Rotary machinery fault diagnosis method based on noise enhancement convolutional neural network

Abstract

The invention discloses a rotating machine fault diagnosis method based on a noise enhancement convolutional neural network, and belongs to the technical field of rotating machine fault diagnosis, and the method comprises the following steps: S1, collecting original vibration signals of a rotating machine during operation in different health states; s2, segmenting the collected original vibration signal by using a sliding window to construct a sample data set, converting each piece of sample data in the sample data set into a two-dimensional image, and constructing a training data set and a test data set; s3, training the noise enhancement convolutional neural network by using the training data set, generating noise through network weight parameters and training rounds in the training process, and injecting the noise into the noise enhancement convolutional neural network to enhance the noise enhancement convolutional neural network; and S4, applying the trained noise enhanced convolutional neural network to test data set analysis so as to obtain a fault diagnosis result of test data. According to the method, the generalization of the convolutional neural network model can be enhanced, and the fault diagnosis precision is improved.

Description

Technical Field

[0001] The invention relates to the technical field of rotating machinery fault diagnosis, and in particular to a rotating machinery fault diagnosis method based on a noise enhanced convolutional neural network. Background Art

[0002] With the rapid development of industry, rotating machinery is widely used in various modern production and manufacturing fields due to its high efficiency and continuous working characteristics. Rolling bearings are the core components of rotating machinery. As a key component in rotating machinery that bears loads, transmits motion and maintains equipment accuracy, rolling bearings are prone to failures under long-term working conditions. Rolling bearing failures have always been the main source of failures in rotating machinery. Since the health of rolling bearings is related to the operating status of the entire equipment, rolling bearing fault diagnosis and maintenance have become a key link to ensure the stable operation of the equipment, which is of great significance to ensuring the smooth progress of industrial production.

[0003] With the continuous development of artificial intelligence and the improvement of hardware computing power, deep learning has been widely used in image processing, speech recognition, natural language processing and other fields. Various deep learning methods have also appeared in the field of rolling bearing fault diagnosis. Models such as convolutional neural networks, recursive neural networks, and long short-term memory networks have made rolling bearing fault diagnosis methods more and more diverse and intelligent. In particular, convolutional neural networks have become a classic model in the field of rolling bearing fault diagnosis due to their feature extraction capabilities. However, the application effect of basic models is often not good. How to improve the model and improve the fault diagnosis performance has become an urgent problem to be solved. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a rotating machinery fault diagnosis method based on noise enhanced convolutional neural network. By improving the convolutional neural network, the generalization of the convolutional neural network model can be enhanced and the accuracy of rotating machinery fault diagnosis can be improved.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A method for fault diagnosis of rotating machinery based on noise enhanced convolutional neural network comprises the following steps:

[0007] S1. Collect the original vibration signals of rotating machinery when it is running in different health states;

[0008] S2, segmenting the collected original vibration signal using a sliding window to construct a sample data set, converting each sample data in the sample data set into a two-dimensional image, and constructing a training data set and a test data set;

[0009] S3, training the noise enhanced convolutional neural network using the training data set, and during the training process, generating noise through network weight parameters and training rounds, and injecting the noise enhanced convolutional neural network into the noise enhanced convolutional neural network to enhance it;

[0010] S4. Use the trained noise-enhanced convolutional neural network to analyze the test data set to obtain the fault diagnosis results of the test data.

[0011] A further improvement of the technical solution of the present invention is that in S2, each sample in the sample data set is converted into a two-dimensional image, specifically:

[0012] Grayscale transformation is used to convert each sample data in the sample data set into a grayscale image.

[0013] A further improvement of the technical solution of the present invention is that in S3, the noise generated by the network weight parameters and the training rounds is injected into the noise-enhanced convolutional neural network to enhance it, comprising the following steps:

[0014] In each round of training, the noise intensity parameter is determined by the network weight parameter and the training round number;

[0015] Generate a noise matrix that conforms to Gaussian distribution according to the noise intensity parameter and the preset mean parameter;

[0016] The noise matrix is ​​injected into the convolutional neural network to enhance it.

[0017] A further improvement of the technical solution of the present invention is that in S3, the noise enhanced convolutional neural network includes:

[0018] Convolution module: consists of the first convolution layer, the first BN layer, and the first pooling layer. The output of the first BN layer is activated by ReLU and then input into the first pooling layer for feature extraction of input sample data.

[0019] Convolutional noise module: It consists of a second convolutional layer, a second BN layer, and a second pooling layer. The output of the second BN layer is injected into the noise matrix after ReLU activation and then input into the second pooling layer. The noise matrix is ​​determined by the weight parameters of the second convolutional layer and the training rounds.

[0020] Parallel activation noise adding module: composed of a third convolutional layer, a third BN layer, and a third pooling layer. The output of the third BN layer is injected into P independent and identically distributed noise matrices and then activated by ReLU respectively, and the mean matrix of the obtained P feature matrices is input into the third pooling layer; wherein the P independent and identically distributed noise matrices are determined by the weight parameters of the third convolutional layer and the training rounds;

[0021] Fully connected module: It consists of two fully connected layers and is used to output the rotating machinery fault diagnosis results.

[0022] A further improvement of the technical solution of the present invention is that the noise matrix is ​​determined by the weight parameters of the second convolutional layer and the training rounds, comprising the following steps:

[0023] The weight tensor corresponding to each convolution kernel of the second convolution layer is flattened into a one-dimensional vector, and each one-dimensional vector is converted into a single value through a fully connected layer, and the ratio with the training round is used as the first noise intensity parameter;

[0024] Each first noise intensity parameter and a preset mean parameter generate a noise channel that conforms to a Gaussian distribution;

[0025] Multiple noise channels are concatenated into a noise matrix; wherein the number of channels of the noise matrix is ​​the same as the number of convolution kernels of the second convolution layer; and the size of the noise matrix is ​​the same as the output size of the second BN layer.

[0026] A further improvement of the technical solution of the present invention is that the P independent and identically distributed noise matrices are determined by the weight parameters of the third convolutional layer and the training rounds, comprising the following steps:

[0027] The weight tensor corresponding to each convolution kernel of the third convolution layer is flattened into a one-dimensional vector, and each one-dimensional vector is converted into a single value through a fully connected layer, and the ratio with the training round is used as the second noise intensity parameter;

[0028] Each second noise intensity parameter and a preset mean parameter generate a noise channel that conforms to a Gaussian distribution;

[0029] Multiple noise channels are concatenated into a noise matrix, and P independent and identically distributed noise matrices are cyclically generated; wherein the number of channels of each noise matrix is ​​the same as the number of convolution kernels of the third convolutional layer; and the size of each noise matrix is ​​the same as the output size of the third BN layer.

[0030] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0031] 1. The rotating machinery fault diagnosis method proposed in the present invention improves the performance of the convolutional neural network by injecting noise.

[0032] 2. The present invention designs a method for determining injected noise, in which the injected noise is determined jointly by network weights and training rounds, and the network weight parameters are combined with the injected noise. As the number of training rounds increases, the intensity of the injected noise decreases, and no additional parameters are introduced to generate noise.

[0033] 3. The present invention proposes a combined noise adding method, which injects noise into the convolution operation and activation function in different ways, so that the performance of the convolutional neural network in the classification task is improved, the generalization of the model is enhanced, and the fault diagnosis accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the rotating machinery fault diagnosis method based on noise enhanced convolutional neural network proposed in the present invention;

[0035] Figure 2 It is a structural diagram of the convolution noise adding module proposed in the present invention;

[0036] Figure 3 It is a structural diagram of the parallel activation noise adding module proposed in the present invention;

[0037] Figure 4(a) shows the t-SNE visualization result of the output features of the unnoised convolutional neural network model, and Figure 4(b) shows the t-SNE visualization result of the output features of the noise-enhanced convolutional neural network model;

[0038] Figure 5 It is the confusion matrix of the classification results obtained by the method proposed in this invention. DETAILED DESCRIPTION

[0039] The present invention is further described in detail below with reference to the accompanying drawings and embodiments:

[0040] The terms "first", "second", etc. are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second", etc. may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0041] like Figure 1 As shown, a rotating machinery fault diagnosis method based on noise enhanced convolutional neural network includes the following steps:

[0042] S1. Collect the original vibration signals of rotating machinery when it is running in different health states;

[0043] S2, segmenting the collected original vibration signal using a sliding window to construct a sample data set, converting each sample data in the sample data set into a two-dimensional image, and constructing a training data set and a test data set;

[0044] S3, using the training data set to train the noise enhanced convolutional neural network, and during the training process, generating noise through network weight parameters and training rounds, and injecting it into the noise enhanced convolutional neural network to enhance it;

[0045] S4. Use the trained noise-enhanced convolutional neural network to analyze the test data set to obtain the fault diagnosis results of the test data.

[0046] The noise-enhanced convolutional neural network consists of:

[0047] Convolution module: It consists of the first convolution layer, the first BN layer, and the first pooling layer. The output of the first BN layer is activated by ReLU and then input into the first pooling layer for feature extraction of input sample data.

[0048] Convolutional noise module: It consists of the second convolutional layer, the second BN layer, and the second pooling layer. The output of the second BN layer is injected into the noise matrix after ReLU activation and input into the second pooling layer. The noise matrix is ​​determined by the weight parameters of the second convolutional layer and the training rounds.

[0049] Parallel activation noise adding module: It consists of the third convolutional layer, the third BN layer, and the third pooling layer. The output of the third BN layer is injected into P independent and identically distributed noise matrices and activated by ReLU respectively. The mean matrix of the obtained P feature matrices is input into the third pooling layer. The P independent and identically distributed noise matrices are determined by the weight parameters of the third convolutional layer and the training rounds.

[0050] Fully connected module: It consists of two fully connected layers and is used to output the rotating machinery fault diagnosis results.

[0051] In one embodiment, a rotating machinery fault diagnosis method based on noise enhanced convolutional neural network comprises the following steps:

[0052] S1, first use the vibration acceleration sensor to collect the vibration data of the rolling bearing in different health states, the driving motor speed is 2000r / min, and the sampling frequency is set to 12.8kHz; the equipment under test is a gear-rolling bearing fault simulation test bench, and its health state includes four health states: rolling element fault (BF), rolling bearing inner ring fault (IF), rolling bearing outer ring fault (OF), and normal state (NC).

[0053] S2, use a sliding window with a length of 1024 and a step size of 580 to overlap and intercept the original vibration signal to obtain multiple sub-signals to construct a sample data set; for each sample in the sample data set, reshape the sample data with a length of 1024 to obtain an image of size 32×32, and use grayscale transformation to convert each sample data in the sample data set into a grayscale image:

[0054]

[0055] Where X represents the grayscale image, X(i) represents the pixel in the grayscale image, x(i) represents the data point in the sampling segment, and x min and x max Represent the maximum and minimum values ​​of the data in the sampling points respectively;

[0056] The sample data set is randomly divided to construct a training data set and a test data set. In an optional implementation, the training data set includes 400 grayscale images of each health state, and the test data set includes 150 grayscale images of each health state.

[0057] The convolution module includes the first convolution layer, the first BN layer and the first pooling layer. This module is used to perform preliminary feature extraction on the input image. The first convolution layer has 3 input channels, 16 output channels, and a convolution kernel size of 7×7. The output of the first BN layer is activated by ReLU and then input to the first pooling layer. The first pooling layer can be a maximum pooling layer.

[0058] The structure of the convolution noise module is as follows: Figure 2 As shown, it includes a second convolutional layer, a second BN layer, and a second pooling layer; the input channel of the second convolutional layer is 16, the output channel is 32, and the convolution kernel size is 3×3; the feature matrix obtained by ReLU activation of the output of the second BN layer is injected with noise as the input of the second pooling layer, and the second pooling layer can be a maximum pooling layer.

[0059] The noise is generated by flattening the weight tensor corresponding to each convolution kernel of the second convolutional layer into a one-dimensional vector, converting each one-dimensional vector into a single value through a fully connected layer, and using the ratio of the training round as the noise intensity parameter σ. σ and the preset mean parameter μ generate a noise matrix that conforms to the Gaussian distribution, and its probability density function is:

[0060]

[0061] σ=(ω / T) 2

[0062] Among them, σ is the noise intensity, ω is the single value transformed by the weight parameter of the convolution layer, T is the neural network training round, and μ is the preset mean parameter.

[0063] During the training process, the weight parameters of the convolutional layer tend to be stable, and as the number of training rounds increases, the intensity of the Gaussian noise decreases, and the network model is fixed. μ is the noise mean, which can adjust the average size of the Gaussian noise and affect the performance of the convolutional neural network. In the implementation of the present invention, μ is set to 0.

[0064] In the embodiment of the present invention, in order to ensure noise injection, since the number of convolution kernels contained in the convolution layer is the same as the number of channels of the feature matrix output by the convolution layer, a noise intensity value is generated by each convolution kernel, and this noise intensity value and the noise mean jointly determine a channel of the injected noise matrix, thereby ensuring that the number of channels of the injected noise matrix is ​​the same as the number of channels of the feature matrix, and the size of the generated noise matrix is ​​the same as the size of the feature matrix.

[0065] Specifically, taking the convolution noise module as an example, the output channel of the second convolution layer is 32, that is, the second convolution layer contains 32 convolution kernels, and the number of channels of the output matrix of the second convolution layer is 32. The BN layer does not change the number of channels and size of the feature matrix; when generating the noise matrix, 32 convolution kernels are flattened and passed through the fully connected layer to obtain 32 noise intensity parameters. Each noise intensity parameter and the preset mean parameter generate the noise of a channel, and the noise of these 32 channels are spliced ​​together to obtain the final injected noise matrix. When generating the noise matrix, obtain the size of the feature matrix to ensure that the size of the noise matrix is ​​the same as that of the feature matrix.

[0066] The parallel activation noise adding module structure is as follows Figure 3 As shown, it includes a third convolutional layer, a third BN layer, and a third pooling layer; the input channel of the third convolutional layer is 32, the output channel is 64, and the convolution kernel size is 3×3; the output of the third BN layer is injected into P independent and identically distributed noise matrices and then activated by ReLU respectively, and the mean of the P noise-added feature matrices is input into the third pooling layer. The third pooling layer can be an average pooling layer.

[0067] The method for determining the noise intensity parameter σ is the same as that of the convolution noise module. σ and the preset mean parameter μ are circulated P times to generate P noise matrices that conform to the Gaussian distribution. To ensure noise injection, the number of channels of each noise matrix is ​​the same as the number of convolution kernels of the third convolution layer, and the size of each noise matrix is ​​the same as the output size of the third BN layer. In the implementation, the number of convolution kernels of the third convolution layer, the number of output channels of the third BN layer, and the number of channels of the noise matrix are all 64.

[0068] The activation function of the parallel activation noise adding module can be expressed as:

[0069]

[0070] Among them, X2″ is the feature matrix output by the third BN layer, X2″′(N) is the feature matrix after noise addition, and P is the number of independent and identically distributed noise matrices. is a noise matrix. In an optional implementation of the present invention, P is set to 2.

[0071] The fully connected layer outputs the fault diagnosis results as a linear classifier. The output dimension of the fully connected layer is the same as the number of fault categories. The output value represents the original prediction result of the model on the input data in different categories. The category corresponding to the maximum output value is used as the model's diagnosis result for the input data.

[0072] S3, use the training data set to train the constructed model, and determine the noise through the network weight parameters and training rounds, and inject it into the model to enhance it. In the first round of training, the network parameters are randomly initialized, and the injected noise is determined by the randomly initialized network weights. In each subsequent round of training, the model parameters are optimized using the cross entropy loss and stochastic gradient descent algorithm, and the injected noise is determined by the updated network weights and training rounds. The batch size of the training is 10, the learning rate is 0.01, the weight decay parameter is set to 0.0001, and the other parameters are set to 0 by default.

[0073] S4, using the trained model to test sample sets for rotating machinery fault diagnosis, thereby obtaining fault diagnosis results of the test data.

[0074] The "t-SNE" method is used to visualize the fault features extracted by the model on the test sample set, and a convolutional neural network model consisting of three unnoised convolution modules is set for comparison. The results are shown in Figure 4. It can be seen that in the classification result of the unnoised convolutional neural network in Figure 4 (a), the degree of aliasing of the data of label 0 and label 1 is high, and some data corresponding to label 0 and data corresponding to label 3 are misclassified; while in the result corresponding to the model proposed in Figure 4 (b), the classification effect is generally good, the classification of label 2 and label 3 data is completely correct, and compared with the unnoised model, the degree of aliasing of the data of label 0 and label 1 is low, the data distribution is denser, and the feature extraction effect is better.

[0075] The classification results of the method of the present invention are represented by a confusion matrix, such as Figure 5 As shown in the figure, the confusion matrix is ​​an error matrix. The performance of the learning algorithm can be judged through the confusion matrix, which clearly reflects the degree of consistency between the true value and the predicted value.

[0076] In summary, the present invention can enhance the convolutional neural network by adopting a rotating machinery fault diagnosis method based on a noise-enhanced convolutional neural network, and inject noise in different ways in the convolution operation and activation function, so that the convolutional neural network improves the performance in classification tasks, the generalization of the model is enhanced, and the fault diagnosis accuracy is improved. In addition, the present invention provides a feasible noise determination method and noise addition method for using a noise-enhanced convolutional neural network.

Claims

1. A method for fault diagnosis of rotating machinery based on noise enhanced convolutional neural network, characterized in that: The steps include: S1. Collect the original vibration signals of rotating machinery when it is running in different health states; S2, segmenting the collected original vibration signal using a sliding window to construct a sample data set, converting each sample data in the sample data set into a two-dimensional image, and constructing a training data set and a test data set; S3, training the noise enhanced convolutional neural network using the training data set, and during the training process, generating noise through network weight parameters and training rounds, and injecting the noise enhanced convolutional neural network into the noise enhanced convolutional neural network to enhance it; S4. Use the trained noise-enhanced convolutional neural network to analyze the test data set to obtain the fault diagnosis results of the test data.

2. The method for fault diagnosis of rotating machinery based on noise enhanced convolutional neural network according to claim 1, characterized in that: In S2, each sample in the sample data set is converted into a two-dimensional image, specifically: Grayscale transformation is used to convert each sample data in the sample data set into a grayscale image.

3. The rotating machinery fault diagnosis method based on noise enhanced convolutional neural network according to claim 1 is characterized in that: In S3, the noise is generated by network weight parameters and training rounds, and injected into the noise-enhanced convolutional neural network to enhance it, including the following steps: In each round of training, the noise intensity parameter is determined by the network weight parameter and the training round number; Generate a noise matrix that conforms to Gaussian distribution according to the noise intensity parameter and the preset mean parameter; The noise matrix is ​​injected into the convolutional neural network to enhance it.

4. The method for fault diagnosis of rotating machinery based on noise enhanced convolutional neural network according to claim 3 is characterized in that: In S3, the noise enhanced convolutional neural network includes: Convolution module: consists of the first convolution layer, the first BN layer, and the first pooling layer. The output of the first BN layer is activated by ReLU and then input into the first pooling layer for feature extraction of input sample data. Convolutional noise module: It consists of a second convolutional layer, a second BN layer, and a second pooling layer. The output of the second BN layer is injected into the noise matrix after ReLU activation and then input into the second pooling layer. The noise matrix is ​​determined by the weight parameters of the second convolutional layer and the training rounds. Parallel activation noise adding module: composed of a third convolutional layer, a third BN layer, and a third pooling layer. The output of the third BN layer is injected into P independent and identically distributed noise matrices and then activated by ReLU respectively, and the mean matrix of the obtained P feature matrices is input into the third pooling layer; wherein the P independent and identically distributed noise matrices are determined by the weight parameters of the third convolutional layer and the training rounds; Fully connected module: It consists of two fully connected layers and is used to output the rotating machinery fault diagnosis results.

5. The method for fault diagnosis of rotating machinery based on noise enhanced convolutional neural network according to claim 4 is characterized in that: The noise matrix is ​​determined by the weight parameters of the second convolutional layer and the training rounds, comprising the following steps: The weight tensor corresponding to each convolution kernel of the second convolution layer is flattened into a one-dimensional vector, and each one-dimensional vector is converted into a single value through a fully connected layer, and the ratio with the training round is used as the first noise intensity parameter; Each first noise intensity parameter and a preset mean parameter generate a noise channel that conforms to a Gaussian distribution; Multiple noise channels are concatenated into a noise matrix; wherein the number of channels of the noise matrix is ​​the same as the number of convolution kernels of the second convolution layer; and the size of the noise matrix is ​​the same as the output size of the second BN layer.

6. The method for fault diagnosis of rotating machinery based on noise enhanced convolutional neural network according to claim 4, characterized in that: The P independent and identically distributed noise matrices are determined by the weight parameters of the third convolutional layer and the training rounds, comprising the following steps: The weight tensor corresponding to each convolution kernel of the third convolution layer is flattened into a one-dimensional vector, and each one-dimensional vector is converted into a single value through a fully connected layer, and the ratio with the training round is used as the second noise intensity parameter; Each second noise intensity parameter and a preset mean parameter generate a noise channel that conforms to a Gaussian distribution; Multiple noise channels are concatenated into a noise matrix, and P independent and identically distributed noise matrices are cyclically generated; wherein the number of channels of each noise matrix is ​​the same as the number of convolution kernels of the third convolutional layer; and the size of each noise matrix is ​​the same as the output size of the third BN layer.

Images