, the full life cycle data, and then the data under full engine conditions is screened. For the healthy blade root bolt data, the collection time is required to be the same as the degraded data, and the preload data under full engine conditions must also be screened. Finally, in the same time period, these bolt preload data are screened according to the collected blade load data.

[0008] Next, the data of some deteriorated bolts is generated through the data generation module. The process is as follows: Figure 2 As shown in Figure 1, first, after the data screening is completed, the number of blades where the deteriorated bolts meet the conditions is counted and assigned a value K. At the same time, the collected healthy bolt data is also divided into K parts according to the blades where they are located, and FFT conversion is performed. The short-time Fourier transform STFT (K i ) is subtracted from the STFT characteristics of its early healthy data to obtain the degradation characteristic component in the data, and then it is added to the STFT data of the healthy bolt of the corresponding blade to obtain the bolt degradation characteristic data simulated on the healthy bolt. And so on, K groups of simulated bolt degradation characteristic data are obtained to obtain the generated data set.

[0009] Then, the collected real data and generated simulated data are divided in the dataset construction module. The dataset is divided into training set, validation set, and test set according to a certain ratio.

[0010] Then, the model is trained through the model training module, and the divided data set is converted into a time-frequency graph and input into the model. The model is iterated and finally a better result is obtained. It is worth noting that the model introduces some domain expert knowledge in the process of model training to target the unique changing operating conditions of the fan operation. This domain expert knowledge can be explained as experienced engineers who can judge the time interval of stable operation of the fan based on the frequency changes in the time-frequency graph and obtain data features under relatively stable conditions. This invention concretizes this experience into a method of combining convolution with multi-head attention calculation, and introduces it into the process of model diagnosis to improve the diagnostic performance of the model. The process is as follows: Figure 3As shown, at the same time, considering the difference between generated data and real data, the invention also introduces an additional domain classifier to judge whether the data is real data, and controls the update of weight parameters in the iterative process through double Loss.

[0011] Finally, the trained model is tested in the data testing module, that is, the performance of the model is verified through the data in the test set. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 Data amplification-based blade root bolt degradation diagnosis method and system overall process Figure 2 Data generation module flow chart Figure 3 Model training module flow chart DETAILED DESCRIPTION

[0012] The monitoring data of relevant components are collected in the database through the data collection module, including the preload data of the blade root bolts and the blade load data. When collecting data, the wind turbine number where the bolts, blades and other components are located, the data timestamp, and detailed information such as the operating conditions should be marked;

[0013] Furthermore, the data is filtered according to the corresponding requirements in the data screening module to obtain data that meets more detailed conditions. Among them, the bolt preload data should be provided by the bolts at the root of the wind turbine blades, and only the data under the full-load condition is filtered. Then, the bolt data is filtered according to the load data of the wind turbine blade where the bolt is located, that is, the working condition of the bolt is judged by the load state of the blade. When filtering the deteriorated bolt data, the time range should include the whole process of the bolt from healthy operation to complete deterioration.

[0014] Furthermore, the number of blades where the deteriorated bolts are located is counted, represented by K, and then the fault data of K groups is synthesized based on the health data of the corresponding units through the data generation module. The process is as follows:

[0015] Step 1: Get the deterioration characteristics of the bolt;

[0016] F ti =STFT(data ti )-STFT(data (t-Δt)i )

[0017] In the formula

[0018] data represents the data of the deteriorated bolts;

[0019] STFT stands for Short Time Fourier Transform;

[0020] t represents a certain moment in the bolt deterioration process;

[0021] Δt represents the period from initial deterioration to complete deterioration of the bolt;

[0022] i represents the data corresponding to the deteriorated bolts on different blades;

[0023] F ti It means that on the same bolt, the STFT data at a certain moment during the degradation period is subtracted from the STFT data at a time Δt earlier than that moment to obtain the degradation characteristics of the bolt;

[0024] Step 2: Perform normal distribution screening on degradation characteristics and remove outliers;

[0025]

[0026] In the formula

[0027] It indicates the value retained after removing outliers;

[0028] Step 3: Add the degradation characteristics to the normal signal data to obtain the time-frequency spectrum data of the generated signal;

[0029]

[0030] In the formula

[0031] C i represents the generated STFT data obtained;

[0032] H ti Indicates the STFT data of the healthy bolt on the blade at the time corresponding to the deteriorated bolt;

[0033] Step 4: Generate deteriorated bolt dataset;

[0034] S={C it ∣i∈K}

[0035] In the formula

[0036] S represents the generation of deteriorated bolt data set;

[0037] Furthermore, the generated degraded data and the real healthy data are divided into training set, validation set and test set according to a certain ratio. Note that the data in the test set should not contain the generated data, but should be all real data, and the generated data corresponding to the degraded data cannot appear in the training set and validation set.

[0038] Furthermore, the constructed data set is input into the model for training. Specifically, the invention introduces domain expert knowledge in view of the impact of variable operating conditions on data features. That is, experienced engineers will judge the time interval of stable operation of the fan based on the frequency changes in the time-frequency diagram to obtain data features under relatively stable operating conditions. Based on this, a method for extracting stable signal feature weights through convolution is designed to guide the multi-head attention mechanism for diagnosis. First, the input short-time Fourier transform data is converted into a time-frequency diagram.

[0039] Furthermore, the domain expert knowledge is visualized through one-dimensional convolution to extract relatively stable working condition features in the time-frequency graph, which are used as weights to guide the multi-head attention mechanism for diagnosis. The principle is to use the powerful extraction ability of one-dimensional convolution for time-invariant signal features to focus on the features of stable working conditions in the time-frequency graph, that is, to divide the feature map after one-dimensional convolution into regions, and calculate the average value of the features in each region, and finally normalize the average values ​​of all regions as a weight and add it to the calculation process of the multi-head attention. The process is shown below.

[0040] first step:

[0041] Because the input time-frequency graph is in color, the input channel is C = 3. Then one-dimensional convolution is applied to extract features, and the convolution kernel size is k. The process can be expressed as:

[0042]

[0043] In the formula

[0044] F i,j,m Represents the value of the feature map after convolution at position (i, j) and channel m;

[0045] I i,j+n-1,c Represents the value of the input time-frequency graph at position (i, j+n-1) and channel c;

[0046] C represents the number of channels of the input image;

[0047] k represents the size of the convolution kernel;

[0048] W n,c,m Represents the weight of the convolution kernel for position n, input channel c, and output channel m;

[0049] b m Represents the bias of output channel m.

[0050] Step 2: Fuse the channels of the output feature map F to obtain a feature map with a channel number of 1. The process is as follows:

[0051]

[0052] In the formula

[0053] G i,j Represents the value of the fused feature map at position (i, j);

[0054] F i,j,m Represents the value of the original multi-channel feature map at position (i, j) and channel m;

[0055] M represents the number of channels of the original feature map;

[0056] α m Represents the weighting coefficient of channel m.

[0057] Step 3: Divide the fused single-channel feature map G into n windows along the width direction.

[0058] Step 4: Calculate the average value a of the data in each window i i , the process is as follows:

[0059]

[0060] In the formula

[0061] a i represents the average value of the i-th window;

[0062] G x+(i-1)·n,y Represents the pixel value at position (x+(i-1)n,y) in the feature map G;

[0063] n×m represents the size of the window.

[0064] Step 5: Normalize these average values ​​[a1, a2, …, an], the process is as follows:

[0065]

[0066] in

[0067] w i represents the normalized weight of the i-th feature window;

[0068] n represents the number of data;

[0069] Represents the sum of all window means.

[0070] Furthermore, the weight value obtained by one-dimensional convolution is used as the influence of different working conditions on the degradation feature and multiplied by the key K in the multi-head attention calculation process to calculate the contribution of data under different working conditions to the identification of degradation features. The calculation process is as follows:

[0071] Step 1: The time-frequency graph is divided into a series of image blocks, which are then flattened and mapped into a high-dimensional space. These embedding vectors can be expressed as:

[0072] X=[x1,x2,...,xz]

[0073] Step 2: For each input embedding block, generate query Q, key K and value V matrices:

[0074] Q=XW Q ,K=XW K ,V=XW V W Q ,W K ,W V is a learnable weight matrix.

[0075] Step 3: Perform Hadamard product on the weight value W obtained by convolution and the key K to obtain the weighted key Kw K w =W⊙K

[0076] In the formula

[0077] W represents the regional weight value obtained after the time-frequency graph is convolved;

[0078] K represents the key K matrix.

[0079] Step 4: Calculate the dot product between the query and the key to get the attention score, then scale it:

[0080]

[0081] In the formula

[0082] Q represents the query matrix (Query);

[0083] Kw represents the key matrix (Key) with weights K = Q⊙ω (where q is the regional weight value obtained by one-dimensional convolution;

[0084] V represents the value matrix (Value);

[0085] d k Represents the dimension of the key vector, used for scaling to avoid too large dot product values;

[0086] QK w T Represents the dot product of the query and the weighted key matrix to generate the attention score;

[0087] Softmax transforms features into probability distribution.

[0088] Step 5: Split the input into multiple different heads to capture different parts of the information:

[0089] MultiHead(Q,K,V)=Concat(head 1 ,head 2 ,…,head h )W O

[0090] In the formula

[0091] The calculation for each head is:

[0092] head i =Attention(Q i ,K i ,V i )-W O

[0093] W O is the linear transformation matrix of the multi-head output.

[0094] Furthermore, after each multi-head attention calculation, standard layer normalization and multi-layer perceptron processing output features are performed, and then the next layer of multi-head attention calculation process is entered. After multiple cycles, the final feature vector is obtained for the classifier to distinguish.

[0095] Furthermore, in order to solve the difference between generated data and real data, a data classifier is added on the basis of the traditional label classifier, that is, to judge whether the data is generated data according to the feature vector and return a Loss L , and classify the data in the label classifier to obtain the classification results and Loss f , the result is judged by the dual classifier, that is, when the discriminator cannot determine whether the data is generated, the classifier can correctly classify the data and iterate the model weights through the dual loss. The process is as follows:

[0096] Step 1: Use the label classifier to predict the label based on the feature vector f(x) provided by the attention calculation, and the loss function Loss L calculate:

[0097]

[0098] In the formula

[0099] Classifier L It is the network layer for label classification;

[0100] C is the number of categories;

[0101] y i is the actual label;

[0102] is the predicted probability of the i-th class.

[0103] Step 2: According to the feature vector f(x) provided by the attention calculation, the data source is predicted through the data classifier, and the loss function Loss F calculate:

[0104]

[0105] In the formula

[0106] Classifier f is the network layer used for data classification;

[0107] D is the number of data source categories;

[0108] y f is the actual data source label;

[0109] is the predicted probability of the data source label.

[0111] Furthermore, the designed model and the weights saved during training are used to evaluate the test set data to obtain the true effect of the model on the new data, thereby achieving the purpose of monitoring the status of the blade root bolts.

Claims

1. The present invention proposes a variable operating condition blade root bolt degradation diagnosis method and system based on data amplification, which is characterized by: In view of the current situation of scarce data on deterioration of fan blade root bolts, the characteristic signals in the deteriorated bolt samples are extracted and superimposed with the healthy bolt sample signals to obtain degradation simulation signals of different bolts. This is used as sample expansion to provide data for deep learning methods to learn. Secondly, in view of the common phenomenon that operating conditions change with wind speed during fan operation, a deep learning model that extracts and integrates expert knowledge is designed to alleviate the problem that the degradation characteristics in the data change with changes in operating conditions. Finally, considering the difference between the generated signal and the real signal, the model iteration process and final results are optimized by introducing double classifiers and double Loss methods.

2. According to claim 1, it is characterized in that The characteristic signals existing in the deteriorated bolt samples are extracted and superimposed with the healthy bolt sample signals to generate degradation simulation signals of different bolts. This method is established by the following equation: Step 1: Get the deterioration characteristics of the bolts; F ti =STFT(data ti )-STFT(data (t-Δt)i ) In the formula data represents the data of the deteriorated bolts; STFT stands for Fourier transform; t represents a certain moment in the bolt deterioration process; Δt represents the period from initial deterioration to complete deterioration of the bolt; i represents the data corresponding to the deteriorated bolts on different blades; F ti It means that on the same bolt, the STFT data at a certain moment during the degradation period is subtracted from the STFT data at a time Δt earlier than that moment to obtain the degradation characteristics of the bolt; Step 2: Perform normal distribution screening on degradation characteristics and remove outliers; In the formula It indicates the value retained after removing outliers; Step 3: Add the degradation feature to the normal signal frequency domain data to obtain the spectrum data of the generated signal; In the formula C i represents the generated STFT data obtained; H ti Indicates the STFT data of the healthy bolt on the blade at the time corresponding to the deteriorated bolt; Step 4: Generate deteriorated bolt dataset; S={C it ∣i∈K} In the formula S represents the generation of deteriorated bolt data set.

3. According to claim 1, it is characterized in that A deep learning model for extracting expert knowledge is designed, which is established by the following equation: Step 1: Use a one-dimensional convolution kernel to extract time-invariant features in the data: In the formula F i,j,m Represents the value of the feature map after convolution at position (i, j) and channel m; I i,j+n-1,c Represents the value of the input time-frequency graph at position (i, j+n-1) and channel c; C represents the number of channels of the input image; k represents the size of the convolution kernel; W n,c,m Represents the weight of the convolution kernel for position n, input channel c, and output channel m; b m Represents the bias of output channel m. Step 2: Fuse the channels of the output feature map F to obtain a feature map with a channel number of 1. The process is as follows: In the formula G i,j Represents the value of the fused feature map at position (i, j); F i,j,m Represents the value of the original multi-channel feature map at position (i, j) and channel m; M represents the number of channels of the original feature map; α m Represents the weighting coefficient of channel m. Step 3: Divide the fused single-channel feature map G into n windows along the width direction. Step 4: Calculate the average value a of the data in each window i i , the process is as follows: In the formula a i represents the average value of the i-th window; G x+(i-1)·n,y Represents the pixel value at position (x+(i-1)n,y) in the feature map G; n×m represents the size of the window. Step 5: Normalize these average values ​​[a1, a2, …, an] to get the weight values ​​corresponding to different windows. The process is as follows: In the formula w i represents the normalized weight of the i-th feature window; n represents the number of data; Represents the sum of all window means.

4. According to claim 1, it is characterized in that In order to integrate the extracted expert knowledge into the attention calculation process, the weight value obtained by convolution is Hadamard product with the key K to obtain the weighted key Kw. This method is established by the following equation: K w =W⊙K In the formula W represents the regional weight value obtained after the time-frequency graph is convolved; K represents the key K matrix.

5. According to claim 1, it is characterized in that On the basis of the traditional label classifier, a data classifier is added to optimize the model iteration process and the final result through the dual classifier and dual loss method. The method is established through the following equation: Step 1: Use the label classifier to predict the label based on the feature vector f(x) provided by the attention calculation, and the loss function Loss L calculate: In the formula Classifier L is a classifier for label classification; C is the number of categories; y i is the actual label; is the predicted probability of the i-th class. Step 2: According to the feature vector f(x) provided by the attention calculation, the data source is predicted through the data classifier, and the loss function Loss F calculate: In the formula Classifier f is a classifier used to classify data; D is the number of data source categories; y f is the actual data source label; is the predicted probability of the data source label.

Images