CN110702411B - Residual error network rolling bearing fault diagnosis method based on time-frequency analysis - Google Patents

Abstract

The invention relates to a fault diagnosis method for a rolling bearing with residual network based on time-frequency analysis. The graph is converted into a two-dimensional grayscale time-frequency graph; S2, use the residual network to extract features from the signal, and diagnose the fault type of the bearing; the input of the residual network is the grayscale time-frequency graph generated in step S1, and the output is The result of the fault diagnosis. The invention adopts the short-time Fourier transform to convert the bearing vibration data into a time-frequency diagram, which can clearly reflect the time-domain and frequency-domain characteristics when the faulty bearing vibrates, and facilitates the accurate diagnosis of different fault types by the network. Since the time-frequency signal contains both the time domain and frequency domain information of the bearing, and the deepening of the network layer of the residual network will not lead to the problem of gradient disappearance or gradient explosion, this method can obtain higher accuracy for bearing fault diagnosis. 's accuracy.

Description

A Fault Diagnosis Method for Rolling Bearings with Residual Networks Based on Time-Frequency Analysis

technical field

The invention relates to the field of bearing fault diagnosis, and more particularly, to a fault diagnosis method of a residual network rolling bearing based on time-frequency analysis.

Background technique

Rolling bearings are an important part of mechanical components. In large equipment or production lines, a large number of rolling bearings are running at any time. Once the rolling bearing has a serious failure, it will lead to the difficulty in controlling the accuracy of the product, and even stop the mechanical equipment or production line. Therefore, it is very important to diagnose the fault of rolling bearing. When the bearing fails, it can be found and repaired in time, which is also of great help to the operation reliability of the equipment.

For the fault diagnosis of rolling bearings, traditional methods can be used, such as feature extraction and fault classification of vibration signals. However, this method requires the relevant personnel to have rich prior knowledge, and when the vibration signal is mixed with the noise signal, the difficulty of signal feature extraction increases accordingly. Since fault diagnosis is to identify different fault types, and deep learning has made good achievements in image recognition, the deep learning method can be applied to the fault identification of rolling bearings. The residual network solves the problem of the degradation of deep learning due to the deepening of the network layer. Therefore, using the residual network for the fault diagnosis of rolling bearings can extract the high-dimensional features of the signal by deepening the network layer and increase the recognition accuracy.

However, the vibration signal of the rolling bearing is a signal about time, that is, the vibration signal of the rolling bearing only contains the information in the time domain and does not contain the information in the frequency domain. If the vibration signal of the rolling bearing is directly input into the residual network, the signal characteristics will be lost and the accuracy of diagnosis will be reduced. Therefore, the short-time Fourier transform can be used to convert the time-domain signal of the bearing into a time-frequency signal. The signal is a two-dimensional signal whose abscissa represents the time domain and the ordinate represents the frequency domain. In this way, the input data of the residual network will contain both time domain information and frequency domain information. The extraction of fault features by the network will be more comprehensive, which is beneficial to increase the accuracy of fault diagnosis.

Application No. 201710747694.9, named as a fault diagnosis method for rolling bearings based on convolutional neural network, using the method of short-time Fourier transform to transform the rolling bearing from time domain information to frequency domain information, this method can extract the frequency domain features of the signal , but this method uses an ordinary convolutional neural network in the subsequent processing, and this kind of network will have the problem of gradient disappearance or gradient explosion when the number of network layers is deepened. Therefore, the number of network layers is limited, the high-dimensional features of the bearing signal cannot be extracted, and the diagnostic accuracy is also limited.

Application No. 201810339956.2, the name is a method for establishing an intelligent diagnosis model for rolling bearings based on convolutional neural networks. Although this method converts the one-dimensional signal of the bearing into a two-dimensional signal, the conversion method is arranged in order, so that the stacked two-dimensional signals The signal has no actual physical meaning and cannot reflect the time and frequency domain information of bearing faults. Therefore, the accuracy of bearing fault diagnosis cannot be improved.

Huang Chicheng's master's thesis of Zhejiang University, "Research on the Fault Diagnosis and Optimization Method of Rolling Bearing Combining Time-Frequency Analysis and Convolutional Neural Networks" uses ResNet18 to diagnose the fault of the bearing's time-frequency signal. This method is directly identified in the residual network. The network structure is not modified according to the bearing characteristics, and there is no visual analysis of the network training process.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a fault diagnosis method for a residual network rolling bearing based on time-frequency analysis, which can achieve a high accuracy and high accuracy when using the method to diagnose the bearing fault.

The technical solution adopted by the present invention to solve the technical problem is: constructing a residual network rolling bearing fault diagnosis method based on time-frequency analysis, comprising the following steps:

S1. Collect vibration signal data, convert the vibration time domain signal of the rolling bearing into a time-frequency map by using short-time Fourier transform, and convert the time-frequency map into a two-dimensional grayscale time-frequency map;

S2. Use the residual network to perform feature extraction on the signal, and diagnose the fault type of the bearing; the input of the residual network is the grayscale time-frequency map generated in step S1, and the output is the result of fault diagnosis.

In the above scheme, in the step S1, each grayscale time-frequency image is labeled and divided into a training set and a test set.

In the above solution, step S3 is also included: input the data of the training set in step S1 into the residual network for training, the loss function adopts the cross entropy loss function, and the optimization method is the Adam algorithm; after the training is completed, the feature maps of different network layers are drawn, At the same time, the t-SNE algorithm is used to reduce the dimension and visualize the output of different network layers, and observe the relationship between different network layers.

In the above solution, step S4 is also included: input the test set in step S1 into the residual network to test the accuracy of the residual network.

In the above scheme, the short-time Fourier transform of the vibration time-domain signal z(t) is:

Among them, t is the time, f is the frequency, γ(t) is the window function, t'-t represents the sliding window, * represents the complex conjugate, and z(t) is the signal.

In the above scheme, the residual network is a 20-layer residual network, the convolution kernel of the first layer of the residual network is a 5×5 convolution kernel, and the subsequent network layers are a 3×3 convolution kernel.

Implementing the residual network rolling bearing fault diagnosis method based on time-frequency analysis of the present invention has the following beneficial effects:

1. The present invention uses short-time Fourier transform to convert the bearing vibration data into a time-frequency diagram, which can clearly reflect the time-domain and frequency-domain characteristics of the faulty bearing during vibration, which is convenient for the network to accurately diagnose different fault types. Since the time-frequency signal contains both the time-domain and frequency-domain information of the bearing, and the deepening of the network layer of the residual network does not cause the problem of gradient disappearance or gradient explosion, this method can obtain higher accuracy for bearing fault diagnosis. 's accuracy.

2. The present invention uses the t-SNE visualization algorithm to output the feature maps of different network layers, which can clearly show the changes of network characteristics when different network layers are diagnosing faulty bearings, which facilitates the adjustment of network parameters.

Description of drawings

The present invention will be further described below in conjunction with the accompanying drawings and embodiments, in which:

Figure 1 is a time-frequency diagram of different fault types;

Figure 2 shows the accuracy and loss function values on the training set;

Figure 3 is an input network feature map;

Fig. 4 is the output feature map of the first stage;

Fig. 5 is the output feature map of the second stage;

Fig. 6 is the output feature map of the third stage;

Fig. 7 is the output feature map of the fourth stage;

Figure 8 is a schematic diagram of the dimensionality reduction visualization of the original data;

Figure 9 is a schematic diagram of the first stage output parameter dimensionality reduction visualization;

Figure 10 is a schematic diagram of the second-stage output parameter dimensionality reduction visualization;

Figure 11 is a schematic diagram of the third-stage output parameter dimensionality reduction visualization;

Figure 12 is a schematic diagram of the fourth-stage output parameter dimensionality reduction visualization.

Detailed ways

In order to have a clearer understanding of the technical features, objects and effects of the present invention, the specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

The fault diagnosis method of the residual network rolling bearing based on the time-frequency analysis of the present invention comprises the following steps:

S1. Vibration signal data collection and processing;

The data set used this time is the bearing data set of Western Reserve University. In this example, a total of 10 fault types are set, as shown in Table 1. Then the data of each fault type is divided into 2048×1 samples, and finally each sample is converted into a 64×64 time-frequency grayscale image by using short-time Fourier transform.

Given a window function γ(t) with a very short time width and sliding the window, the short-time Fourier transform of the signal z(t) is:

Among them, t is the time, f is the frequency, γ(t) is the window function, t'-t represents the sliding window, * represents the complex conjugate, and z(t) is the signal.

The vibration time-domain signal of the rolling bearing is converted into a time-frequency diagram by means of short-time Fourier transform. Then, the time-frequency graph is converted into a two-dimensional time-frequency grayscale graph as shown in FIG. 1 . The brighter the color in the time-frequency grayscale graph represents the larger the amplitude, and the darker the color represents the smaller the amplitude. In the time-frequency grayscale image, not only the time domain information of the vibration signal can be seen, but also the frequency domain information of the vibration signal at different times. Finally, each time-frequency grayscale image is labeled and divided into training set and test set.

There are 1000 samples for each failure type, and 900 of them are randomly selected as the training set and the remaining 100 as the test set. These time-frequency grayscale images include the time domain signal and frequency domain signal when the bearing vibrates, which makes the characteristics of the bearing vibration more obvious and facilitates the network to identify the fault type.

Table 1 Fault types of bearings

S2, residual network construction;

Residual network is a convolutional neural network that directly connects the parameters to the subsequent network layers, which can effectively avoid the problem of network degradation caused by network deepening. So the residual network can use deeper network layers than the general convolutional neural network.

According to the residual block of the residual network, a 20-layer residual network is constructed. In order to extract the low-frequency information of bearing vibration, the convolution kernel of the first layer of the network is a larger convolution kernel of 5 × 5, and the subsequent network layers are is a 3×3 convolution kernel. The input of this network is the grayscale time-frequency map generated in S1, and the output is the result of fault diagnosis.

The residual network structure is shown in Table 2:

Table 2 Residual network structure

S3, the training of residual network;

The data from the training set in S1 is fed into the residual network for training. The number of training steps is set to 500 times, the batch-size of training is 50, the learning rate is 0.01, and the training algorithm of the network adopts the Adam algorithm. The training process of the residual network is shown in Figure 2. As the iteration progresses, the network quickly reaches a stable state, the accuracy rate on the training set also reaches 100%, and the loss function gradually decreases.

After the network is trained, the weight parameters of each layer will change. Figure 3 is a time-frequency diagram of the input network for diagnosis. Figures 4 to 7 are the output of the time-frequency diagram through the first to fourth stages. feature map. Each of these boxes represents an output channel. When the time-frequency map passes through the first stage, the outline of the time-frequency map can be roughly seen. However, as the network layer increases, the network begins to extract high-dimensional features. It can be seen from the figure that as the network deepens, the number of output channels of the network increases, and the features of the deep network are more abstract. This feature is mainly used for computer identification of different fault types.

Figure 8 is the image after dimensionality reduction and visualization of the original data using the t-SNE algorithm, and Figures 9 to 12 are the images of the output features from the first stage to the fourth stage after dimensionality reduction and visualization. Each color in the figure represents a fault type, and the farther apart each color is, the better the fault classification effect is. It can be seen that the raw data in Figure 8 has intersections and adjacent parts between various fault types, so the raw data cannot separate fault types. In Figures 9 to 12, with the deepening of the network layer, various fault types begin to separate slowly. By the time of the last stage of output, all fault types can be separated, and each fault type is far apart, which is not easy to occur. interference. It shows that the classification effect of the network is very good.

S4, residual network test;

After the residual network training is completed, the test set in S1 is input to the network for testing, and its classification accuracy reaches 100%. In order to enable the network to have a stable diagnosis effect even in a noisy environment, this example adds -4dB, -2dB, 0dB, 2dB, and 4dB of Gaussian noise to the dataset for testing. The test results are shown in Table 3:

Table 3 Test results under different noise environments

It can be seen from Table 3 that the network also has a high diagnostic accuracy in a high noise environment (-4dB), and as the noise decreases, the test accuracy gradually increases, and the loss function value gradually decreases. When the noise is The accuracy rate can reach 100% at 4dB, indicating that the characteristic interference of the noise environment on various fault types can be ignored at this time, and it will not affect the diagnosis results of the bearing by the network.

S5. The residual network output is then used for fault diagnosis.

The embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative rather than restrictive. Under the inspiration of the present invention, without departing from the scope of protection of the present invention and the claims, many forms can be made, which all belong to the protection of the present invention.

Claims (2)

1. a residual network rolling bearing fault diagnosis method based on time-frequency analysis, is characterized in that, comprises the following steps: S1. Collect vibration signal data, convert the vibration time domain signal of the rolling bearing into a time-frequency map by using short-time Fourier transform, and convert the time-frequency map into a two-dimensional grayscale time-frequency map; S2. Use the residual network to perform feature extraction on the signal, and diagnose the fault type of the bearing; the input of the residual network is the grayscale time-frequency map generated in step S1, and the output is the result of the fault diagnosis; In the step S1, each grayscale time-frequency map is labeled and divided into a training set and a test set; It also includes step S3: input the data of the training set in step S1 into the residual network for training, the loss function adopts the cross entropy loss function, and the optimization method is the Adam algorithm; after the training is completed, the feature maps of different network layers are drawn, and the t- The SNE algorithm performs dimensionality reduction and visualization on the output of different network layers, and observes the relationship between different network layers; The short-time Fourier transform of the vibration time-domain signal z(t) is:

Among them, t is the time, f is the frequency, γ(t) is the window function, t'-t represents the sliding window, * represents the complex conjugate, and z(t) is the signal; The residual network is a 20-layer residual network, the convolution kernel of the first layer of the residual network is a 5×5 convolution kernel, and the subsequent network layers are a 3×3 convolution kernel. 2. The residual network rolling bearing fault diagnosis method based on time-frequency analysis according to claim 1, further comprising step S4: inputting the test set in step S1 into the residual network to test the accuracy of the residual network .

Images