A rotating machinery fault diagnosis method based on one-dimensional attention mechanism and convolutional neural network

Abstract

The application discloses a rotating machinery fault diagnosis method based on a one-dimensional attention mechanism and a convolutional neural network, which comprises the following steps: collecting original waveform data of rotating machinery in operation to obtain original vibration data; performing data set expansion on the original data; and establishing a novel one-dimensional convolutional neural network model comprising a one-dimensional channel and a time attention mechanism.In the network model of the application, single-channel data can form a multi-channel feature map through one-dimensional convolution, and then the weight of each channel is reasonably changed, so that the one-dimensional channel attention mechanism can adaptively improve the sensitivity of the network to different channels, thereby optimizing the network performance; and the channel attention mechanism and the time attention mechanism are used in combination, which can greatly improve the diagnosis accuracy and the training speed.The application is an original intelligent diagnosis model, which improves the network performance through the attention mechanism, so that the network is more suitable for a fault diagnosis task, and intelligent fault identification is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rotating machinery fault diagnosis, and particularly relates to a rotating machinery fault diagnosis method based on one-dimensional attention mechanism and convolutional neural network. BACKGROUND

[0002] Various rotating machines represented by bearings and gears are the most widely used components in industrial equipment. Once a fault occurs, it will lead to equipment failure, resulting in economic losses and even safety accidents. Therefore, it is of great practical significance to carry out fault diagnosis and condition monitoring.

[0003] Traditional intelligent fault diagnosis systems mainly include data acquisition, feature extraction and fault recognition. Among them, feature extraction and fault recognition are the most important and have a great influence on the accuracy of fault diagnosis. However, the traditional fault diagnosis system has the following shortcomings: 1) the diagnosis performance depends on the design of the feature extraction method by domain experts, and for each new fault diagnosis task, the method suitable for the task must be redesigned, resulting in an increase in time and labor costs; 2) manual feature extraction is difficult to comprehensively represent the running state and kinematics of rotating machines, which further leads to a decrease in fault diagnosis accuracy; 3) the traditional fault diagnosis system is too dependent on prior knowledge from design to use, and it takes a long time to train fault diagnosis personnel; 4) traditional intelligent recognition models such as support vector machine and random forest are difficult to learn enough features, which leads to an ineffective improvement in recognition accuracy.

[0004] Deep learning theory has been successful in the fields of computer vision, natural language processing, speech recognition, and protein spatial structure prediction. Convolutional neural networks are widely used in various classification models due to their powerful feature extraction capabilities.

[0005] There have been fault diagnosis methods based on attention mechanism, but basically one-dimensional signals need to be converted into two-dimensional pictures through signal analysis methods for fault diagnosis. Common conversion methods include fast Fourier transform, discrete wavelet transform, etc., which generate two-dimensional three-channel RGB pictures by using the horizontal axis to represent time, the vertical axis to represent frequency, and color to represent amplitude. Such pictures are called time-frequency spectra. However, this method has certain defects: the resolution of the time-frequency spectrum generated by the fast Fourier transform is not adjustable in time and frequency domains, so the resolution needs to be set again for different devices and parts, which leads to poor generalization of the diagnosis network; the discrete wavelet transform needs to be decomposed according to the signal characteristics, which greatly reduces the model generalization ability; in addition, there are other conversion methods, but information loss occurs during signal conversion, and the model complexity is increased, which leads to a decrease in overall network efficiency and an increase in memory usage. SUMMARY

[0006] In view of the above defects of the prior art, the application provides a rotating machinery fault diagnosis method based on a one-dimensional attention mechanism and a convolutional neural network, which adds a channel attention mechanism and a time attention mechanism to avoid various problems in the process of converting a one-dimensional signal into a two-dimensional signal, while ensuring the training speed and diagnosis accuracy.

[0007] To achieve the above object, the application provides a rotating machinery fault diagnosis method based on a one-dimensional attention mechanism and a convolutional neural network, comprising:

[0008] Collecting a vibration signal of the rotating machinery during operation;

[0009] Overlapping and slicing the vibration signal to expand a data set;

[0010] Extracting a vibration feature from the expanded vibration signal;

[0011] Inputting the vibration feature into a deep neural network model for training to obtain an intelligent diagnosis model;

[0012] Using the intelligent diagnosis model to diagnose a rotating machinery fault.

[0013] Further, the vibration signal of the rotating machinery during operation is collected, specifically: collecting the vibration signal of the rotating machinery during operation at a sampling frequency f s Collecting an acceleration signal set X of the rotating machinery in different rotating speeds and different working conditions under each fault state n = {x n1 ,x n2 ,...x nm}, wherein n represents a fault state number, and m represents a sampling point number of a column of acceleration signals under a certain fault state.

[0014] Further, the vibration signal is overlapped and sliced to expand the data set, specifically: for a column of acceleration signals, 2048 points are taken as a group to cut into a single sample, and the starting point of each sample is spaced by 256 points; a column of acceleration signals can be expanded into data, and the expanded data is divided into several batches, wherein [] represents rounding.

[0015] Further, the vibration feature is input into a deep neural network model for training, specifically: the vibration feature is input into a one-dimensional convolutional layer a in the deep neural network model, the input channel number of the layer is 1, the output channel number is 128, the kernel size is 64, the step is 4, the activation function is ReLU, and a batch normalization layer is added;

[0016] The one-dimensional convolution is described as:

[0017]

[0018] wherein and respectively represent the weight and bias of each neuron in the i-th convolution kernel of the l-th layer, and σ represents an activation function, x l (j) represents the input of the j-th neuron of the l-th layer, represents a convolution operation;

[0019] The activation function ReLU, i.e. linear rectifier unit, is expressed as:

[0020]

[0021] ReLU has the following advantages compared with sigmoid and tanh: ① the problem of gradient disappearance basically does not occur, ② the overfitting phenomenon can be alleviated, and the training speed is more stable;

[0022] The batch normalization is described as:

[0023]

[0024] wherein μ represents the sample mean, σ represents the sample variance, γ is a "stretching parameter", and β is an "offset parameter", and γ and β are parameters participating in training; it should be noted that the batch normalization layer is located before the activation function;

[0025] The feature map obtained after processing by the one-dimensional convolution layer a is input into the maximum pooling layer a, and the pooling kernel size in the layer is 3 and the pooling step is 2;

[0026] The maximum pooling is expressed as:

[0027]

[0028] wherein p represents the pooling output, n represents the pooling starting position, k represents the pooling kernel size, and z(i) represents the input of the i-th neuron; the feature map after pooling can remove redundant information, reduce the network complexity, and reduce the memory usage;

[0029] The feature map obtained after processing by the maximum pooling layer a is input into the one-dimensional convolution layer b, the input channel number of the layer is 128, the output channel number is 64, the kernel size is 3, the step is 2, the activation function is the ReLU function, and the batch normalization layer is added;

[0030] The feature map obtained after processing by the one-dimensional convolution layer b is input into the maximum pooling layer b, and the pooling kernel size in the layer is 3 and the pooling step is 2;

[0031] The feature map processed by the max pooling layer b is input into a one-dimensional convolution layer c, the input channel number of the layer is 64, the output channel number is 32, the kernel size is 3, the step is 2, and the activation function is a ReLU function;

[0032] The feature map processed by the one-dimensional convolution layer c is input into a max pooling layer c, the pooling kernel size of the layer is 3, and the pooling step is 2;

[0033] The feature map processed by the max pooling layer c is input into a channel attention module to adjust the weight of each channel;

[0034] The feature map after adjusting the channel weight is input into a time attention module to adjust the time weight;

[0035] The feature map after adjusting the time weight is flattened;

[0036] The flattened feature map is input into a full connection layer a, the output channel of the full connection layer a is 1024, the activation function is a ReLU function, and the Dropout technology is used to suppress overfitting of the layer;

[0037] The feature map processed by the full connection layer a is input into a full connection layer b, the output channel of the full connection layer b is 512, the activation function is a ReLU function, and the Dropout technology is used to suppress overfitting of the layer;

[0038] The feature map processed by the full connection layer b is input into a full connection layer c, and the output channel of the full connection layer c is the number of fault categories;

[0039] The cross-entropy loss function is used to evaluate the error between the estimated probability distribution and the actual probability distribution, and then the error is optimized using the SGD algorithm, and the above training steps are repeatedly cycled until the error is less than a set value, the weight file is saved, and an intelligent diagnosis model is obtained.

[0040] Further, the expression of the cross-entropy loss function is:

[0041]

[0042] Where N is the total number of samples, M is the number of categories, y ic 1 when the real category of sample i is the same as the predicted category, otherwise 0, p ic represents the probability that sample i belongs to category c. The cross-entropy loss function is a convex function, so a global optimal solution can be obtained.

[0043] Further, the specific implementation steps in the channel attention module are:

[0044] The feature map input into the channel attention module is globally max-pooled;

[0045] Then the number of channels is reduced to 1 / 4 through a full connection layer, and the reduced feature map is input into an activation function ReLU;

[0046] Then the number of channels is restored through another full connection layer, and the sigmoid function is used to maintain the weight value between (0, 1).

[0047] The channel weight value is multiplied by the channel of the input signal to obtain the output of the channel attention module.

[0048] Further, the global maximum pooling can unify the values of each column in each channel to the maximum value of the column, so as to avoid the influence of the difference between the original values on the subsequent operation. Its expression is:

[0049]

[0050] Wherein, n represents the feature map number, c represents the channel number, and w represents the column number.

[0051] Further, the specific implementation steps in the time attention module are:

[0052] The channel dimension of the feature map input into the time attention module is globally maximum pooled and globally average pooled;

[0053] The globally maximum pooled feature map and the globally average pooled feature map are spliced by column;

[0054] One-dimensional convolution operation is performed on the spliced feature map to obtain the time weight;

[0055] The sigmoid activation function is used to normalize the time weight;

[0056] The input feature map is multiplied by the normalized time weight to obtain the output of the time attention module.

[0057] Compared with the prior art, the network model of the present application has the advantages that: in the network model of the present application, single-channel data can form a multi-channel feature map through one-dimensional convolution, and then the weights of each channel are reasonably changed, so that the one-dimensional channel attention mechanism can adaptively improve the sensitivity of the network to different channels, thereby optimizing the network performance; at the same time, in actual working scenarios, rotating mechanical movement often contains random interference, and the one-dimensional time attention mechanism can make the network model automatically focus on the vibration signals containing fault features, and can effectively extract fault features. The present application mixes the channel attention mechanism and the time attention mechanism, which can greatly improve the diagnostic accuracy and training speed. The measured data is divided into a training set and a test set, and the model is trained and tested; if the model test precision reaches the set threshold, the training is considered complete. The intelligent diagnosis model of the present application improves the network performance through the attention mechanism, making it more suitable for fault diagnosis tasks, and realizes intelligent fault recognition. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is the flowchart of the present application;

[0059] Figure 2 is the data expansion diagram of the present application;

[0060] Figure 3 is the overall framework diagram of the neural network;

[0061] Figure 4 is the channel attention mechanism diagram;

[0062] Figure 5 is the time attention mechanism diagram;

[0063] Figure 6 is the training process diagram. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and are not used to limit the present application, that is, the described examples are only a part of the examples of the present application, but not all the examples.

[0065] Example 1

[0066] As shown in Figure 1 , the present application provides a rotating machinery fault diagnosis method based on one-dimensional attention mechanism and convolutional neural network, which specifically comprises:

[0067] S1. Collect the vibration signal of the rotating machinery during operation;

[0068] Specifically, the embodiment adopts the bearing data set disclosed by Case Western Reserve University. The fault classification mode is shown in Table 1,

[0069] Table 1 Classification of the bearing fault data set of Case Western Reserve University

[0070]

[0071] S2. Overlapping slicing is performed on the vibration signal to expand the data set;

[0072] Specifically, for a column of acceleration signals, each sample is cut into a group of 2048 points, and each sample starts at an interval of 256 points. A column of acceleration signals can be expanded to data, and the expanded data is divided into several batches, where [] represents rounding.

[0073] S3. Vibration features are extracted from the expanded vibration signal;

[0074] S4. The vibration features are input into a deep neural network model for training to obtain an intelligent diagnosis model, specifically including

[0075] S4.1 The vibration features are input into a one-dimensional convolution layer a in the deep neural network model, the input channel number of which is 1, the output channel number is 128, the kernel size is 64, the step is 4, the activation function is ReLU, and a batch normalization layer is added.

[0076] The one-dimensional convolution is described as:

[0077]

[0078] wherein and respectively represent the weight and bias of each neuron in the i-th convolution kernel of the l-th layer, σ represents the activation function, and x l (j) represents the input of the j-th neuron of the l-th layer, represents the convolution operation;

[0079] The activation function ReLU, i.e., linear rectifier unit, is expressed as:

[0080]

[0081] ReLU has the following advantages compared with sigmoid and tanh: ①, the problem of gradient disappearance basically does not occur, ②, the overfitting phenomenon can be alleviated, and the training speed is more stable.

[0082] The batch normalization is described as:

[0083]

[0084] wherein μ denotes sample mean, σ denotes sample variance; γ is a "stretching parameter", β is an "offset parameter", γ and β are parameters involved in training; the batch normalization layer is located before the activation function;

[0085] S4.2 The feature map obtained by processing through the one-dimensional convolutional layer a is input into the max pooling layer a, and the pooling kernel size in the layer is 3 and the pooling step is 2;

[0086] The max pooling is expressed as:

[0087]

[0088] wherein p denotes the pooling output, n denotes the pooling starting position, k denotes the pooling kernel size, and z(i) denotes the input of the i-th neuron; the feature map after pooling can remove redundant information, reduce network complexity, and reduce memory usage.

[0089] S4.3 The feature map obtained by processing through the max pooling layer a is input into the one-dimensional convolutional layer b, the input channel number of the layer is 128, the output channel number is 64, the kernel size is 3, the step is 2, the activation function is ReLU function, and a batch normalization layer is added;

[0090] S4.4 The feature map obtained by processing through the one-dimensional convolutional layer b is input into the max pooling layer b, and the pooling kernel size in the layer is 3 and the pooling step is 2;

[0091] S4.5 The feature map obtained by processing through the max pooling layer b is input into the one-dimensional convolutional layer c, the input channel number of the layer is 64, the output channel number is 32, the kernel size is 3, the step is 2, and the activation function is ReLU function;

[0092] S4.6 The feature map obtained by processing through the one-dimensional convolutional layer c is input into the max pooling layer c, and the pooling kernel size in the layer is 3 and the pooling step is 2;

[0093] S4.7 The feature map obtained by processing through the max pooling layer c is input into the channel attention module to adjust the weight of each channel;

[0094] Specifically, the implementation process of the channel attention module is as follows:

[0095] S4.71 The feature map input into the channel attention module is globally max-pooled;

[0096] S4.72 Then, the channel number is reduced to 1 / 4 through a fully connected layer, and the reduced feature map is input into the activation function ReLU;

[0097] S4.73 The channel number is restored through another fully connected layer, and the sigmoid function is used to maintain the weight between (0, 1).

[0098] S4.74 multiply each channel weight with each channel of the input signal to obtain the output of the channel attention module.

[0099] In summary, the channel attention module works as follows: after global max pooling, the shape of the feature map is set to Cx1x1, at this time each channel has the same weight. In the training process (i.e. S4.1-S4.12), each channel will be given different weights, the channel weight that has a positive impact on the output accuracy will be increased, otherwise the weight will be reduced, thereby improving the model accuracy.

[0100] S4.8 input the feature map after adjusting the channel weight into the time attention module to adjust the time weight;

[0101] Specifically, the implementation process of the time attention module is as follows:

[0102] S4.81 perform global max pooling and global average pooling on the channel dimension of the feature map input into the time attention module;

[0103] S4.82 concatenate the feature map after global max pooling with the feature map after global average pooling by column;

[0104] S4.83 perform one-dimensional convolution operation on the concatenated feature map to obtain the time weight;

[0105] S4.84 normalize the time weight using the sigmoid activation function;

[0106] S4.85 multiply the input feature map with the normalized time weight to obtain the output of the time attention module.

[0107] In summary, the time attention module works as follows: the shape of the input feature map is CxT=NxM, where C represents the number of channels and T represents the number of feature map points in the time dimension. In order to reduce the complexity of the model and retain the necessary time dimension information, first, global max pooling and global average pooling are performed on the channel dimension, then the pooled feature maps are concatenated by column, at this time the shape of the feature map channel is set to CxT=2xM; then one-dimensional convolution is used on the new feature map to extract features and the weight in the time dimension is maintained between (0, 1) through the sigmoid function; finally, the obtained weight is multiplied by the input feature map to obtain the output of the time attention module. In the training process (i.e. S4.1-S4.12), the part containing the fault feature in the time dimension will be given a higher weight, thereby improving the network training speed and model accuracy.

[0108] S4.9 flatten the feature map after adjusting the time weight;

[0109] S4.10 The flattened feature map is input into a full connection layer a, which outputs 1024 channels with a ReLU function as the activation function, and the Dropout technique is used to suppress overfitting for this layer;

[0110] S4.11 The feature map processed by the full connection layer a is input into a full connection layer b, which outputs 512 channels with a ReLU function as the activation function, and the Dropout technique is used to suppress overfitting for this layer;

[0111] S4.12 The feature map processed by the full connection layer b is input into a full connection layer c, which outputs the number of fault categories;

[0112] S4.13 The error between the estimated probability distribution and the actual probability distribution is evaluated using the cross-entropy loss function, and the error is optimized using the SGD algorithm. The training steps S4.1-S4.12 are repeatedly cycled until the error is less than a set value, the weight file is saved, and the intelligent diagnosis model is obtained.

[0113] S5. The intelligent diagnosis model is used for rotating machinery fault diagnosis.

[0114] Specifically, a data set signal is randomly extracted, a plurality of segments are intercepted and input into the intelligent diagnosis model, the model gives a fault diagnosis result, and the probability of a correct diagnosis result can be obtained according to the probability and statistics theory.

[0115] The foregoing description of specific exemplary embodiments of the application is intended to be illustrative only and is not intended to limit the application to the precise forms described. Many modifications and variations are possible in light of the above teachings without departing from the spirit or essential characteristics of the application. The exemplary embodiments were chosen and described in order to explain the principles of the application and its practical application and to allow others skilled in the art to understand the application for various exemplary embodiments with various modifications being applicable. The scope of the application is intended to be defined by the claims and their equivalents.

Claims

1. A rotating machinery fault diagnosis method based on one-dimensional attention mechanism and convolutional neural network, characterized in that, The method comprises the following steps: collecting vibration signals of a rotating machine during operation; overlapping and slicing the vibration signals to expand the data set; extracting vibration features from the expanded vibration signals; inputting the vibration features into a deep neural network model for training to obtain an intelligent diagnosis model; using the intelligent diagnosis model to diagnose the rotating machine fault; training the deep neural network model by inputting the vibration features into a one-dimensional convolution layer a in the deep neural network model, wherein the input channel number of the one-dimensional convolution layer a is 1, the output channel number is 128, the kernel size is 64, the step is 4, the activation function is ReLU, and a batch normalization layer is added; the one-dimensional convolution is described as: wherein with respectively represent the weight and bias of each neuron in the i-th convolutional kernel of the l-th layer, σ represents an activation function, x l (j) represents the input of the j-th neuron of the l-th layer, represents a convolution operation; the activation function ReLU is a linear rectifier unit, and its expression is: the batch normalization is described as: wherein μ represents the sample mean, σ represents the sample variance, γ is a "stretching parameter", β is an "offset parameter", and γ and β are parameters participating in training; the batch normalization layer is located before the activation function; the feature map obtained after the one-dimensional convolution layer a is processed is input into a max-pooling layer a, wherein the pooling kernel size in the max-pooling layer a is 3, and the pooling step is 2; the max-pooling is expressed as: wherein p represents the pooling output, n represents the pooling starting position, k represents the pooling kernel size, and z(i) represents the input of the i-th neuron; the feature map obtained after the max-pooling layer a is processed is input into a one-dimensional convolution layer b, wherein the input channel number of the one-dimensional convolution layer b is 128, the output channel number is 64, the kernel size is 3, the step is 2, the activation function is ReLU function, and a batch normalization layer is added; the feature map obtained after the one-dimensional convolution layer b is processed is input into a max-pooling layer b, wherein the pooling kernel size in the max-pooling layer b is 3, and the pooling step is 2; the feature map obtained after the max-pooling layer b is processed is input into a one-dimensional convolution layer c, wherein the input channel number of the one-dimensional convolution layer c is 64, the output channel number is 32, the kernel size is 3, the step is 2, and the activation function is ReLU function; the feature map obtained after the one-dimensional convolution layer c is processed is input into a max-pooling layer c, wherein the pooling kernel size in the max-pooling layer c is 3, and the pooling step is 2; the feature map obtained after the max-pooling layer c is processed is input into a channel attention module to adjust the channel weight; the feature map after adjusting the channel weight is input into a time attention module to adjust the time weight; the feature map after adjusting the time weight is flattened; the flattened feature map is input into a full connection layer a, wherein the output channel of the full connection layer a is 1024, the activation function is ReLU function, and the Dropout technology is used to suppress overfitting for the layer; the feature map obtained after the full connection layer a is processed is input into a full connection layer b, wherein the output channel of the full connection layer b is 512, the activation function is ReLU function, and the Dropout technology is used to suppress overfitting for the layer; the feature map obtained after the full connection layer b is processed is input into a full connection layer c, wherein the output channel of the full connection layer c is the number of fault categories; a cross-entropy loss function is used to evaluate the error between the estimated probability distribution and the actual probability distribution, and then the error is optimized by using an SGD algorithm, and the training steps are repeatedly cycled until the error is less than a set value, a weight file is saved, and an intelligent diagnosis model is obtained; the specific implementation steps in the channel attention module are: The feature map of the input channel attention module is globally maximum-pooled; Then, the channel number is reduced to 1 / 4 through a fully connected layer, and the reduced feature map is input into an activation function ReLU; Then, the channel number is restored through another fully connected layer, and the sigmoid function is used to maintain the weight between (0, 1); The channel weight is multiplied by each channel of the input signal to obtain the output of the channel attention module; The specific implementation steps in the time attention module are as follows: The channel dimension of the feature map input into the time attention module is globally maximum-pooled and globally average-pooled; The globally maximum-pooled feature map and the globally average-pooled feature map are spliced by column; One-dimensional convolution operation is performed on the spliced feature map to obtain time weight; The sigmoid activation function is used to normalize the time weight; The input feature map is multiplied by the normalized time weight to obtain the output of the time attention module.

2. The rotating machinery fault diagnosis method based on one-dimensional attention mechanism and convolutional neural network according to claim 1, characterized in that, The vibration signal of the rotating machinery in operation is collected, specifically: collecting the vibration signal of the rotating machinery in operation at a sampling frequency f s The acceleration signal set X of the rotating machinery in each fault state under different rotating speeds and different working conditions is collected n ={x n1 ,x n2 ,...x nm}, wherein n represents the fault state number, and m represents the sampling point number of a column of acceleration signals under a certain fault state.

3. The rotating machinery fault diagnosis method based on one-dimensional attention mechanism and convolutional neural network according to claim 1, characterized in that, The expression of the cross-entropy loss function is: where N is the total number of samples, M is the number of classes, y ic 1 if the true class of sample i is the same as the predicted class, otherwise 0, p ic denotes the probability that sample i belongs to class c.

4. The rotating machinery fault diagnosis method based on one-dimensional attention mechanism and convolutional neural network according to claim 3, characterized in that, The expression of the global maximum pooling is: Wherein, n represents the feature map number, c represents the channel number, and w represents the column number.

Images