Rolling bearing fault diagnosis method based on double fusion attention mechanism

Abstract

The invention discloses a rolling bearing fault diagnosis method based on a double fusion attention mechanism, and the method comprises the steps: collecting a vibration signal, processing the vibration signal to form a two-dimensional time-frequency diagram, and dividing the formed two-dimensional time-frequency diagram into a data set; dual fusion attention based on channel attention and space attention is provided by improving a traditional mixed attention mechanism; constructing a dual fusion attention-multi-scale feature network fault diagnosis model based on dual fusion attention, and training the constructed dual fusion attention-multi-scale feature network fault diagnosis model through a data set; and finally, inputting a rolling bearing signal under a strong noise background into the trained double fusion attention-multi-scale feature network fault diagnosis model to obtain a fault diagnosis result. According to the invention, the problem of low identification accuracy in a complex background in the existing monitoring and diagnosis of the rolling bearing of the machinery is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical equipment status monitoring and fault diagnosis, and specifically relates to a rolling bearing fault diagnosis method based on a dual fusion attention mechanism. Background Art

[0002] In the field of artificial intelligence, attention mechanism has become an important part of neural network structure. This mechanism can learn from the operation of human visual system, so that the model can dynamically focus on the specific part of the input related to the current task, thereby improving the performance of the model. By introducing attention mechanism into traditional convolutional neural network, dynamic convolutional network can dynamically adjust the weight of each convolutional layer according to the input, aggregate multiple parallel convolutional layers in a nonlinear way, thereby enhancing the expression and extraction ability of features. In addition, attention long short-term memory network combines attention mechanism with long short-term memory network. By paying attention to the input data, the long short-term memory network can focus more on important information, thereby improving the performance of memory network and improving the accuracy of information update of forget gate, output gate and input gate structure. Attention mechanism empowers deep learning, which can help deep learning process complex data more effectively and capture key information, thereby improving the overall accuracy and reliability. Especially in the field of fault diagnosis, the application of attention mechanism provides a new breakthrough for the accurate identification and location of mechanical faults.

[0003] Rolling bearings are widely used in various types of mechanical equipment. They are usually in high-speed operation and in large numbers. The health status of these bearings directly determines the stable operation of the equipment, which is of vital importance to ensure the safety and stability of the equipment. Therefore, the research on rolling bearing fault diagnosis is particularly important, which can not only effectively prevent potential operating failures, but also ensure the efficient and reliable operation of the equipment. The deep learning fault classification method encounters some difficulties in practical applications: for example, the working environment is often mixed with a lot of interference elements such as noise, resulting in large fluctuations in the quality of the collected sensor data; at the same time, in the face of the diversity of working conditions and complex backgrounds, the existing models show insufficient adaptability and reduced recognition accuracy, which significantly restricts the performance of the deep learning model. Summary of the invention

[0004] The purpose of the present invention is to provide a rolling bearing fault diagnosis method based on a dual fusion attention mechanism, which solves the problem of low recognition accuracy in complex backgrounds in existing mechanical rolling bearing monitoring and diagnosis.

[0005] The technical solution adopted by the present invention is a rolling bearing fault diagnosis method based on a dual fusion attention mechanism, which specifically includes the following steps:

[0006] Step 1, collecting vibration signals, processing the vibration signals to form a two-dimensional time-frequency graph, and dividing the formed two-dimensional time-frequency graph into data sets;

[0007] Step 2: Improve the traditional hybrid attention mechanism and propose a dual fusion attention based on channel attention and spatial attention;

[0008] Step 3, constructing a dual fusion attention-multi-scale feature network fault diagnosis model based on dual fusion attention, and training the constructed dual fusion attention-multi-scale feature network fault diagnosis model based on the data set divided in step 1;

[0009] Step 4: Input the rolling bearing signal under the strong noise background into the dual fusion attention-multi-scale feature network fault diagnosis model trained in step 3 to obtain the fault diagnosis result.

[0010] The present invention is also characterized in that

[0011] The vibration signals in step 1 include original bearing vibration signals of outer ring fault, inner ring fault and normal state under strong noise.

[0012] In step 1, the vibration signal is processed to form a two-dimensional time-frequency diagram, which specifically includes the following steps:

[0013] Step 1.1, dividing the collected vibration signal into several segments;

[0014] Step 1.2, use continuous wavelet transform to perform time-frequency analysis on these segmented vibration signals, use Morlet as the wavelet basis, and transform the one-dimensional vibration signal into a two-dimensional time-frequency diagram through wavelet transform.

[0015] The data set in step 1 includes a training set, a validation set, and a test set.

[0016] In step 2, the traditional hybrid attention mechanism is improved by giving an original feature as the input F of the attention module. S , get the channel weight W1 about the original input feature, multiply the channel weight W1 by the original input feature element by element, so that the dimension of the channel weight W1 is the same as the original input feature F S Match it, and then compare it with the original input feature F S Element-by-element multiplication obtains the channel enhancement feature F1;

[0017] The channel enhanced feature F1 is input into the spatial attention module, and the original feature is also input into the spatial attention module to obtain the spatial weight Q1 of the channel enhanced feature F1; multiple spatial weights are fused by element-by-element multiplication to obtain the synthetic spatial weight. Finally, the synthetic spatial weight M t Multiply it element by element with the channel enhancement feature F1 to get the output feature Fe .

[0018] The dual fusion attention based on channel attention and spatial attention in step 2 is specifically to first combine the feature concatenation Concat and Add operations to fuse the features, and then fuse the Concat and Add fusion features for a second time. After convolution dimensionality reduction, the Sigmoid function is used to map the features to the range of 0 to 1 to obtain the weights; the Sigmoid function is specifically shown in the following formula:

[0019]

[0020] Where (x) is the input value, f(x) is the corresponding output value of the Sigmoid function, and (e) is the base of the natural logarithm, also known as the Euler number.

[0021] The dual fusion attention-multi-scale feature network fault diagnosis model includes a sequentially connected input layer, four feature extraction layers, a multi-view convolution module, a Concat layer, a convolution layer, a pooling layer, and a fully connected layer;

[0022] The four feature extraction layers each include a sequentially connected pooling layer, a dual fusion attention convolution layer, a batch normalization layer, and a convolution layer;

[0023] The four feature extraction layers, multi-view convolution module, Concat layer, convolution layer, and pooling layer use the ReLU function. The fully connected layer uses the Softmax function as the activation function. The Softmax function converts the classification results into probabilities to realize the fault diagnosis of rolling bearings. A Dropout layer is added before the Softmax layer, and the Dropout parameter is set to 0.2. Finally, the optimizer is specified through the compile function.

[0024] The multi-view convolution module includes four levels, which correspond to the four feature extraction layers respectively. The first level includes a convolution layer of size 1×1 and a maximum pooling layer, which is used to adjust the number of channels and strengthen the significant features through the maximum pooling layer, so that the model can capture key spatial features.

[0025] In the second level, based on the first level, a 3×3 convolution layer is added to perform dilated convolution with a dilation rate of 1.

[0026] The third level adds a 3×3 dilated convolutional layer based on the second level, with a dilation rate of 3, and removes the maximum pooling layer;

[0027] The fourth level adds a 3×3 atrous convolution layer with a dilation rate of 5 on the basis of the third level, and removes the 1×1 convolution layer.

[0028] The four levels of the multi-view convolution module are sequentially connected with a convolution layer and a pooling layer.

[0029] The receptive field of the third and fourth level atrous convolutional layers is as follows:

[0030] RF i+1 =RF i +(k′-1);

[0031] Where RF i+1 is the receptive field of the current level, RF i is the receptive field of the previous level, and k′ is the size of the convolutional layer after adding the dilation rate;

[0032] The size of the convolutional layer after the dilation rate is as follows:

[0033] k′=k+(k-1)×(d-1);

[0034] Where d is the hole rate size; k is the original convolutional layer size.

[0035] The beneficial effects of the present invention are:

[0036] The rolling bearing fault diagnosis method based on the dual fusion attention mechanism of the present invention focuses on rolling bearing fault diagnosis, and proposes and constructs a DFA-MSFN fault diagnosis network model based on the dual fusion attention mechanism. Among them, multi-scale feature fusion plays a key role. Multi-scale feature fusion can fully extract the feature information of rolling bearing fault signals at different scales, integrate the features of different scales, greatly enhance the richness and robustness of the features, and accurately identify the fault type even under strong noise interference, significantly improving the diagnosis accuracy.

[0037] The present invention introduces the original image into the subsequent attention mechanism, which can focus on key features more accurately, improve the efficiency and accuracy of feature extraction, and successfully solve the problem of serial and parallel connections. At the same time, this innovative design enhances the adaptability and generalization ability of the model, enabling it to better cope with different types of rolling bearing faults and noisy environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flowchart of a rolling bearing fault diagnosis method based on a dual fusion attention mechanism of the present invention;

[0039] Figure 2 It is a schematic diagram of the structure combining channel and spatial attention in the present invention;

[0040] Figure 3 It is a structural diagram of the dual fusion attention-multi-scale feature network fault diagnosis model in the present invention. DETAILED DESCRIPTION

[0041] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] The rolling bearing fault diagnosis method of the present invention is based on the dual fusion attention mechanism, such as Figure 1 As shown, the specific steps include:

[0043] Step 1, collecting vibration signals, and processing the vibration signals to form a two-dimensional time-frequency graph, and dividing the formed two-dimensional time-frequency graph into a data set, the data set including a training set, a validation set and a test set;

[0044] Step 2: Improve the traditional hybrid attention mechanism and propose a dual fusion attention (DFA) based on channel attention and spatial attention.

[0045] Step 3: construct a dual fusion attention-multi-scale feature network (DFA-MSFN) fault diagnosis model based on dual fusion attention, and train the constructed dual fusion attention-multi-scale feature network fault diagnosis model based on the data set divided in step 1;

[0046] Step 4: Input the rolling bearing signal under the strong noise background into the dual fusion attention-multi-scale feature network fault diagnosis model trained in step 3 to obtain the fault diagnosis result.

[0047] In the training of the DFA-MSFN fault diagnosis model, the training set is input into the model in batches. The batch size has a significant impact on the network training speed, convergence and generalization performance. In the range of 2 to 32 batches, as the batch size increases, the training speed increases, and the network model with 32 batches / batch converges the fastest. Compared with the validation set, it is found that the model in this range can also achieve good accuracy on the validation set. However, in the range of 64 to 128 batches, although the convergence speed becomes faster, the accuracy is affected and the generalization ability is reduced. Therefore, choosing a suitable batch size can balance the training speed and convergence accuracy, and a batch size of 32 batches / batch is proposed.

[0048] The present invention significantly improves the diagnostic accuracy through multi-scale feature fusion to solve the problem of poor rolling bearing fault diagnosis effect under strong noise background.

[0049] Example 1

[0050] The rolling bearing fault diagnosis method based on the dual fusion attention mechanism of the present invention specifically includes the following steps:

[0051] Step 1: Collect the original bearing vibration signals of outer ring fault, inner ring fault and normal state under strong noise, segment the collected vibration signals and convert them into time-frequency diagrams, use continuous wavelet transform to perform time-frequency analysis on these segmented vibration signals to obtain two-dimensional time-frequency diagrams, and divide them into training set, test set and validation set.

[0052] In step 1, the vibration signal is processed to form a two-dimensional time-frequency diagram, which specifically includes the following steps:

[0053] Step 1.1, divide the collected vibration signal into several small segments, the duration of each signal segment is set to 1 second, the corresponding specific number of sampling points will be determined according to the sampling frequency, and the overlapping length between adjacent samples will be adjusted as needed;

[0054] Step 1.2, use continuous wavelet transform to perform time-frequency analysis on these segmented vibration signals. Continuous wavelet transform can observe signal characteristics in both time domain and frequency domain, overcoming the global defects of Fourier transform.

[0055] Let ψ(x) be a function that is defined in a local region and tends to zero outside this region, and the mean of the function is zero, that is, it satisfies ψ(x) is called mother wavelet, which includes Morlet wavelet, Daubechies wavelet, Haar wavelet, etc. Morlet wavelet basis is selected as the wavelet basis of wavelet transform, and the one-dimensional vibration signals of bearings with seven fault types are transformed into two-dimensional time-frequency diagrams through continuous wavelet transform; the generated time-frequency diagram can take into account both time domain information and frequency domain information, increasing the richness of information. The specific function expression is as follows:

[0056]

[0057] Where w0 represents the center frequency, ψ p,q A series of wavelet basis functions ψ are obtained by scaling and time shifting the mother wavelet. p,q , the specific form is shown as follows:

[0058]

[0059] Where p is the scale, p>1 means the signal waveform shrinks, conversely, a<1 means the signal waveform stretches; and the scale p further affects the attenuation function, thereby changing the time domain support interval, so that the time can be located, and q is the displacement.

[0060] Note that the signal to be analyzed f(t) belongs to the space L 2 (R), then the continuous wavelet transform of the signal is:

[0061]

[0062] In the formula, It is right p,q Take its complex conjugate according to the definition of inner product in the complex frequency domain.

[0063] The transformation result W f (p,q) means that Morlet selects a basic center frequency, and then obtains a series of different center frequencies through scaling, and then obtains a series of basis functions in different intervals through time shift, and then multiplies them with a certain section of the original signal (corresponding to the interval of the basis function) and integrates them. The frequency corresponding to the extreme value generated is the frequency contained in this interval of the original signal, that is, the continuous wavelet transform can be considered as a bandpass filter, which only allows signals with frequencies close to the wavelet center frequency (after scaling) in the signal to be transformed to pass.

[0064] After the above, we get the relationship between time and frequency of the signal to be analyzed, draw it as an image to get the time-frequency diagram of continuous wavelet transform, and use it as the key object of data mining.

[0065] Step 2: Improve the traditional hybrid attention mechanism and propose a dual fusion attention (DFA) based on channel attention and spatial attention.

[0066] Step 2 proposes dual fusion attention by improving the traditional hybrid attention. Combining channel and spatial attention, the simple serial connection method is abandoned, the original image is involved in weighted calculation, and the key areas of the image are effectively focused. The specific improvements are as follows: Figure 2 As shown, it includes giving an original feature as the input F of the attention module S , get the channel weight W1 about the original input feature, multiply the channel weight W1 by the original input feature element by element, so that the dimension of the channel weight W1 is the same as the original input feature F S Match it, and then compare it with the original input feature F S Element-by-element multiplication obtains the channel enhancement feature F1;

[0067] The channel enhanced feature F1 is input into the spatial attention module, and the original feature is also input into the spatial attention module to obtain the spatial weight Q1 of the channel enhanced feature F1; multiple spatial weights are fused by element-by-element multiplication to obtain the synthetic spatial weight. Finally, the synthetic spatial weight M t Multiply it element by element with the channel enhancement feature F1 to get the output feature F e .

[0068] In step 2, the dual fusion attention based on channel attention and spatial attention is to first combine the feature concatenation Concat and Add operations to fuse the features. The Add operation increases the feature dimension information and improves the bearing fault classification effect. Then, the Concat and Add fusion features are fused again. After convolution dimensionality reduction, the Sigmoid function is used to map the features to the range of 0 to 1 to obtain the weights. Its function is that when the stimulus received by the neuron is greater than the threshold, it will be activated and the neuron will generate an output signal, otherwise the neuron is in an inhibited state. Sigmoid maps the value to the range of (0,1), corresponding to the normalized life. The function is continuous and smooth within the domain of definition. The Sigmoid function is shown in the following formula:

[0069]

[0070] Where (x) is the input value, f(x) is the corresponding Sigmoid function output value, and (e) is the base of the natural logarithm, also known as the Euler number, which is approximately equal to 2.71828.

[0071] Step 3: Since the mechanical equipment works under the background of strong noise, the details and texture features in the low layer of the feature extraction network are easily disturbed by noise. Therefore, a dual fusion attention-multi-scale feature network (DFA-MSFN) fault diagnosis model is constructed based on dual fusion attention, and the constructed dual fusion attention-multi-scale feature network fault diagnosis model is trained based on the data set divided in step 1.

[0072] Furthermore, if Figure 3 As shown in the figure, the dual fusion attention-multi-scale feature network fault diagnosis model includes a sequentially connected input layer, four feature extraction layers, a multi-view convolution module, a Concat layer, a convolution layer, a pooling layer and a fully connected layer; the multi-view convolution module adopts a hierarchical decreasing form.

[0073] Feature extraction layer 1, feature extraction layer 2, feature extraction layer 3 and feature extraction layer 4 all include sequentially connected pooling layers, dual fusion attention convolution layers, batch normalization layers and convolution layers; the multi-view convolution module includes four levels, and the four levels correspond to the four feature extraction layers respectively.

[0074] The operational expression of the convolutional layer is shown below;

[0075]

[0076] Where: d is the number of convolutional layers; M v is the number of feature maps in the d-1th layer; is the vth output feature map of the dth layer; is the vth convolutional layer associated with the i-th feature map in the d-th layer; is the bias of the vth convolutional layer in the dth layer; f(·) is the activation function.

[0077] The four feature extraction layers, multi-view convolution module, Concat layer, convolution layer, and pooling layer use the ReLU function. The ReLU function uses an input size of 224×224×3 and increases the number of feature maps through the convolution layer. The expression of the ReLU function is shown in the following formula:

[0078]

[0079] Furthermore, the pooling layer mainly includes the maximum pooling layer and the average pooling layer. During feature extraction, the average pooling layer focuses on retaining the overall features of the image and reducing the spatial dimension by calculating the average value of the local area.

[0080] Furthermore, the fully connected layer uses the Softmax function as the activation function. The Softmax function converts the classification results into probabilities to realize the fault diagnosis of rolling bearings. The entire DFA-MSFN fault diagnosis model optimizes parameters through forward propagation and back propagation to achieve efficient image classification. A Dropout layer is added before the Softmax layer to prevent the model from overfitting. The Dropout parameter is set to 0.2. Finally, the optimizer is specified through the compile function.

[0081] The optimizers include Adam, RMSprop and Nadam. The performance effects of Adam, RMSprop and Nadam on the DFA-MSFN fault diagnosis model are compared. The convergence speed is the fastest when the Nadam optimizer is used. In comparison with other optimizers, the Nadam optimizer not only performs well on the training set, but also has better performance on the validation set. Because of its adaptive learning rate feature, the learning rate can be automatically adjusted according to the gradient size, further improving the training efficiency of the DFA-MSFN fault diagnosis model.

[0082] That is, in the DFA-MSFN fault diagnosis model, the time-frequency graph is input into the feature extraction layer 1 to obtain feature 1, and then enters the first level of the multi-view convolution module to obtain D1; feature 1 also enters the feature extraction layer 2 to obtain the subsequent levels D2 to D4. At the same time, the outputs of each layer (such as D1, D2, D3, and D4) are concatenated by the Concat operation at their respective levels, and then concatenated with the outputs of other layers through the Concat layer to fuse features of different scales. After multiple convolutions, pooling, and feature fusion, the model finally passes the data to the fully connected layer to obtain the fault category.

[0083] Then, the constructed DFA-MSFN fault diagnosis model is trained based on the data set divided in step 1.

[0084] Step 4: Input the rolling bearing signal under the strong noise background into the dual fusion attention-multi-scale feature network fault diagnosis model trained in step 3 to obtain the fault diagnosis result.

[0085] Example 2

[0086] On the basis of Example 1, the improved traditional hybrid attention in step 2 of the rolling bearing fault diagnosis method based on the dual fusion attention mechanism of the present invention is specifically represented by the following steps:

[0087] Step 2.1: Given an original feature as the input of the attention module, and F S ∈R C×H×W First, F S Input into the channel attention module to obtain the channel weight W1 about the original input feature, and W1∈R C×1×1 ;

[0088] Step 2.2: Multiply the channel weight W1 by the original input feature element by element, and use the broadcast function in data processing to copy it along the second and third dimensions of W1 so that their dimensions are the same as F. S Match it, and then compare it with the original input feature F S The channel enhancement feature F1 is obtained by element-by-element multiplication. The formula is as follows:

[0089]

[0090] Step 2.3: Input the channel enhanced feature F1 into the spatial attention module, and input the original feature into the spatial attention module to obtain the spatial weight Q1 of the channel enhanced feature, and Q1∈R 1×H×W and the original input features F S The spatial weight Q2, and Q2∈R 1×H×W ;

[0091] Step 2.4: Fuse the spatial weight Q1 and the spatial weight Q2 by element-by-element multiplication to obtain the synthetic spatial weight. The formula is as follows:

[0092]

[0093] Finally, the synthetic spatial weight M t Multiply it element by element with the channel enhancement feature F1 to get the output feature F e , the formula is as follows:

[0094]

[0095] Example 3

[0096] On the basis of Example 1, the four levels of the multi-view convolution module in step 3 of the rolling bearing fault diagnosis method based on the dual fusion attention mechanism of the present invention are specifically:

[0097] The first level includes a 1×1 convolutional layer and a maximum pooling layer, which are used to adjust the number of channels to reduce parameters and calculations, and to strengthen significant features through the maximum pooling layer, so that the model can capture key spatial features;

[0098] The second level is based on the first level. The extracted features are convolved with a moderate convolution layer and dilation rate to obtain a wider field of view. A 3×3 convolution layer is added for dilation convolution with a dilation rate of 1.

[0099] In the third level, in high-level feature extraction, the receptive field is expanded by increasing the dilation rate to capture contextual information in a large range of the time-frequency graph, such as the association of frequency intervals or time periods. The third level adds a 3×3 dilated convolution layer with a dilation rate of 3 on the basis of the second level, and removes the maximum pooling layer;

[0100] In the fourth level, the dilation rate is increased to expand the coverage of the convolution layer to obtain global context information such as the frequency distribution of the time-frequency graph, and to improve the network's ability to extract the overall information of the time-frequency graph. On the basis of the third level, a 3×3 dilated convolution layer with a dilation rate of 5 is added, and the 1×1 convolution layer is removed.

[0101] The four levels of the multi-view convolution module are sequentially connected with convolution layers and pooling layers.

[0102] The receptive field of the third and fourth level atrous convolutional layers is as follows:

[0103] RF i+1 =RF i +(k′-1);

[0104] Where RF i+1 is the receptive field of the current level, RF i is the receptive field of the previous level, and k′ is the size of the convolutional layer after adding the dilation rate;

[0105] The size of the convolutional layer after the dilation rate is as follows:

[0106] k′=k+(k-1)×(d-1);

[0107] Where d is the hole rate size; k is the original convolutional layer size.

[0108] Specifically, the feature extraction layer 1 includes a convolution layer (Conv) with an input size of 224×224×3, a convolution kernel size of 3×3, a stride of 1, a padding number of 1, and an output channel number of 8. The convolution layer output is then batch normalized using BatchNorm2d, with an activation function of Relu, and a dual fusion attention module (DFA). Finally, a pooling layer with a size of 3×3 and a stride of 1 is used, and the output size is 222×222×8.

[0109] In the first stage of the multi-view convolution module, the input is globally max-pooled through the channel AvgPool, and the output size is 222×222×1. Then there is a convolution layer (Conv) with an input size of 222×222×1, a convolution kernel size of 1×1, a stride of 1, and an output channel number of 1. The outputs of the above two layers are concatenated through the Concat layer to obtain an output size of 222×222×2;

[0110] In the second stage of the multi-view convolution module, the channel AvgPool output size is 220×220×1. This is followed by a convolutional layer (Conv) with an input size of 220×220×1, a kernel size of 1×1, a stride of 1, and an output channel number of 1, and a dilated convolution layer (Dilated Conv) with a dilation rate of 1, a padding of 'Same', an output channel number of 16, a kernel size of 3×3, and a stride of 1. The outputs of the three layers are concatenated through the Concat layer to obtain an output size of 220×220×18;

[0111] The input size of feature extraction layer 3 is 220×220×16, the convolution kernel size is 3×3, the stride is 1, the padding number is 1, and the number of output channels is 32. After BatchNorm2d, Relu activation function, and DFA, the output size of the pooling layer with a size of 3×3 and a stride of 1 is 218×218×32;

[0112] In the third stage of the multi-view convolution module, the input size is 218×218×32. There is a convolution layer (Conv) with a kernel size of 1×1, a stride of 1, and an output channel of 1, followed by a dilated convolution layer (DilatedConv) with a dilation rate of 1, a padding number of 'Same', an output channel of 32, a kernel size of 3×3, and a stride of 1, and a dilated convolution layer (DilatedConv) with a dilation rate of 3, a padding number of 'Same', an output channel of 32, a kernel size of 3×3, and a stride of 1. The outputs of the three layers are concatenated by Concat to obtain an output size of 218×218×65;

[0113] The input size of feature extraction layer 4 is 218×218×32, the convolution kernel size is 3×3, the stride is 1, the padding number is 1, and the number of output channels is 64. After BatchNorm2d, Relu activation function, DFA, and pooling layer with size 3×3 and stride 1, the output size is 216×216×64;

[0114] In the fourth level of the multi-view convolution module, there are two dilated convolution layers. The first dilated convolution layer has a convolution kernel size of 3×3, a stride of 1, a dilation rate of 3, a padding number of 'Same', and an output channel number of 64. The second dilated convolution layer has a convolution kernel size of 3×3, a stride of 1, a dilation rate of 5, a padding number of 'Same', and an output channel number of 64. The outputs of the two layers are concatenated through the Concat layer to obtain an output size of 216×216×128.

[0115] Furthermore, the effects of multi-scale information fusion using the hierarchical decreasing multi-view convolution module and the non-hierarchical decreasing multi-view convolution are compared. When extracting mid-level features, the large hole rate convolution of the non-hierarchical decreasing multi-view convolution will lead to local information loss and large amount of calculation. The hierarchical decreasing multi-view convolution module can focus on local details and global context information. The improved fault diagnosis accuracy is 0.99, which is much higher than the non-hierarchical decreasing 0.952. Moreover, it also shows better stability and generalization ability in comparison with the validation set. At the same time, the amount of calculation is reduced by about 36% and the time consumed is reduced by 31%.

Claims

1. A rolling bearing fault diagnosis method based on a dual fusion attention mechanism, characterized in that: The specific steps include: Step 1, collecting vibration signals, processing the vibration signals to form a two-dimensional time-frequency graph, and dividing the formed two-dimensional time-frequency graph into data sets; Step 2: Improve the traditional hybrid attention mechanism and propose a dual fusion attention based on channel attention and spatial attention; Step 3, constructing a dual fusion attention-multi-scale feature network fault diagnosis model based on dual fusion attention, and training the constructed dual fusion attention-multi-scale feature network fault diagnosis model based on the data set divided in step 1; Step 4: Input the rolling bearing signal under the strong noise background into the dual fusion attention-multi-scale feature network fault diagnosis model trained in step 3 to obtain the fault diagnosis result.

2. The rolling bearing fault diagnosis method based on the dual fusion attention mechanism according to claim 1 is characterized in that: The vibration signal in step 1 includes original bearing vibration signals of outer ring fault, inner ring fault and normal state under strong noise.

3. The rolling bearing fault diagnosis method based on the dual fusion attention mechanism according to claim 1 is characterized in that: The processing of the vibration signal to form a two-dimensional time-frequency graph in step 1 specifically includes the following steps: Step 1.1, dividing the collected vibration signal into several segments; Step 1.2, use continuous wavelet transform to perform time-frequency analysis on these segmented vibration signals, use Morlet as the wavelet basis, and transform the one-dimensional vibration signal into a two-dimensional time-frequency diagram through wavelet transform.

4. The rolling bearing fault diagnosis method based on the dual fusion attention mechanism according to claim 1 is characterized in that: The data set in step 1 includes a training set, a validation set and a test set.

5. The rolling bearing fault diagnosis method based on dual fusion attention mechanism according to claim 1 is characterized in that: The improved traditional hybrid attention mechanism in step 2 is specifically given an original feature as the input F of the attention module S , get the channel weight W1 about the original input feature, multiply the channel weight W1 by the original input feature element by element, so that the dimension of the channel weight W1 is the same as the original input feature F S Match it, and then compare it with the original input feature F S Element-by-element multiplication obtains the channel enhancement feature F1; The channel enhanced feature F1 is input into the spatial attention module, and the original feature is also input into the spatial attention module to obtain the spatial weight Q1 of the channel enhanced feature F1; multiple spatial weights are fused by element-by-element multiplication to obtain the synthetic spatial weight. Finally, the synthetic spatial weight M t Multiply it element by element with the channel enhancement feature F1 to get the output feature F e .

6. The rolling bearing fault diagnosis method based on the dual fusion attention mechanism according to claim 5 is characterized in that: The dual fusion attention based on channel attention and spatial attention in step 2 is specifically to first combine the feature concatenation Concat and Add operations to fuse the features, then fuse the Concat and Add fusion features for a second time, and after convolution dimensionality reduction, use the Sigmoid function to map the features to the range of 0 to 1 to obtain the weights; the Sigmoid function is specifically shown in the following formula: Where (x) is the input value, f(x) is the corresponding output value of the Sigmoid function, and (e) is the base of the natural logarithm, also known as the Euler number.

7. The rolling bearing fault diagnosis method based on dual fusion attention mechanism according to claim 1 is characterized in that: The dual fusion attention-multi-scale feature network fault diagnosis model includes a sequentially connected input layer, four feature extraction layers, a multi-view convolution module, a Concat layer, a convolution layer, a pooling layer and a fully connected layer; The four feature extraction layers each include a sequentially connected pooling layer, a dual fusion attention convolution layer, a batch normalization layer, and a convolution layer; The four feature extraction layers, multi-view convolution module, Concat layer, convolution layer, and pooling layer use the ReLU function. The fully connected layer uses the Softmax function as the activation function. The Softmax function converts the classification results into probabilities to realize the fault diagnosis of rolling bearings. A Dropout layer is added before the Softmax layer, and the Dropout parameter is set to 0.

2. Finally, the optimizer is specified through the compile function.

8. The rolling bearing fault diagnosis method based on dual fusion attention mechanism according to claim 1 is characterized in that: The multi-view convolution module includes four levels, which correspond to the four feature extraction layers respectively. The first level includes a convolution layer and a maximum pooling layer of size 1×1, which are used to adjust the number of channels and strengthen the significant features through the maximum pooling layer, so that the model can capture key spatial features. In the second level, based on the first level, a 3×3 convolution layer is added to perform dilated convolution with a dilation rate of 1. The third level adds a 3×3 dilated convolutional layer based on the second level, with a dilation rate of 3, and removes the maximum pooling layer; The fourth level adds a 3×3 dilated convolutional layer on the basis of the third level, with a dilation rate of 5, and removes the 1×1 convolutional layer; The four levels of the multi-view convolution module are sequentially connected with a convolution layer and a pooling layer.

9. The rolling bearing fault diagnosis method based on dual fusion attention mechanism according to claim 8 is characterized in that: The receptive field of the third and fourth level hollow convolutional layers is as follows: RF i+1 =RF i +(k′-1); Where RF i+1 is the receptive field of the current level, RF i is the receptive field of the previous level, and k′ is the size of the convolutional layer after adding the dilation rate; The size of the convolutional layer after the dilation rate is as follows: k′=k+(k-1)×(d-1); Where d is the hole rate size; k is the original convolutional layer size.

Images