Bearing fault identification method based on deep learning under noise interference
By constructing a noise-added data set in bearing fault recognition and training a deep learning model, the problem of low bearing fault recognition accuracy under noise interference is solved, and a higher recognition accuracy and simplified feature extraction process is achieved.
Patent Information
- Application Number
- CN202411378799.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-30
- Publication Date
- 2025-03-04
AI Technical Summary
The existing bearing fault identification methods based on deep learning still have room for improvement in recognition accuracy under noise interference, and traditional methods that rely on professional knowledge are cumbersome.
A bearing fault identification method based on deep learning under noise interference is proposed. By constructing a bearing fault vibration signal data set and pre-processing, the original CNN model and bearing fault identification model are trained, and the noise-added data set is used to train the recognition model to improve the recognition accuracy under noise conditions.
The accuracy of bearing fault identification has been improved, especially under noise interference, the recognition accuracy has been improved by 2.31%, overcoming the dependence of traditional methods on professional knowledge and cumbersome problems.
Smart Images
Figure CN119557721A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deep learning and signal recognition, and in particular to a bearing fault recognition method based on deep learning under noise interference. Background Art
[0002] Bearing fault identification methods based on traditional algorithms rely heavily on professional knowledge to process and extract features of raw signals, and the fault identification process is very cumbersome. However, bearing fault identification methods based on deep learning automatically obtain the features of raw vibration signals without any prior knowledge and perform end-to-end fault identification. Its powerful feature extraction capability can obtain more accurate fault identification results.
[0003] Researchers initially used CNN to implement end-to-end rolling bearing fault detection to improve fault identification accuracy and real-time performance. Later, with the continuous evolution of deep convolutional network models, researchers proposed more models and applied them to bearing fault identification. These methods have made some progress, but lack consideration of noise issues in actual working conditions; some researchers have shifted their focus to the problem of bearing fault identification under noisy conditions, and used various deep neural network models, transfer learning methods, attention mechanisms, and a series of extended optimization methods to identify bearing faults under noisy conditions. However, there are relatively few existing studies, and there is still much room for improvement in identification accuracy. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a bearing fault identification method based on deep learning under noise interference.
[0005] In order to achieve the above tasks, the present invention adopts the following technical solutions:
[0006] A bearing fault identification method based on deep learning under noise interference, comprising:
[0007] Construct a bearing fault vibration signal dataset and preprocess it, use the preprocessed bearing fault vibration signal dataset to train the original CNN model, and save the trained original CNN model;
[0008] Construct a noisy bearing fault vibration signal dataset and perform preprocessing; the dataset is constructed by adding noise to the bearing fault vibration signal dataset;
[0009] A bearing fault recognition model is established; the model is composed of an encoder, a decoder and a trained original CNN model connected in sequence; the bearing fault recognition model is trained using a noisy bearing fault vibration signal data set, and the trained bearing fault recognition model is saved for classification and recognition of bearing fault signals of unknown fault types.
[0010] Furthermore, the construction of the bearing fault vibration signal data set and preprocessing includes:
[0011] S11, collecting bearing fault vibration signals, each signal sample including a segment of bearing fault vibration signal data; performing sampling processing on the signal samples, and converting the signal samples into a discrete format;
[0012] S12, performing normalization processing on each signal sample after sampling processing;
[0013] S13, performing a data segmentation operation on the normalized sample signal using a sliding window, setting a window length, and partially overlapping adjacent sliding windows, thereby dividing the sample signal into multiple signal segments;
[0014] S14, performing data rearrangement operation on each divided signal segment, performing dimensionality-upgrading processing on the one-dimensional signal segment, and constructing a two-dimensional grayscale image; each signal sample finally obtains a set of corresponding grayscale images, and the grayscale images of the same group of signal samples are set with the same label.
[0015] Furthermore, the window length is 1024 and the grayscale image size is 32*32.
[0016] Furthermore, the construction of the noisy bearing fault vibration signal data set and preprocessing is specifically as follows:
[0017] The same method as S11 to S14 is used, except that before S13, Gaussian white noise is first added to the normalized sample signal, and the sample data after adding the noise is then subjected to subsequent data segmentation and data rearrangement operations to obtain a noisy sample signal; the data set constructed by such processing is a noisy bearing fault vibration signal data set.
[0018] Furthermore, the signal-to-noise ratios of the added Gaussian white noise are, from strong to weak, -4dB, -2dB, 0dB, 2dB and 4dB, and any one of them is selected to be added to the sample signal.
[0019] Furthermore, the structure of the original CNN model is:
[0020] Convolution layer, convolution kernel size 3*1, channel 16, padding to (1,0), step size 1; batch normalization layer, channel 16; activation layer; convolution layer, convolution kernel size 1*3, channel 16, padding to (0,1) step size 1; batch normalization layer, channel 16; activation layer; maximum pooling layer, the size of the pooling result is 2*2, step size 2; convolution layer, convolution kernel size 3*1, channel 32, padding to (1,0), step size 1; batch normalization layer, channel 32; activation layer; convolution layer, convolution kernel size 1*3, channel 32, padding is (0,1), step size is 1; batch normalization layer, channel 32; activation layer; maximum pooling layer, the size of the pooling result is 2*2, the step size is 2; convolution layer, convolution kernel size 3*1, channel 64, padding is (1,0), the step size is 1; batch normalization layer, channel 64; activation layer; convolution layer, convolution kernel size 1*3, channel 64, padding is (0,1), the step size is 1; batch normalization layer, channel 64; activation layer; average pooling layer, the size of the pooling result is 1*1.
[0021] Furthermore, the structure of the encoder is:
[0022] Convolution layer, convolution kernel size 3*3, channel 64, padding is (1,1), step size is 1; batch normalization layer, channel 64; activation layer; maximum pooling layer, the size of the pooling result is 2*2, step size is 2; convolution layer, convolution kernel size 3*3, channel 128, padding is (1,1) step size is 1; batch normalization layer, channel 128; activation layer; maximum pooling layer, the size of the pooling result is 2*2, step size is 2; convolution Layer, convolution kernel size 3*3, channel 128, padding is (1,1), step size is 1; batch normalization layer, channel 128; activation layer; maximum pooling layer, the size of the pooling result is 2*2, step size is 2; convolution layer, convolution kernel size 3*3, channel 128, padding is (1,1), step size is 1; batch normalization layer, channel 128; activation layer; maximum pooling layer, average pooling layer, the size of the pooling result is 1*1.
[0023] Furthermore, the structure of the decoder is:
[0024] Transposed convolution layer, convolution kernel size 3*3, channels 128, padding to (1,1), output padding to (1,1), stride 2; batch normalization layer, channels 128; activation layer; transposed convolution layer, convolution kernel size 3*3, channels 64, padding to (1,1), output padding to (1,1), stride 2; batch normalization layer, channels 64; activation layer; transposed convolution layer, convolution kernel size 3*3, channels 64, padding to (1,1), output padding to (1,1), stride 2; batch normalization layer, channels 64; activation layer; transposed convolution layer, convolution kernel size 3*3, channels 64, padding to (1,1), output padding to (1,1), stride 2; batch normalization layer, channels 64; activation layer.
[0025] Furthermore, the loss function of the bearing fault identification model during training is:
[0026]
[0027] Among them, y' k represents the predicted fault category, y k represents the label of the noisy sample signal, k represents the data dimension, n represents the number of class faults, and N represents the batch size; represents the eigenvector of the i-th signal sample, represents the d-dimensional real number set; y i Represents the category label of the i-th signal sample; is the weight matrix; W j , W yi represents the jth and yth i Column, b j , b yi is the corresponding bias, the superscript T represents the transpose, and e is a natural constant.
[0028] A terminal device comprises a processor, a memory and a computer program stored in the memory; when the processor is executed by a computer, the bearing fault identification method based on deep learning under noise interference is implemented.
[0029] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the bearing fault identification method based on deep learning under noise interference is implemented.
[0030] Compared with the prior art, the present invention has the following technical features:
[0031] 1. The proposed CNN-based bearing fault identification model improves the accuracy of bearing fault identification and overcomes the problem that traditional bearing fault identification methods rely on professional knowledge and are cumbersome;
[0032] 2. The proposed bearing fault identification model improves the accuracy of bearing fault identification under noise interference and overcomes the problem that noise in actual working conditions affects the accuracy of bearing fault identification;
[0033] 3. The proposed data preprocessing method makes full use of the internal connections of the data through data normalization, data segmentation and data rearrangement, and eliminates the influence of prior knowledge on feature extraction to the greatest extent. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is the overall network structure diagram of the present invention;
[0035] Figure 2 The original signal and the signal after adding noise;
[0036] Figure 3 It is a diagram showing the process of upgrading a one-dimensional time domain signal to a two-dimensional grayscale image;
[0037] Figure 4 This is the structure diagram of the designed CNN network;
[0038] Figure 5 This is the accuracy graph of the original CNN model in the absence of noise;
[0039] Figure 6 This is the accuracy graph of the original CNN model in the presence of noise;
[0040] Figure 7 This is the confusion matrix diagram of the original CNN model in the presence of noise;
[0041] Figure 8 This is the accuracy diagram of the bearing fault identification model in the presence of noise;
[0042] Fig. 9 Confusion matrix diagram of the bearing fault recognition model in the presence of noise. DETAILED DESCRIPTION
[0043] Referring to the accompanying drawings, the present invention provides a bearing fault identification method based on deep learning under noise interference, comprising the following steps:
[0044] Step 1: construct a bearing fault vibration signal dataset and perform preprocessing.
[0045] S11, collecting bearing fault vibration signals, each signal sample includes a section of bearing fault vibration signal data; performing sampling processing on the signal samples, and converting the signal samples into a discrete format.
[0046] S12, normalize each signal sample after sampling:
[0047]
[0048] Among them, X is the signal data obtained after normalization, x i is the signal point in the signal sample, x max 、x min are the maximum and minimum values of the signal points in the sample signal respectively.
[0049] S13, performing data segmentation operation on the normalized sample signal, using a sliding window, setting the window length to 1024, and partially overlapping between adjacent sliding windows, thereby dividing the sample signal into multiple signal segments with a length of 1024, and each signal segment overlaps with the adjacent signal segment, thereby expanding the valid data.
[0050] S14, performing data rearrangement operation on each divided signal segment, performing dimensionality-upgrading processing on the one-dimensional signal segment, and constructing a two-dimensional grayscale image, such as Figure 3 As shown; in this embodiment, the grayscale image size is 32*32; so that each signal sample finally obtains a set of corresponding grayscale images, and the grayscale images of the same set of signal samples are set with the same label.
[0051] Step 2: construct the original CNN model and train the original CNN model using the preprocessed bearing fault vibration signal dataset.
[0052] The original CNN model structure designed in this embodiment is as follows Figure 4 The specific parameters of the original CNN model are shown in Table 1; the calculation formula for the size of the feature map after the convolution operation of each convolution layer is:
[0053] H out =(H in -kernel_size[0]+
[0054] 2×padding[0]) / stride[0]+1
[0055] W out =(W in -kernel_size[1]+
[0056] 2×padding[1]) / stride[1]+1
[0057] Among them, H out and W out Respectively represent the height and width of the feature map output after convolution processing, H in and W inThey represent the height and width of the feature map / grayscale image input to the convolution layer, kernel_size represents the size of the convolution kernel, padding represents padding, stride represents the stride, index [0] represents data in the height direction, and index [1] represents data in the width direction.
[0058] Table 1 Original CNN model parameter values
[0059]
[0060] Among them, the convolution layer parameters (3, 1, 16) indicate that the convolution kernel size used is 3*1 and the number of channels is 16. The maximum pooling layer parameters (2, 2) indicate that the size of the pooling result is 2*2, and the Padding (1, 0) indicates the padding amount in the height and width directions; 1 indicates that the padding amount on the upper and lower sides of the height is 1 row each, and 0 indicates that there is no padding on the left and right sides in the width direction.
[0061] Step 3: Train the original CNN model using the Adam optimizer and save the trained original CNN model; the initial learning rate and weight factor are set to 10. -3 and 10 -4 , the batch size used is 16; the training phase has 25 epochs, the learning rate is reduced by 90% at the 15th and 20th epochs, and the loss function is:
[0062]
[0063] Where n represents the number of class faults; N represents the batch size; represents the eigenvector of the i-th signal sample, represents the d-dimensional real number set; y i Represents the category label of the i-th sample; is the weight matrix; W j , W yi represents the jth and yth i Column, b j , b yi is the corresponding bias, the superscript T represents the transpose, and e is a natural constant.
[0064] Step 4: construct a noisy bearing fault vibration signal dataset; this dataset is constructed by adding noise on the basis of the bearing fault vibration signal dataset, specifically:
[0065] The same method as S11 to S14 is used, except that before S13, Gaussian white noise is first added to the normalized sample signal, and the sample data after adding the noise is then subjected to subsequent data segmentation and data rearrangement operations to obtain a noisy sample signal; the data set constructed by such processing is a noisy bearing fault vibration signal data set.
[0066] In this embodiment, the signal-to-noise ratios of the added Gaussian white noise are, from strong to weak, -4dB, -2dB, 0dB, 2dB and 4dB, and any one of them can be selected and added to the sample signal. Figure 2 These are the normal time domain signal without noise and with noise when the load state is 0, and the signal-to-noise ratio of the added noise is -4db.
[0067] Step 5: Build the encoder and decoder.
[0068] First, the noisy sample signal x is input into the encoder for encoding processing to realize feature extraction. The signal undergoes downsampling and the spatial features are reduced. The feature y=f(x) obtained after the encoder encoding processing is input into the decoder, and upsampled through transposed convolution to increase the width and height of the input to achieve the purpose of noise reduction and then restore the signal. The signal is restored to a sample signal data x=g(y)=g(f(x)) close to noise-free, that is, to make x reproduce the input x as much as possible. In the process of signal recovery, the number of channels is decreasing and the spatial scale is increasing. Here, the encoding of the middle layer, that is, the mapping from the input signal to the encoder, is the most important, that is, to realize automatic feature extraction of the signal. It can be expressed as follows:
[0069] y=f(x)=A(wx+b)
[0070] x=g(y)=A(w'y+b')
[0071] Among them, A represents the activation function relu, w and w' represent the weights of the encoder network and the decoder network respectively, and b and b' represent the bias of the encoder network and the decoder network respectively.
[0072] The specific parameters of the encoder and decoder are shown in Table 2 and Table 3 respectively; the formula for calculating the size of the feature map after the transposed convolution operation is:
[0073] H out =(H in -1)×stride[0]-2×padding[0]+dialation[0]×(kernel_size[0]-1)+out_padding[0]+1
[0074] W out =(W in-1)×stride[1]-2×padding[1]+dialation[1]×kernel_size[1]-1)+out_padding[1]+1
[0075] Among them, index [0] represents the data in the height direction, index [1] represents the data in the width direction, and dialation is the convolution parameter, which is set to 1.
[0076] Table 2 Encoder parameter values
[0077]
[0078] Table 3 Decoder parameter values
[0079]
[0080]
[0081] Among them, the parameter (1, 1) of Out_padding fills several rows or columns in the height and width directions of the calculated output feature map, which is a one-sided padding.
[0082] Step 6: construct a bearing fault recognition model; the model is composed of an encoder, a decoder, and a trained original CNN model connected in sequence. The bearing fault recognition model is trained using a noisy bearing fault vibration signal dataset, and the trained bearing fault recognition model is saved for classification and recognition of bearing fault signals of unknown fault types.
[0083] That is, the noisy sample signal is input into the encoder for encoding processing, and then decoded by the decoder to restore the sample signal that is approximately noise-free. After that, the signal is classified by the trained original CNN model, and finally the category corresponding to the fault is output.
[0084] The Adam optimizer is used for training, and the initial learning rate and weight factor are set to 10 -3 and 10 -4 , the batch size used is 16; the training phase has 25 epochs, the learning rate is reduced by 90% at the 15th and 20th epochs, and the loss function is:
[0085]
[0086] Among them, y' k represents the predicted fault category, y k represents the label of the noisy sample signal, k represents the data dimension, and the other parameters have the same meaning as before.
[0087] The accuracy diagrams of the bearing fault identification model provided by the present invention in the absence of noise and in the presence of noise are as follows: Figure 5 and Figure 6 As shown; the confusion matrix diagram in the case of noise is as follows Figure 7 As shown; the accuracy diagram and confusion matrix diagram of the bearing fault recognition model in the presence of noise are shown as follows: Figure 8 and Fig. 9 It can be seen that the original CNN model designed in the present invention obtains a fault recognition accuracy of 99.83% in the absence of noise, and the designed bearing fault recognition model obtains a fault recognition accuracy of 97.14% in the case of -4db signal-to-noise ratio noise, which is 2.31% higher than the accuracy of the original CNN model in the same noise condition.
[0088] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application 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 make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.
Claims
1. A bearing fault identification method based on deep learning under noise interference, characterized in that: include: Construct a bearing fault vibration signal dataset and preprocess it, use the preprocessed bearing fault vibration signal dataset to train the original CNN model, and save the trained original CNN model; Construct a noisy bearing fault vibration signal dataset and perform preprocessing; This dataset is constructed by adding noise to the bearing fault vibration signal dataset; A bearing fault recognition model is established; the model is composed of an encoder, a decoder and a trained original CNN model connected in sequence; the bearing fault recognition model is trained using a noisy bearing fault vibration signal data set, and the trained bearing fault recognition model is saved for classification and recognition of bearing fault signals of unknown fault types.
2. The bearing fault identification method based on deep learning under noise interference according to claim 1 is characterized in that: The method of constructing a bearing fault vibration signal data set and preprocessing the data set includes: S11, collecting bearing fault vibration signals, each signal sample including a segment of bearing fault vibration signal data; performing sampling processing on the signal samples, and converting the signal samples into a discrete format; S12, performing normalization processing on each signal sample after sampling processing; S13, performing a data segmentation operation on the normalized sample signal using a sliding window, setting a window length, and partially overlapping adjacent sliding windows, thereby dividing the sample signal into multiple signal segments; S14, performing data rearrangement operation on each divided signal segment, performing dimensionality-upgrading processing on the one-dimensional signal segment, and constructing a two-dimensional grayscale image; each signal sample finally obtains a set of corresponding grayscale images, and the grayscale images of the same group of signal samples are set with the same label.
3. The bearing fault identification method based on deep learning under noise interference according to claim 2 is characterized in that: The window length is 1024 and the grayscale image size is 32*32.
4. The bearing fault identification method based on deep learning under noise interference according to claim 2 is characterized in that: The construction of the noisy bearing fault vibration signal data set and preprocessing is specifically as follows: The same method as S11 to S14 is used, except that before S13, Gaussian white noise is first added to the normalized sample signal, and the sample data after adding the noise is then subjected to subsequent data segmentation and data rearrangement operations to obtain a noisy sample signal; the data set constructed by such processing is a noisy bearing fault vibration signal data set.
5. The bearing fault identification method based on deep learning under noise interference according to claim 4 is characterized in that: The signal-to-noise ratios of the added Gaussian white noise are, from strong to weak, -4dB, -2dB, 0dB, 2dB and 4dB. Any one of them is selected and added to the sample signal.
6. The bearing fault identification method based on deep learning under noise interference according to claim 1 is characterized in that: The structure of the original CNN model is: Convolution layer, convolution kernel size 3*1, channels 16, padding (1,0), step size 1; batch normalization layer, channels 16; activation layer; Convolution layer, convolution kernel size 1*3, channels 16, padding (0,1) step size 1; batch normalization layer, channels 16; activation layer; Maximum pooling layer, the size of the pooling result is 2*2, the step size is 2; convolution layer, the convolution kernel size is 3*1, the channel is 32, the padding is (1,0), and the step size is 1; batch normalization layer, the channel is 32; activation layer; Convolution layer, convolution kernel size 1*3, channel 32, padding (0,1), step size 1; batch normalization layer, channel 32; activation layer; Maximum pooling layer, the size of the pooling result is 2*2, the step size is 2; convolution layer, the convolution kernel size is 3*1, the channel is 64, the padding is (1,0), and the step size is 1; batch normalization layer, the channel is 64; activation layer; Convolutional layer, convolution kernel size 1*3, channel 64, padding (0,1), stride 1; Batch normalization layer, channel 64; activation layer; average pooling layer, the size of the pooling result is 1*1.
7. The bearing fault identification method based on deep learning under noise interference according to claim 1 is characterized in that: The structure of the encoder is: Convolution layer, convolution kernel size 3*3, channel 64, padding is (1,1), step size is 1; batch normalization layer, channel 64; activation layer; maximum pooling layer, the size of the pooling result is 2*2, step size is 2; convolution layer, convolution kernel size 3*3, channel 128, padding is (1,1) step size is 1; batch normalization layer, channel 128; activation layer; maximum pooling layer, the size of the pooling result is 2*2, step size is 2; convolution Layer, convolution kernel size 3*3, channel 128, padding is (1,1), step size is 1; batch normalization layer, channel 128; activation layer; maximum pooling layer, the size of the pooling result is 2*2, step size is 2; convolution layer, convolution kernel size 3*3, channel 128, padding is (1,1), step size is 1; batch normalization layer, channel 128; activation layer; maximum pooling layer, average pooling layer, the size of the pooling result is 1*1.
8. The bearing fault identification method based on deep learning under noise interference according to claim 1 is characterized in that: The structure of the decoder is: Transposed convolution layer, convolution kernel size 3*3, channels 128, padding to (1,1), output padding to (1,1), stride 2; batch normalization layer, channels 128; activation layer; transposed convolution layer, convolution kernel size 3*3, channels 64, padding to (1,1), output padding to (1,1), stride 2; batch normalization layer, channels 64; activation layer; transposed convolution layer, convolution kernel size 3*3, channels 64, padding to (1,1), output padding to (1,1), stride 2; batch normalization layer, channels 64; activation layer; transposed convolution layer, convolution kernel size 3*3, channels 64, padding to (1,1), output padding to (1,1), stride 2; batch normalization layer, channels 64; activation layer.
9. The bearing fault identification method based on deep learning under noise interference according to claim 1 is characterized in that: The loss function of the bearing fault identification model during training is: Among them, y' k represents the predicted fault category, y k represents the label of the noisy sample signal, k represents the data dimension, n represents the number of class faults, and N represents the batch size; represents the eigenvector of the i-th signal sample, represents the d-dimensional real number set; y i Represents the category label of the i-th signal sample; is the weight matrix; W j , W yi represents the jth and yth i Column, b j 、b yi is the corresponding bias, the superscript T represents the transpose, and e is a natural constant.
10. A computer-readable storage medium, wherein a computer program is stored in the medium; characterized in that: When the computer program is executed by a processor, the bearing fault identification method based on deep learning under noise interference according to any one of claims 1 to 9 is implemented.