Wind turbine bearing fault diagnosis method based on time-frequency domain convolutional network and deep forest

Abstract

The present invention discloses a wind turbine bearing fault diagnosis method based on a time-frequency domain convolutional network and a deep forest, relates to the field of fault diagnosis technology, solves the technical problem of low wind turbine bearing fault diagnosis accuracy under complex working conditions and high noise conditions, and the key points of the technical solution are to add fault labels to the wind turbine bearing vibration data to form a data set, and divide the data set into a training set and a test set; input the training set and the test set data into a time-frequency domain convolutional network to extract deep features; input the extracted deep feature set into the cascade forest part training of the deep forest to form an identity mapping between vibration data and fault labels; input the test set data into the trained model to obtain the corresponding fault type label. The combination of the time-frequency domain convolutional network and the deep forest is conducive to the automatic extraction of features, can effectively extract effective information from the wind turbine bearing features under high noise and complex working conditions, and achieve higher fault diagnosis accuracy.

Description

Technical Field

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a wind turbine bearing fault diagnosis method based on time-frequency domain convolutional networks and deep forests. Background Art

[0002] Due to long-term operation under heavy load and high speed environment, wind turbine bearings have become one of the most prone to failure in wind turbines, so early fault diagnosis of wind turbine bearings is of great significance. However, due to the diversity and randomness of the working environment of wind turbines, as well as the interference of noise signals on vibration signals, wind turbine bearing fault diagnosis has become a difficult problem to solve, which seriously affects the normal operation and maintenance of wind turbines.

[0003] At present, scholars at home and abroad have done a lot of research on the problem of wind turbine bearing fault diagnosis. However, faced with the complex and changeable operation conditions of wind turbines and the existing high noise problems, the accuracy of wind turbine fault diagnosis using conventional machine learning algorithm modeling is low, and the parameter adjustment process of the currently popular deep learning model is too cumbersome and time-consuming, which poses a severe challenge to the safe operation of wind turbines. Summary of the invention

[0004] The present application provides a wind turbine bearing fault diagnosis method based on time-frequency domain convolutional network and deep forest, the technical purpose of which is to improve the fault diagnosis accuracy of wind turbine bearings under complex working conditions and high noise conditions.

[0005] The above technical objectives of this application are achieved through the following technical solutions:

[0006] A wind turbine bearing fault diagnosis method based on time-frequency domain convolutional network and deep forest, comprising:

[0007] S1: Obtain the original vibration signal of the wind turbine bearing and the corresponding working condition, speed, load and fault type label; window the original vibration signal to obtain multiple fault data of a single fault source, preprocess the multiple fault data and add corresponding fault type labels to form a data set, and divide the data set into a training set and a test set;

[0008] S2: inputting the training set and the test set into a time-frequency domain convolutional network for feature extraction to obtain respective deep-level feature sets;

[0009] S3: inputting the deep feature set into a first deep forest model to train the first deep forest model to obtain a second deep forest model;

[0010] S4: inputting the deep feature set of the test set into the second deep forest model to obtain the corresponding fault type label and complete fault diagnosis;

[0011] The first deep forest model includes at least two cascade layers, adjacent cascade layers are connected by autoencoders, and each cascade layer includes a random forest, a completely random forest, an XGboost, and a Gaussian Bayes classifier;

[0012] Inputting the deep feature set into a first deep forest model to train the first deep forest model to obtain a second deep forest model includes:

[0013] S31: inputting the deep feature set into the first cascade layer of the first deep forest model for diagnosis, and obtaining a diagnosis result of the first cascade layer;

[0014] S32: combining the diagnosis result with the deep feature set and inputting the result into an autoencoder to obtain enhanced features;

[0015] S33: inputting the enhanced features to the next cascade layer for diagnosis, and obtaining the diagnosis result of the next cascade layer;

[0016] S34: Repeat steps S32 to S33 until the accuracy of the first deep forest model no longer increases, and the trained model is output as the second deep forest model.

[0017] The beneficial effects of this application are:

[0018] The wind turbine bearing fault diagnosis method based on time-frequency domain convolutional network and deep forest in the present application, in the process of constructing the mapping between bearing vibration data and fault type labels, realizes the acquisition of the global receptive field in the time domain through convolution in the frequency domain, constructs richer deep features, improves the effectiveness of the feature set, and thus achieves the accuracy of wind turbine bearing fault diagnosis, and has strong practicality and high use value.

[0019] At the same time, this application introduces multiple different primary diagnostic models in each layer of the deep forest, and uses autoencoders to map the parameters of the intermediate links to the same space, which can not only avoid the drawbacks of single sub-model diagnosis but also speed up the intermediate training speed and improve the diagnostic efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flowchart of this application;

[0021] Figure 2 This is the time-frequency domain convolutional network structure diagram of this application;

[0022] Figure 3 This is a structural diagram of a deep forest model according to an embodiment of the present application;

[0023] Figure 4 This is a time-frequency domain convolution internal parameter setting diagram of an embodiment of the present application;

[0024] Figure 5 This is a data type diagram of an embodiment of the present application;

[0025] Figure 6 This is a schematic diagram of wind turbine bearing fault diagnosis results according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] The technical solution of the present application will be described in detail below with reference to the accompanying drawings.

[0027] A wind turbine bearing fault diagnosis method based on time-frequency domain convolutional network and deep forest in this embodiment, see Figure 1 , including the following processes:

[0028] S1: Collect data to obtain the original vibration signal of the wind turbine bearing and the corresponding operating conditions, speed, load and fault type labels; window the original vibration signal to obtain multiple fault data of a single fault source, preprocess the multiple fault data and add corresponding fault type labels to form a data set, and divide the data set into a training set and a test set.

[0029] S2: Feature extraction: input the training set and the test set into the time-frequency domain convolutional network for feature extraction to obtain their respective deep-level feature sets.

[0030] The structure of the time-frequency domain convolutional network is as follows Figure 2 As shown in , the time-frequency domain convolutional network includes time-domain and frequency-domain convolutional layers, an attention mechanism layer, and a convolution-pooling layer; the time-domain and frequency-domain convolutional layers include a time-domain convolution unit, a frequency-domain convolution unit, and an average pooling layer; the convolution-pooling layer includes two consecutive convolution-pooling units and a fully connected layer. The parameters of the time-frequency domain convolutional network can be referred to Figure 4 to set it up.

[0031] As a specific embodiment, the specific feature extraction process is as follows:

[0032] S21: directly convolve the vibration signal through a time domain convolution unit to obtain a time domain sequence with a local receptive field;

[0033] The vibration signal is transformed by DFT through the frequency domain convolution unit to obtain the amplitude signal in the frequency domain, and then the amplitude signal is convolved to obtain the local receptive field frequency domain sequence, and finally the local receptive field frequency domain sequence is transformed by IDFT to obtain the time domain sequence of the global receptive field;

[0034] S22: splicing the local receptive field time domain sequence and the global receptive field time domain sequence, downsampling the spliced ​​time domain sequence through an average pooling layer, and then extracting features using an attention mechanism to obtain preliminary features;

[0035] S23: The preliminary features are subjected to two consecutive convolution-pooling units for deep feature extraction, and finally output through a fully connected layer to obtain the deep-level features.

[0036] S3: Model training, inputting the deep feature set into the first deep forest model to train the first deep forest model to obtain a second deep forest model. The structure of the first deep forest model is as follows Figure 3 As shown, the extracted deep feature set is input into the first layer of the cascade forest of the first deep forest model for diagnosis, and the diagnosis result of the first layer model is combined with the original input feature set and input into the autoencoder to obtain a data set in the same mapping space. The data set is input into the second layer to obtain the corresponding diagnosis result, and then combined with the original input feature set and input into the autoencoder to obtain a new data set and input into the third layer, and so on until the model accuracy no longer increases and the trained second deep forest model is output.

[0037] That is, the training process includes:

[0038] S31: inputting the deep feature set into the first cascade layer of the first deep forest model for diagnosis, and obtaining a diagnosis result of the first cascade layer;

[0039] S32: combining the diagnosis result with the deep feature set and inputting the result into an autoencoder to obtain enhanced features;

[0040] S33: inputting the enhanced features to the next cascade layer for diagnosis, and obtaining the diagnosis result of the next cascade layer;

[0041] S34: Repeat steps S32 to S33 until the accuracy of the first deep forest model no longer increases, and the trained model is output as the second deep forest model.

[0042] S4: Test prediction, input the deep feature set of the test set into the second deep forest model, obtain the corresponding fault type label, and complete fault diagnosis.

[0043] The time-frequency domain convolutional network proposed in this application improves on the traditional one-dimensional convolutional neural network by adding a frequency domain convolution part, so as to obtain the global receptive field in the time domain and achieve more accurate feature extraction of vibration signals. At the same time, the time-frequency domain convolutional network also has an exceptionally good performance in processing vibration data in high noise conditions and complex working conditions, which helps to weaken the impact of high noise signals and working condition changes, thereby achieving efficient diagnosis of wind turbine bearing faults.

[0044] In the improved deep cascade forest of this embodiment, each layer is composed of random forest, completely random forest, XGboost and Gaussian Bayes classifier. The cascade layer accepts the original input deep feature data set and the enhanced vector of the previous cascade forest, outputs the enhanced feature vector of this layer through the intra-layer model analysis, and then splices it with the original feature and inputs it into the autoencoder to obtain the mapping data in the same space and input the data to the next layer of diagnosis. After multiple layers of cascade forests, until the model accuracy converges, the model training is completed.

[0045] Among them, the fully random forest in the deep forest model is a special case of the random forest, and the model parameters in all layers of the deep forest model remain consistent.

[0046] In this embodiment, the deep forest model includes at least two cascade layers, adjacent cascade layers are connected through autoencoders, and each cascade layer includes a random forest, a completely random forest, an XGboost and a Gaussian Bayes classifier.

[0047] The autoencoder is a stacked autoencoder and the parameters of each layer of the autoencoder are consistent.

[0048] The model parameters used in this embodiment are as follows: the learning rate of the random forest is 0.15, the depth of the tree is 11, the number is 200, the learning rate of the XGboost model is 0.1, the number of learners is 400, and the Gaussian Bayes classifier is the default parameters.

[0049] In this embodiment, the length of the input vibration data is 1200, including vibration data of more than one circle of the bearing, and a total of 6000 test samples and training samples are set.

[0050] The diagnostic method of this embodiment, in the process of constructing an identity mapping between vibration data and fault types, utilizes a time-frequency domain convolutional network to extract deep-level features of vibration data under complex working conditions and high noise, which can weaken the adverse effects of complex working conditions and noise, and utilizes an improved deep forest to improve diagnostic accuracy, thereby achieving efficient diagnosis of wind turbine bearing faults.

[0051] The effectiveness of the technical solution of the present application is illustrated below by an embodiment.

[0052] The Case Western Reserve University public bearing data set is used as the verification data of the present invention, such as Figure 5 As shown, the data contains ten bearing conditions including nine fault types and one normal data type, and -2db noise is added to the vibration data.

[0053] The fault diagnosis model based on time-frequency domain convolutional network and deep forest in this application is used for fault diagnosis, and the corresponding fault diagnosis results are as follows: Figure 6 As shown. Figure 6 It can be seen that the diagnostic results of this application can basically distinguish these ten different faults and have a high diagnostic accuracy.

[0054] The fault diagnosis method of the present application is based on time-frequency domain convolutional networks and deep forest models. Compared with traditional machine learning methods, it can not only effectively extract features under complex working conditions and high noise conditions, but also combine multiple models to achieve efficient diagnosis of wind turbine bearings, which can meet the requirements of smart power plants for intelligent diagnosis and wide applicability of wind turbine bearings.

[0055] Those skilled in the art can understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A wind turbine bearing fault diagnosis method based on time-frequency domain convolutional network and deep forest, characterized in that: include: S1: Obtain the original vibration signal of the wind turbine bearing and the corresponding operating condition, speed, load and fault type label; Windowing the original vibration signal to obtain multiple fault data of a single fault source, preprocessing the multiple fault data and adding corresponding fault type labels to form a data set, and dividing the data set into a training set and a test set; S2: inputting the training set and the test set into a time-frequency domain convolutional network for feature extraction to obtain respective deep-level feature sets; S3: Inputting the deep feature set into the first deep forest model to train the first deep forest model to obtain a second deep forest model; S4: inputting the deep feature set of the test set into the second deep forest model to obtain the corresponding fault type label and complete fault diagnosis; The first deep forest model includes at least two cascade layers, adjacent cascade layers are connected by autoencoders, and each cascade layer includes a random forest, a completely random forest, an XGboost, and a Gaussian Bayes classifier; Inputting the deep feature set into a first deep forest model to train the first deep forest model to obtain a second deep forest model includes: S31: inputting the deep feature set into the first cascade layer of the first deep forest model for diagnosis, and obtaining a diagnosis result of the first cascade layer; S32: combining the diagnosis result with the deep feature set and inputting the result into an autoencoder to obtain enhanced features; S33: inputting the enhanced features to the next cascade layer for diagnosis, and obtaining the diagnosis result of the next cascade layer; S34: Repeat steps S32 to S33 until the accuracy of the first deep forest model no longer increases, and the trained model is output as the second deep forest model.

2. The fault diagnosis method according to claim 1, characterized in that: The time-frequency domain convolutional network includes a time domain and frequency domain convolutional layer, an attention mechanism layer, and a convolution-pooling layer; the time domain and frequency domain convolutional layer includes a time domain convolution unit, a frequency domain convolution unit, and an average pooling layer; the convolution-pooling layer includes two consecutive convolution-pooling units and a fully connected layer.

3. The fault diagnosis method according to claim 2, characterized in that: The training set and the test set both include vibration signals, and the step S2 includes: S21: directly convolve the vibration signal through a time domain convolution unit to obtain a time domain sequence with a local receptive field; The vibration signal is transformed by DFT through the frequency domain convolution unit to obtain the amplitude signal in the frequency domain, and then the amplitude signal is convolved to obtain the local receptive field frequency domain sequence, and finally the local receptive field frequency domain sequence is transformed by IDFT to obtain the time domain sequence of the global receptive field; S22: splicing the local receptive field time domain sequence and the global receptive field time domain sequence, downsampling the spliced ​​time domain sequence through an average pooling layer, and then extracting features using an attention mechanism to obtain preliminary features; S23: The preliminary features are subjected to two consecutive convolution-pooling units for deep feature extraction, and finally output through a fully connected layer to obtain the deep-level features.

4. The fault diagnosis method according to claim 1, characterized in that: The autoencoder is a stacked autoencoder and the parameters of each layer of the autoencoder are kept consistent.

Images