A bearing fault diagnosis classification method based on multi-scale features and attention

Abstract

The application relates to a bearing fault diagnosis classification method based on multi-scale features and attention, which comprises the following steps: S1, collecting time domain signal data of a bearing in real time through an acceleration sensor; S2, extracting, analyzing and processing the pretreated data through a pre-constructed multi-scale feature classification module to obtain first data; S3, inputting the first data into a Transformer attention mechanism learning module for learning to obtain second data; and S4, inputting the second data into a full connection layer to output a diagnosis classification result of bearing faults; the application uses a convolutional neural network structure to construct a deep multi-scale feature extraction module, adopts a strategy of different size convolution kernels to mine shallow fault feature information, then introduces a pure attention mechanism to deeply filter fault features, retains the most representative features of the same fault in different working conditions, and can complete the classification of bearing faults in different working conditions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bearing fault diagnosis classification and equipment health management, and particularly relates to a bearing fault diagnosis classification method based on multi-scale features and attention. BACKGROUND

[0002] Rolling bearings are one of the most important components in rotating machinery and equipment, but rolling bearings are easily damaged by working environment during working process, thereby causing mechanical failure. According to statistics, 40% of motor failures are bearing failures. Due to complex working environment, rolling element bearings are inevitably subject to failure, and the environment in which the failure occurs is unpredictable. The most common method to prevent possible damage is to monitor vibration in real time when the rotating mechanism is running. Intelligent fault diagnosis methods are applied to identify fault types using state signals collected by sensors. Traditional intelligent fault diagnosis methods can be divided into two steps: feature extraction and classification. In the bearing intelligent fault diagnosis algorithm, feature extraction based on signal processing is a necessary link, but this link is time-consuming and requires high expert prior experience, cannot meet the diagnosis performance requirements under different working conditions, and will seriously burden the real-time performance and universality of the diagnosis system, and has been unable to meet the growing requirements of mechanical big data.

[0003] In recent years, more and more related researches of deep learning are applied in smart factories and big data fault diagnosis. The characteristics of this kind of technology are that the most relevant features can be automatically extracted from original vibration time domain signals or time-frequency domain information, and in the process of continuous parameter optimization and learning, the most relevant features for bearing fault diagnosis can be captured, which replaces the tedious feature engineering of traditional algorithms. However, some deep learning related methods are single in feature filtering, and some key information is missing due to noise in the network training process, so they are often only suitable for a type of bearing fault detection under a specific working condition, and have great limitations, and cannot meet the needs under different complex working conditions. SUMMARY

[0004] On the basis of the prior art, the application provides a bearing fault diagnosis classification method under multiple working conditions based on multi-scale feature extraction and attention mechanism, which can extract key information from a large amount of data under different working conditions, and realize bearing fault diagnosis under multiple working conditions.

[0005] The application achieves the above-mentioned purposes through the following technical solutions:

[0006] A bearing fault diagnosis classification method based on multi-scale features and attention, comprising the following steps:

[0007] S1, collecting time domain signal data of the bearing in real time through an acceleration sensor, pre-processing, and obtaining pre-processed data;

[0008] S2, extracting and analyzing the preprocessed data through a pre-constructed multi-scale feature classification module to obtain first data, wherein the pre-constructed multi-scale feature classification module sequentially comprises a convolution layer with a large convolution kernel and three convolution layers with the same small convolution kernel;

[0009] S3, inputting the first data into a Transformer attention mechanism learning module for learning to obtain second data;

[0010] S4, inputting the second data into a full connection layer to obtain a final four-dimensional classification result vector, and outputting a diagnosis classification result of the bearing fault.

[0011] As a further optimization scheme of the present application, the preprocessing of S1 comprises the following steps:

[0012] A time sliding window with a size of 1024 is used to slide along the time axis, and the sliding step is the window size, and each sliding generates a time domain signal input X, X = [ , ,……, ] Wherein n is 1024;

[0013] The input time domain signal X is normalized to [-1, 1], and the formula for normalization is as follows:

[0014] ;

[0015] Obtain a time domain feature set input sequence as preprocessed data.

[0016] As a further optimization scheme of the present application, S2 specifically comprises:

[0017] The preprocessed data is input into the multi-scale feature extraction module, and sequentially passes through a convolution layer with a convolution kernel size of 15 and three consecutive convolution layers with a convolution kernel size of 3;

[0018] Wherein, the output of the first convolution layer is , the output of the second convolution layer is , the output of the third convolution layer is , the output of the fourth convolution layer is , as the first data;

[0019] The one-dimensional forward propagation of the neuron input from the convolution layer l-1 to the layer l is expressed by the following formula:

[0020] ;

[0021] Wherein, is the a bias scalar of the neuron; an output of a neuron of a previous layer; an output of a neuron of a previous layer; a kernel from a neuron of a previous layer to a neuron of the l-th layer a neuron of the l-th layer ; is a ReLU activation function.

[0022] As a further optimization scheme of the present application, the S3 is specifically:

[0023] The first data is identified as an input sequence using an attention mechanism, and is regarded as a token sequence

[0024] A trainable and randomly initialized token-like head is added to the head of the first data, that is, to compose [ , ], position encoding information is added , and + is obtained

[0025] Enter the multi-head attention mechanism module in the Transformer structure, and perform self-attention calculation on it according to the following formula:

[0026] ;

[0027] wherein, , and are parameter matrices to be learned, softmax= , Z is a vector, and Z i , Z j is an element.

[0028] In order to make the vector information of different subspaces shared, further parallel attention calculation, that is, multi-head attention, is adopted, and calculation is performed according to the following formula:

[0029] ;

[0030] wherein is , is a parameter matrix.

[0031] The output feature vector after the Transformer structure is taken as the second data.​​

[0032] As a further optimization scheme of the application, the regularization layer and the ReLU activation function are contained in each convolutional layer, and the maximum pooling and adaptive pooling operations are performed on each convolutional layer.

[0033] The application has the beneficial effects that:

[0034] A deep multi-scale feature extraction model is constructed using a convolutional neural network structure, aiming to perform data preprocessing on input time domain signals under different working conditions, adopt a strategy of different size convolution kernels to mine shallow fault feature information, and then introduce a pure attention mechanism to deeply filter fault features and retain the most representative features of the same fault in different working conditions. The global context capture of the attention mechanism and the local dependence capture of the convolution are combined, which can complete the classification of bearing faults under different working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a time domain feature schematic diagram for outer ring fault;

[0036] Figure 2 is a time domain feature schematic diagram for inner ring fault;

[0037] Figure 3 is a time domain feature schematic diagram for roller fault;

[0038] Figure 4 is a bearing fault diagnosis and classification method flowchart based on multi-scale features and attention of the application;

[0039] Figure 5 is a time domain signal division schematic diagram using a sliding window;

[0040] Figure 6 is a deep learning network model structure diagram;

[0041] Figure 7 is a standard Transformer structure;

[0042] Figure 8 is a Case Western Reserve University bearing fault data acquisition platform site diagram;

[0043] Figure 9 is a 12k drive end bearing fault data classification result confusion matrix;

[0044] Figure 10 is a 12k fan end bearing fault data classification result confusion matrix;

[0045] Figure 11 is a 48k drive end bearing fault data classification result confusion matrix. DETAILED DESCRIPTION

[0046] The application will be described in further detail below with reference to the drawings. It is necessary to point out here that the following detailed description is only used to further illustrate the application and cannot be understood as limiting the scope of protection of the application, and the skilled in the art can make some non-essential improvements and adjustments to the application according to the above application content.

[0047] In order to reduce the computational overhead, the time-domain signal on the time scale is taken as the model input, and no much signal processing is needed for the data except removing the direct current component caused by the sensor deviation. Among them, the time-domain signal as the input is the vibration original signal collected from the acceleration sensor installed on the bearing seat of the mechanical equipment, and the time-domain feature extraction comes from the time-domain signal, that is, taking time as an independent variable to reflect the relationship between the signal amplitude and the time variable. The time-domain feature has a clear trend of change over time, and the multi-layer convolution layer design can automatically extract different fault features without additional data preprocessing and statistical feature extraction steps.

[0048] Principle of time-frequency signal containing fault characteristics

[0049] When the rolling bearing fails in use, the amplitude time-domain signal of the same type of fault type will often exhibit some identical characteristics, and these characteristics contain effective information that helps to classify and judge the fault type. The three fault modes of the bearing will exhibit different fault characteristics in the vibration time-domain signal, i.e. the outer ring, the inner ring and the roller. Among them, the impact of various faults will cause the vibration time-domain signal to exhibit a relatively regular periodic change with different vibration characteristics. Next, the bearing vibration time-domain signals under the conditions of outer ring, inner ring and roller faults will be compared and analyzed.

[0050] Analysis of time-domain signal of outer ring fault

[0051] When the rolling bearing works, the position of the outer ring changes relatively little, and only when the rolling element rotates to the outer ring fault position will it cause an impact. At this time, the vibration time-domain signal has obvious characteristics, which is shown as a periodic impact response characteristic in the time-domain waveform graph, as shown in Figure 1 .

[0052] Analysis of time-domain signal of inner ring fault

[0053] Compared with the outer ring fault, the main change of the inner ring fault is in its rotation state. Due to the regular rotation of the inner ring, the periodic change of the system pulse excitation force and the load pressure borne by the fault position is caused. In summary, because of these characteristics of rotation, the inner ring is far from the vibration sensor and close to the vibration sensor, so it has a certain impact on the time-domain signal, as shown in Figure 2 .

[0054] Analysis of time-domain signal of roller fault

[0055] Rollers are loaded with friction and radial, axial forces on the raceway between the inner and outer rings during operation, which will produce periodic impact. At the same time, the rollers are uniformly separated by the cage, and the rotation is realized, so the load of the fault part will appear periodic change, and the change of load distribution will cause the amplitude fluctuation of certain characteristics of time domain signal, but due to the complexity of roller fault, the change of time domain characteristics of fault is not as obvious as the signal change of inner and outer ring fault, as shown in Figure 3 .

[0056] Bearing fault diagnosis classification method based on multi-scale feature and attention

[0057] The bearing fault diagnosis classification method based on multi-scale feature and attention, as shown in Figure 4 , specifically comprises the following steps:

[0058] The first part: using time sliding window to construct time domain feature set input

[0059] The time domain signal collected by the sensor is preprocessed to construct the time domain feature set input. First, the time domain feature changes obviously with time, which is often used as a characteristic parameter. A time sliding window with a size of 1024 is used to slide along the time axis, and the sliding step is the window size. Each sliding generates a signal input X. Assuming that the vibration time domain signal is X = [ , ,……, ], wherein n is 1024. As shown in Figure 5 , the time domain signal is divided using the sliding window, and adjacent different colors represent independent characteristic input samples.

[0060] In order to facilitate the training of the network, each input signal X is subjected to "[-1, 1]" normalization processing, and the formula of the normalization processing is as follows:

[0061] (1)

[0062] The second part: using CNN to construct multi-scale feature extractor and data processing flow

[0063] Firstly, the multi-scale feature extraction module is composed of a CNN network, as shown in Figure 6As shown in the top section, convolutional layers (Conv), pooling layers (MaxPooling, AdaptiveMaxPooling), regularization layers (BN), and activation functions (ReLU) are used. The input data is the collected vibration time-domain signal. A time sliding window slides along the time series, and the data within the window is regularized before being input into the feature fusion unit. Among them, the convolutional layers are mainly used to process the raw data and are responsible for the initial feature extraction. The pooling layer can be regarded as a special type of convolutional layer, and the regularization layer is used to accelerate the training of network parameters.

[0064] One-dimensional forward propagation from the neuron input in convolutional layer l-1 to layer l can be expressed by the following formula:

[0065] (2)

[0066] in, For the first The bias scalar of each neuron; For the next level The output of the nth neuron; from the nth neuron in the previous layer The first neuron to the first neuron in layer l one neuron The kernel is used to determine the input of this layer. ; This is the ReLU activation function.

[0067] After the initial processing, the input sequence X undergoes a series of convolutional pooling operations. First, the first convolutional layer uses a large kernel (15 pixels wide) to quickly integrate the feature information of the input sequence, improving data processing efficiency and obtaining coarse-grained feature information after the first processing. Compared to small kernels, wide kernels better suppress high-frequency noise. After the first convolutional layer, its output... The vector 1024×1 is expanded to 1010×16.

[0068] Subsequently, three more convolutional layers with the same kernel size of 3 are added to deepen the network. These multiple convolutional layers can filter coarse-grained feature data, better suppressing high-frequency noise and overfitting. Their output... , , The values ​​are 504×32, 504×64, and 500×128, respectively. Furthermore, smaller convolutional kernels can be used for multi-layer nonlinear mapping, enabling correlation calculations on data from different time domains at a finer granular scale. Simultaneously, max pooling and adaptive pooling operations are employed in each layer to enhance the identification capability of fault features. Fault features extracted after deep convolutional operations are more robust, improving the detection and classification accuracy of the final neural network.

[0069] Part 3: Fault Diagnosis and Classification Using the Transformer Attention Mechanism

[0070] The Transformer structure is introduced into the classifier, and an attention mechanism is used to identify the deep feature sequence processed by the multi-scale feature extractor as input, which is then treated as a token sequence. . Figure 7 It demonstrates a standard Transformer architecture, whose core module is a multi-head attention mechanism module responsible for feature extraction.

[0071] Then, a trainable and randomly initialized token header is used. Add to the sequence head, i.e., compose =[ , ].

[0072] In order for the model to utilize the sequence order, additional information about the relevant or absolute position of the markers must be added; that is, positional encoding information needs to be added. Finally obtained = + .

[0073] Enter the multi-head attention mechanism module in the Transformer structure and perform self-attention calculations according to the following formula:

[0074] (3)

[0075] in , and This is the parameter matrix to be learned. softmax= Z is a vector. i Z j It is an element.

[0076] To ensure that vector information from different subspaces is shared, parallel attention computation, also known as multi-head attention, is further employed, calculated according to the following formula:

[0077] (4)

[0078] in for , It is a parameter matrix.

[0079] The output feature vector after passing through the Transformer structure to a fully connected layer to obtain a final four-dimensional classification result vector.

[0080] Embodiments

[0081] The experimental data comes from the current popular Case Western Reserve University rolling bearing fault data set, Figure 8 The experimental platform of this data set is shown. Four kinds of working condition bearing fault data under different frequencies are selected from it, i.e. all bearing fault data under four different rotating speeds. The fault data can be divided into four categories of normal, roller fault, inner ring fault and outer ring fault. Each category of bearing will face different fault types and fault intensities during operation. For detailed data set information, please see Table 1. The horizontal original vibration signals selected from these data are used to verify the classification accuracy of rolling bearing fault diagnosis under complex working conditions.

[0082] Table 1 Detailed information of the verification data set

[0083] ;

[0084] In order to facilitate understanding, the real data set will be analyzed and verified as follows:

[0085] Step 1: Collect time domain signal sample data of bearings under different working conditions. Divide the training set and test set according to the sample size of 1024:2 for each category of data set. The training set is randomly divided, so the training set and test set both contain corresponding sample data of normal, roller fault, inner ring fault and outer ring fault under different working conditions.

[0086] Step 2: Perform preprocessing operation on the divided data set to construct time domain feature set input sequence. Perform regularization operation of formula 1 on each sample to obtain algorithm model input = [ , , …, ], where n is 1024.

[0087] Step 3: Set the network training parameters, the process of network training is to update the parameters through the back propagation algorithm. Since the bearing fault diagnosis problem is a classification problem, the cross-entropy loss function is adopted as the loss function. The more accurate the prediction result of the network output is, the lower the loss error is: the training process of the network is actually to constantly pursue the minimization of the loss function, and the "Adam" optimization algorithm is selected in the application, which can adaptively adjust the learning rate and iteratively optimize the network parameters, and the iteration number is set to 100 and the learning rate is set to 0.001, the loss function adopts the cross-entropy function, and the activation function adopts the ReLU and Softmax functions. The training batch size batch is set to 64, that is, the training samples are randomly divided into small batches with a fixed length of 64, and in each epoch, a total of 64 sample data are sent into the network.

[0088] Step 4: Establish a deep learning network, including a multi-scale feature extraction module based on CNN, a Transformer attention mechanism learning module and a fully connected classification layer model.

[0089] Step 5: Input all the training sample sets into the established deep learning network for training, and use the test set to evaluate the performance of the model at each time during the training. The cross-entropy loss function is used to optimize the algorithm parameters in the way of back propagation, so as to continuously adjust all the parameters of the network, and finally obtain an offline model suitable for a certain sampling frequency, which is a bearing fault diagnosis classification model.

[0090] Step 6: Deploy the trained offline model to an online platform to diagnose the health conditions of various rotating machinery equipment under different working conditions, and can give the most likely fault type when a fault occurs.

[0091] Table 2 is the overall diagnosis performance of the application on the test data.

[0092] ;

[0093] Figures 9-11 The confusion matrix of the test results is given, which are the classification results of 12k drive end bearing fault data, 12k fan end bearing fault data and 48k drive end bearing fault data in turn, and the total number of samples to be tested is 1186, 1181 and 3226 respectively. From Table 2 and Figures 9-11It can be seen that the classification accuracy of the bearing fault classification method under multiple working conditions based on multi-scale feature extraction and attention mechanism can reach more than 90%, among which for 12k driving end bearing fault data, normal state, inner ring fault and roller fault can be effectively and accurately classified, when 12k fan end bearing fault data is classified, only one sample of roller fault is misclassified in 1181 test samples, and the rest are correctly predicted; in the classification performance of 48k driving end bearing fault data, the diagnosis accuracy is as high as 97.09%. The model proposed in the patent technology has strong reliability, and has high accuracy for fault diagnosis under multiple working conditions.

[0094] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it cannot be understood as limiting the scope of the patent of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, several modifications and improvements can be made, which belong to the protection scope of the present application.

Claims

1. A bearing fault diagnosis and classification method based on multi-scale features and attention, characterized in that, Includes the following steps: S1. Real-time acquisition of bearing time-domain signal data via accelerometer, preprocessing, and obtaining preprocessed data; The preprocessing of S1 includes the following steps: A time sliding window of size 1024 is used to slide along the time axis, with a sliding step size equal to the window size. Each slide generates a segment of time-domain signal input X, where X = [ , , ..., ], where n is 1024; For each segment of the input time-domain signal X, a [-1, 1] regularization process is performed. The formula for the regularization process is as follows: ； Obtain the temporal feature set input sequence as preprocessing data; S2. The preprocessed data is extracted and analyzed by a pre-constructed multi-scale feature classification module to obtain the first data. The pre-constructed multi-scale feature classification module includes a convolutional layer with a large convolutional kernel and three convolutional layers with the same small convolutional kernel. Specifically, S2 refers to: The preprocessed data is input into the multi-scale feature extraction module and passes through a convolutional layer with a kernel size of 15 and three consecutive convolutional layers with a kernel size of 3 in sequence. The output after the first convolutional layer The output is after the second convolutional layer. The output is after the third convolutional layer. The output is after the fourth convolutional layer. As the first data; One-dimensional forward propagation from the neuron input in convolutional layer l-1 to layer l is expressed by the following formula: ； in, For the first The bias scalar of each neuron; For the next level The output of the nth neuron; from the nth neuron in the previous layer The first neuron to the first neuron in layer l one neuron The kernel is used to determine the input of this layer. ; It is the ReLU activation function; S3. After inputting the first data into the Transformer attention mechanism learning module for learning, the second data is obtained; Specifically, S3 is: The first data is identified as an input sequence using an attention mechanism and treated as a token sequence. ; A trainable and randomly initialized token header Add to the first data header, i.e., compose =[ , Add location encoding information ,get = + ; Enter the multi-head attention mechanism module in the Transformer structure and perform self-attention calculation according to the following formula: ； in, , and Let softmax be the parameter matrix to be learned. Z is a vector. i Z j It is an element; To enable the sharing of vector information across different subspaces, parallel attention computation, also known as multi-head attention, is further employed, calculated according to the following formula: ； in for , For parameter matrices; The output feature vector after passing through the Transformer structure , as the second data; S4. Input the second data into a fully connected layer to obtain the final four-dimensional classification result vector, and output the diagnostic classification result of bearing fault.

2. The bearing fault diagnosis and classification method based on multi-scale features and attention as described in claim 1, characterized in that, Each convolutional layer contains a regularization layer and a ReLU activation function, and each convolutional layer performs max pooling and adaptive pooling operations.

Images