A bearing fault diagnosis method based on multi-input CNN

Abstract

The present invention discloses a bearing fault diagnosis method based on multi-input CNN. The method comprises the following steps: first, using an accelerometer to collect bearing vibration signals; then converting the vibration signals into time-frequency diagrams through short-time Fourier transform and wavelet transform, labeling 5 types of samples respectively, and dividing the samples into data sets and test sets; then sending the training set into the built CNN model, continuously performing forward propagation and reverse propagation to update parameters until the total number of iterations set in advance is reached, and saving the model obtained at this time; then using the test set to test the diagnostic performance of the bearing fault diagnosis model; finally using this model to perform fault analysis, thereby realizing bearing fault diagnosis. The bearing fault diagnosis method of the present invention converts the original vibration data into time-frequency diagrams through short-time Fourier transform and continuous wavelet transform, and finally inputs them into the network at the same time, thereby more fully extracting the original signal features and realizing accurate diagnosis of bearing faults.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearing fault detection, and in particular to a bearing fault diagnosis method based on multi-input CNN. Background Art

[0002] Rolling bearings are core components in mechanical equipment. Bearing failure may cause serious casualties and economic losses. Therefore, bearing fault diagnosis and ensuring the smooth operation of bearings are essential steps to maintain the safe and stable operation of modern mechanical equipment.

[0003] Traditional bearing fault diagnosis methods focus on manual feature extraction, but these processes are too cumbersome and lack intelligence. It is difficult to extract features manually when the amount of data is particularly large. With the rise of big data-driven artificial intelligence technology, deep learning has been widely used in feature learning, pattern recognition, data mining and other fields with its powerful learning ability. Fault diagnosis has also made new research progress with the application of deep learning technology. As a typical deep learning technology, convolutional neural network has made some progress in fault diagnosis.

[0004] Since the actual measured bearing fault signals usually contain strong noise, in order to better extract signal features, most models usually use time-frequency conversion methods to convert one-dimensional vibration signals into two-dimensional time series signals. However, most of these models only use one time-frequency conversion method, which cannot fully extract the features of the original signal and there is a risk of losing important information, resulting in insufficient training accuracy. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a bearing fault diagnosis method based on multi-input CNN, by using two time-frequency conversion methods, continuous wavelet transform (CWT) and short-time Fourier transform (STFT), to accurately convert the original vibration signal into a time domain signal, so that the original signal features can be fully extracted and the training accuracy of the network is improved.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A bearing fault diagnosis method based on multi-input CNN includes the following steps:

[0008] Step 1:

[0009] The vibration signal collected by the accelerometer is a one-dimensional signal. The vibration signal is converted into a time-frequency diagram through short-time Fourier transform and continuous wavelet transform. The sample label is given and divided into training set and test set.

[0010] Step 2:

[0011] Construct a convolutional neural network model, including: a feature extraction layer, a feature overlay layer, and a feature classification layer. The feature extraction layer includes two channels that use convolution kernels of different sizes to extract time-frequency graph features obtained by different time-frequency conversion methods. The feature overlay layer overlays the two output feature graphs of the feature extraction layer in the channel dimension. The feature classification layer classifies fault categories through a Softmax classifier.

[0012] Step 3:

[0013] Initialize the parameters of the neural network, initialize the convolution kernel, bias, and weights to a random number, and then set hyperparameters such as learning rate and number of iterations.

[0014] Step 4:

[0015] Use the samples in the training set to train the constructed network, and update the network parameters through forward propagation and back propagation until the model reaches the preset cycle;

[0016] Step 5:

[0017] The test set samples are input into the trained network to obtain the diagnosis results.

[0018] In step 1, the data is preprocessed and the data set and test set are divided as follows:

[0019] The data is divided into the number of sample points collected in one rotation, and the amplitude of the data is normalized to [-1, 1]. Since the data set used has 5 categories, each of which includes 500 samples, there are a total of 2500 samples. The training data and test data are randomly selected with a ratio of 4:1.

[0020] The step 2 is specifically as follows:

[0021] (2-1): The feature extraction layer consists of two channels. The convolution kernel size and step size in the two channels are different. For a given sample Represents the sample, L represents the corresponding label, and the working process of the convolutional layer is described by the following formula:

[0022] F i =CONV(S*γ i conv)

[0023] Where CONV(*) represents the convolution operation of CNN, γ i conv represents the parameters of the i-th convolutional layer, F i Represents the extracted spatial feature information.

[0024] (2-2): The feature superposition layer superimposes the two output feature maps of the feature extraction layer in the channel dimension, and finally outputs a feature map with 28 channels and a length and width of 32×32.

[0025] (2-3): The feature classification layer classifies the fault category through the Softmax classifier. The Softmax classifier process is described by the following formula:

[0026]

[0027] Where * represents the Softmax classifier, represents the Softmax parameter of the jth class, SC j Represents the predicted probability distribution of different categories.

[0028] The step three is specifically as follows:

[0029] The hyperparameters of the model are set as follows: the learning rate is set to 0.005; the number of training rounds is 50; 64 samples are randomly shuffled and selected as a batch in each round; and the Adam optimization algorithm is selected as the optimization algorithm.

[0030] Beneficial effects of the present invention:

[0031] The present invention discloses a bearing fault diagnosis method based on multi-input CNN. The method converts a one-dimensional original vibration signal into a time-frequency signal through short-time Fourier (STFT) and continuous wavelet transform (CWT) respectively and saves it as an image as the input of the model. By adjusting the hyperparameters of the model, bearing fault diagnosis is realized. Compared with other models that use a time-frequency conversion method, the original vibration signal characteristics can be more fully extracted. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is the algorithm flow chart of this method.

[0033] Figure 2 It is the time-frequency diagram after short-time Fourier transform.

[0034] Figure 3 It is the time-frequency diagram after continuous wavelet transform.

[0035] Figure 4 This is the network structure diagram of this method. DETAILED DESCRIPTION

[0036] The present invention is described in detail below in conjunction with the accompanying drawings:

[0037] like Figure 1 A bearing fault diagnosis method based on a multi-input CNN of the present invention comprises:

[0038] (1) The vibration signal collected by the accelerometer is a one-dimensional signal. The vibration signal is converted into a time-frequency diagram through short-time Fourier transform and continuous wavelet transform. Figure 2 is the time-frequency diagram after short-time Fourier transform under a certain fault. Figure 3 It is a time-frequency diagram after continuous wavelet transformation under a certain fault. Given the time-frequency diagram label, it is divided into training set and test set;

[0039] In this embodiment, the Southeast University gearbox dataset is directly used to describe the specific implementation steps of the present invention.

[0040] The data also comes from the Southeast University gearbox data set, and the bearing fault is tested by manually setting a constant speed workbench. Five different fault states of data are selected.

[0041] The bearing vibration data collected above are labeled according to their damage types, as shown in Table 1 below;

[0042] Table 1 A sample data set

[0043]

[0044] The above table divides the loss state into 5 categories, each of which includes 500 samples, so there are 2500 samples in total. The training data and test data are randomly selected with a ratio of 4:1.

[0045] (2) Build a neural network model based on the improved CNN, and use the labeled bearing vibration data to train the model, where the model includes: feature extraction layer, feature superposition layer, and feature classification layer.

[0046] (2-1) The feature extraction layer consists of two channels. The convolution kernel size and step size in the two channels are different. For a given sample Represents the sample, L represents the corresponding label, and the process is described by the following formula:

[0047] F i =CONV(S*γ I conv)

[0048] Where CONV(*) represents the convolution operation of CNN, γ i conv represents the parameters of the i-th convolutional layer, F i Represents the extracted spatial feature information.

[0049] (2-2) The feature superposition layer superimposes the two output feature maps of the feature extraction layer in the channel dimension, and finally outputs a feature map with 28 channels and a length and width of 32×32.

[0050] (2-3): The feature classification layer classifies the fault category through the Softmax classifier. The Softmax classifier process is described by the following formula:

[0051]

[0052] Where * represents the Softmax classifier, represents the Softmax parameter of the jth class, SC j Represents the predicted probability distribution of different categories.

[0053] (3) The training samples are passed into the network according to the number set in advance for forward propagation. After passing through the classification layer, the predicted value is obtained, and then the loss between the predicted value and the true value is calculated. The error is back-propagated using the optimization algorithm to update the network parameters. The hyperparameters of the model are set as follows: the learning rate is set to 0.005; the number of training rounds is 50; 64 samples are randomly shuffled and extracted as a batch in each round; the optimization algorithm selects the Adam optimization algorithm.

[0054] (4) Input the test set into the trained model to obtain the diagnosis results and complete the bearing fault diagnosis.

Claims

1. Bearing fault diagnosis method based on multi-input CNN, characterized in that: Follow these steps: a. The vibration signal collected by the accelerometer is a one-dimensional signal. The vibration signal is converted into a time-frequency diagram through short-time Fourier transform and continuous wavelet transform. The sample label is given and divided into training set and test set. b. Build a convolutional neural network model, including: a feature extraction layer, a feature superposition layer, a feature classification layer, and the feature extraction layer includes two channels that use convolution kernels of different sizes to extract the time-frequency graph features obtained by different time-frequency conversion methods, as follows: For a given sample S represents the sample, L represents the corresponding label, and the process is described by the following formula: F i =CONV(S*γ i conv) (1) Where CONV(*) represents the convolution operation of CNN, γ i conv represents the parameters of the i-th convolutional layer, F i The feature superposition layer superimposes the two output feature maps of the feature extraction layer on the channel dimension, and the feature classification layer classifies the fault category through the Softmax classifier; c. Initialize the parameters of the neural network, initialize the convolution kernel, bias, and weights to a random number, and then set hyperparameters such as learning rate and number of iterations; d. Use the samples in the training set to train the constructed network, and update the network parameters through forward propagation and back propagation until the model reaches the preset cycle; e. The test set samples are input into the trained network to obtain the diagnosis results.

2. The bearing fault diagnosis method based on multi-input CNN according to claim 1 is characterized in that: In step a, the data is preprocessed and divided into training set and test set as follows: The data is divided into the number of sample points collected in one revolution, and the amplitude of the data is normalized to [-1, 1]. Since the data set used has 5 categories, each category includes 500 samples, there are a total of 2500 samples. The training data and test data are randomly selected with a ratio of 4:

1.

3. The bearing fault diagnosis method based on multi-input CNN according to claim 1 is characterized in that: In step b, the feature overlay layer is as follows: The feature overlay layer overlays the two output feature maps of the feature extraction layer in the channel dimension, and finally outputs a feature map with 28 channels and a length and width of 32×32.

4. The bearing fault diagnosis method based on multi-input CNN according to claim 1 is characterized in that: In step b, the feature classification layer is as follows: The feature classification layer classifies the fault category through the Softmax classifier. The Softmax classifier process is described by the following formula: Where * represents the Softmax classifier, represents the Softmax parameter of the jth class, SC j Represents the predicted probability distribution of different categories.

5. The bearing fault diagnosis method based on multi-input CNN according to claim 1 is characterized in that: In step c, the constructed network is trained using samples in the training set until the model reaches the preset cycle as follows: The learning rate is set to 0.005; the number of training rounds is 50; 64 samples are randomly shuffled and selected as a batch in each round; the optimization algorithm is the Adam optimization algorithm.

Images