Rolling bearing fault diagnosis method based on MSDFM-ASTRSB under small sample

Abstract

The invention discloses a rolling bearing fault diagnosis method based on MSDFM-ASTRSB under a small sample, and the method comprises the steps: collecting one-dimensional vibration signals of a rolling bearing under different fault states through a plurality of sensors; a one-dimensional vibration signal of a single sensor in a single fault state is converted into a corresponding two-dimensional time-frequency diagram through continuous wavelet transform, and the two-dimensional time-frequency diagrams of a plurality of sensors are fused to obtain a multi-sensor data fusion image in the single fault state. Further obtaining multi-sensor data fusion images in all fault states and taking the images as a data set; constructing an MSDFM-ASTRSB rolling bearing fault diagnosis model based on an asymmetric soft threshold residual shrinkage block and a position interaction self-attention mechanism, and training the fault diagnosis model by using a training set in the data set; and using the trained fault diagnosis model to complete fault diagnosis of the rolling bearing. According to the invention, the defects of training sample shortage and single data are overcome, and the accuracy of rolling bearing fault diagnosis is improved.

Description

Technical Field

[0001] The invention relates to the technical field of rolling bearing fault diagnosis, and more specifically to a rolling bearing fault diagnosis method based on MSDFM-ASTRSB under a small sample. Background Art

[0002] With the advent of the information age, rotating machinery and equipment are widely used in important fields such as transportation, aerospace, etc. As an important part of rotating machinery, the operating status of rolling bearings directly affects the entire rotating machinery. Once a failure occurs, it may cause significant casualties and property losses. Therefore, it is very important to study rapid and efficient rolling bearing fault diagnosis methods.

[0003] Traditional fault diagnosis methods rely heavily on artificial prior knowledge, which has great uncertainty and leads to poor generalization ability. However, deep learning methods overcome the limitations of artificial feature selection in traditional fault diagnosis methods by adaptively selecting features, and have good application prospects in the field of fault diagnosis.

[0004] As a typical representative of deep learning algorithms, convolutional neural networks can adaptively extract feature information and have achieved great success in the field of fault diagnosis. Zhao et al. solved the problem of complex models of some rolling bearing fault diagnosis methods by constructing deep separable convolution blocks and non-local feature perception modules in "Rolling bearing fault diagnosis model based on DSCB-NFAM". Guo et al. constructed a one-dimensional wide convolutional neural network and introduced a convolutional block attention module in "Rolling Bearing Fault Diagnosis MethodBased on Attention CNN and BiLSTM Network", weighted different feature dimensions, and then used BiLSTM to further extract features, and finally used the Softmax classifier for fault diagnosis. However, although the rolling bearing fault diagnosis method based on CNN has a good diagnostic effect, it is easy to experience gradient vanishing phenomenon as the network structure deepens. The residual neural network uses identity mapping to enable the input information to bypass one or more layers and pass directly to deeper layers. During back propagation, it can well alleviate the problem of gradient vanishing that deep convolutional networks are prone to, significantly improve the training effect of deep networks, and has been widely used in the field of rolling bearing fault diagnosis. Chang et al. optimized the residual structure of the residual neural network in "Intelligent Fault Diagnosis of RollingBearings Using Efficient and Lightweight ResNet Networks Based on anAttention Mechanism", introduced a convolutional attention module (CBAM) to improve the feature extraction efficiency of the network, and used the ELU with better performance to replace the ReLU activation function to improve the network's expression ability. Finally, the CBAM-ResNet18 network was used for fault diagnosis. Li et al. first compressed and reconstructed the vibration signal in "Joint SemitensorCompressive Sensing With Multiscale Pyramidal Residual Network for RollingBearing Fault Diagnosis", and then combined the residual module and the multi-scale network module to establish a multi-scale pyramid residual network for fault diagnosis. Finally, two experimental data were used to verify the effectiveness of the proposed method. The above methods use identity mapping to construct a deeper network structure. Although they achieve better fault detection results while avoiding gradient disappearance, the deeper network structure leads to a surge in the number of parameters and higher hardware requirements.

[0005] The residual shrinkage network reduces unnecessary information and parameters by introducing a shrinkage mechanism, while maintaining the network depth and expression ability, reducing the complexity of the model and the requirements for hardware. In "Afault diagnosis method of rolling bearing based on improved deep residual shrinkage networks", Tong et al. proposed a pseudo soft threshold block and an adaptive slope block to eliminate the signal distortion caused by the soft threshold function, and used them to construct an improved residual shrinkage unit, adaptively set the optimal threshold and slope, and verified the robustness of the proposed method in a strong noise environment. In "An intelligent diagnosis method of rolling bearing based on multi-scale residual shrinkage convolutional neural network", Zhao et al. constructed a multi-scale residual shrinkage layer based on multi-scale learning and residual shrinkage blocks, and designed a separable convolution block using residual connections and separable convolutions. Finally, a rolling bearing fault diagnosis model was constructed, and the noise resistance and generalization ability of the proposed method were verified by two bearing data sets.

[0006] Although the above methods have strong noise resistance and excellent robustness, they have the following problems: (1) The above methods all use sufficient data to train the proposed network model, without considering the problem of fewer fault samples in real industrial engineering. (2) The above methods all use single sensor data for fault diagnosis, without considering the problem that single sensor data cannot fully represent the vibration characteristics of rolling bearings. (3) The above methods do not consider the distribution of positive and negative features in fault information, and use a global threshold function, which limits the learning ability of the model. Moreover, all information between thresholds is compressed to 0, which may cause the loss of some information. Summary of the invention

[0007] In view of this, the present invention provides a rolling bearing fault diagnosis method based on MSDFM-ASTRSB under a small sample, which at least solves some of the technical problems in the background technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] The present invention discloses a rolling bearing fault diagnosis method based on MSDFM-ASTRSB under a small sample, comprising the following steps:

[0010] S1: Use multiple sensors to collect one-dimensional vibration signals of rolling bearings under various fault conditions;

[0011] S2: The one-dimensional vibration signal collected by a single sensor under a single fault state is converted into a corresponding two-dimensional time-frequency diagram through continuous wavelet transform;

[0012] S3: Fusing the two-dimensional time-frequency images of multiple sensors under the same fault state to obtain a multi-sensor data fusion image under a single fault state;

[0013] S4: repeat step S3, and use the multi-sensor data fusion images in all fault states as a data set, and divide the data set into a training set, a verification set, and a test set;

[0014] S5: Based on the asymmetric soft threshold residual shrinkage block and the position interaction self-attention mechanism, a MSDFM-ASTRSB rolling bearing fault diagnosis model is constructed, and the MSDFM-ASTRSB rolling bearing fault diagnosis model is trained with the training set obtained in step S4;

[0015] S6: Use the trained MSDFM-ASTRSB rolling bearing fault diagnosis model to complete the rolling bearing fault diagnosis.

[0016] Furthermore, step S2 specifically includes the following steps:

[0017] S21, normalizing a one-dimensional vibration signal collected by a single sensor under a single fault state;

[0018] S22, dividing the normalized one-dimensional vibration signal by using an overlapping sampling method to obtain a one-dimensional sample data set of a single sensor under a single fault state;

[0019] S23, performing continuous wavelet transform on the one-dimensional sample data set obtained in step S22 to obtain wavelet coefficients of the input one-dimensional sample data, and mapping the wavelet coefficients into corresponding two-dimensional time-frequency diagrams.

[0020] Furthermore, step S21 performs normalization processing on the one-dimensional vibration signal collected by a single sensor under a single fault state, specifically including using the following normalization formula:

[0021]

[0022] in, is the i-th data in the normalized one-dimensional vibration signal, X i is the i-th original data in the one-dimensional vibration signal before normalization, X min is the minimum value of the one-dimensional vibration signal before normalization, X max is the maximum value of the one-dimensional vibration signal before normalization.

[0023] Furthermore, in step S22, the normalized one-dimensional vibration signal is divided using an overlapping sampling method, which specifically includes the following steps:

[0024] The normalized one-dimensional vibration signal is overlapped and sampled using a sliding window with a size of 1028 and a step size of 64 to obtain a one-dimensional sample data set of a single sensor under a single fault state.

[0025] Furthermore, step S23 specifically includes the following steps:

[0026] S231, select cmor3-3 wavelet basis function in the composite Morlet wavelet function, analyze the time-frequency characteristics of the one-dimensional sample data set, the expression of the cmor3-3 wavelet basis function is:

[0027]

[0028] Among them, σ represents the bandwidth parameter of the wavelet, π represents pi, t represents the sampling time, i represents the imaginary unit, f represents the center frequency of the wavelet, and exp(·) represents the exponential function.

[0029] S232, perform continuous wavelet transform on the one-dimensional sample data set, obtain the wavelet coefficients of the input one-dimensional sample data, and map the wavelet coefficients of the input one-dimensional sample data into a two-dimensional time-frequency graph; wherein the specific expression of continuous wavelet transform is:

[0030]

[0031] Among them, CWT(a,b) represents the wavelet transform of the input one-dimensional sample data x(t), x(t) represents the input one-dimensional sample data, a represents the scale parameter, b represents the translation parameter, ψ * represents the conjugate of the wavelet function.

[0032] Furthermore, the two-dimensional time-frequency graphs of multiple sensors under the same fault state are fused, which specifically includes the following steps:

[0033] Obtain the pixel matrix corresponding to the two-dimensional time-frequency diagram of all single sensors under the same fault state;

[0034] The pixel matrices corresponding to the two-dimensional time-frequency graphs of multiple single sensors under the same fault state are fused in the height direction to obtain a pixel fusion matrix;

[0035] The pixel fusion matrix is ​​mapped into a multi-sensor fusion two-dimensional time-frequency diagram under a single fault state.

[0036] Further, the MSDFM-ASTRSB rolling bearing fault diagnosis model constructed in step S4 includes a feature extraction module, an asymmetric soft threshold residual shrinkage block, a position interaction self-attention mechanism module and a Softmax classifier;

[0037] The feature extraction module is used to extract features from the multi-sensor data fusion image in the input training set, and obtain shallow feature information corresponding to each multi-sensor data fusion image;

[0038] The asymmetric soft threshold residual shrinkage block is used to further extract the shallow feature information of each multi-sensor data fusion image and output the deep feature information;

[0039] The position interaction self-attention mechanism module is used to further process the deep feature information to obtain the output information of the position interaction self-attention mechanism module;

[0040] The Softmax classifier is used to map the output information of the position interaction self-attention mechanism module to obtain the fault diagnosis result of the rolling bearing.

[0041] Furthermore, the asymmetric soft threshold residual shrinkage block is used to further extract the shallow feature information of each multi-sensor data fusion image and output the deep feature information, which specifically includes the following steps:

[0042] Divide shallow feature information into positive feature information and negative feature information;

[0043] For positive feature information, a global average pooling operation is performed to obtain a positive soft vector; the obtained positive soft vector is input into two fully connected layers for feature combination and conversion; the feature combination and conversion results are input into the Sigmoid function to obtain a positive dynamic adjustable factor; the obtained positive soft vector is multiplied by the positive dynamic adjustable factor to obtain a positive threshold;

[0044] For negative feature information, a global average pooling operation is performed to obtain a negative soft vector; the obtained negative soft vector is input into two fully connected layers for feature combination and conversion; the feature combination and conversion results are input into the Sigmoid function to obtain a negative dynamic adjustable factor; the obtained negative soft vector is multiplied by the negative dynamic adjustable factor to obtain a negative threshold;

[0045] The shallow feature information is globally averaged and then passed through two fully connected layers and a Sigmoid layer to obtain an adjustable factor between the positive and negative thresholds.

[0046] An asymmetric soft threshold function is constructed according to the positive threshold, the negative threshold and the adjustable factor between the positive and negative thresholds, and the characteristic information output by the asymmetric soft threshold function is obtained. The mathematical expression of the asymmetric soft threshold function is:

[0047]

[0048] In the formula, y represents the feature information output by the asymmetric soft threshold function; ω1 represents the positive threshold; ω2 represents the negative threshold; α represents the adjustable factor between the positive and negative thresholds; x represents the shallow feature information of the input;

[0049] The feature information output by the asymmetric soft threshold function is added to the original information of the corresponding multi-sensor data fusion image to obtain the deep feature information of the multi-sensor data fusion image.

[0050] Furthermore, the position interaction self-attention mechanism module further processes the deep feature information to obtain output information of the position interaction self-attention mechanism module, which specifically includes the following steps:

[0051] S51, performing global average pooling operations in H dimensions and W dimensions on the deep feature information output by the asymmetric soft threshold residual shrinkage block, respectively, to obtain a horizontal vector representation corresponding to the H dimension and a vertical vector representation corresponding to the W dimension;

[0052] S52, performing two-layer convolution operations on both the horizontal vector representation and the vertical vector representation, so as to sequentially scale and restore the number of channels of the vectors;

[0053] S53, respectively passing the horizontal vector representation and the vertical vector representation after the channel number is restored through the Sigmoid function to obtain a horizontal position vector corresponding to the horizontal vector representation and a vertical position vector corresponding to the vertical vector representation;

[0054] S54, multiplying the lateral position vector by the deep feature information to obtain lateral position information of the deep feature information;

[0055] Multiplying the longitudinal position vector by the deep feature information to obtain the longitudinal position information of the deep feature information;

[0056] S55, performing a linear transformation on the lateral position information obtained in step S54 to obtain a query matrix, a key matrix, and a value matrix corresponding to the lateral position information;

[0057] Performing a linear transformation on the longitudinal position information obtained in step S54 to obtain a query matrix, a key matrix and a value matrix corresponding to the longitudinal position information;

[0058] S56, performing a point multiplication operation on the query matrix of the longitudinal position information and the transposed matrix of the key matrix of the lateral position information to obtain first correlation information, and performing a Softmax function on the first correlation information to obtain a first correlation matrix; then multiplying the value matrix of the lateral position information by the first correlation matrix to obtain a self-attention score of the lateral position information;

[0059] Perform a point multiplication operation on the query matrix of the horizontal position information and the transposed matrix of the key matrix of the vertical position information to obtain the second correlation information, and obtain the second correlation matrix by the Softmax function of the second correlation information; then multiply the value matrix of the vertical position information by the second correlation matrix to obtain the self-attention score of the vertical position information;

[0060] S57. Add the self-attention score of the horizontal position information and the self-attention score of the vertical position information, and perform a residual connection with the input deep feature information to obtain the output information of the position interaction self-attention mechanism.

[0061] The present invention also discloses a rolling bearing fault diagnosis device based on MSDFM-ASTRSB under a small sample, including a computer system. When executed, the computer system can implement any one of the rolling bearing fault diagnosis methods based on MSDFM-ASTRSB under a small sample described in the present invention.

[0062] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a rolling bearing fault diagnosis method and device based on MSDFM-ASTRSB under a small sample, which has the following beneficial effects:

[0063] The present invention uses multi-sensor data to train a rolling bearing fault diagnosis model, thereby solving the technical defects of sample shortage and single data in the rolling bearing fault diagnosis model training process.

[0064] The rolling bearing fault diagnosis model based on MSDFM-ASTRSB established in the present invention uses an asymmetric soft threshold module and a position interactive self-attention module, which reduces feature redundancy while avoiding the loss of some useful information, thereby improving the accuracy of rolling bearing fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0066] Figure 1 A schematic flow chart of the rolling bearing fault diagnosis method based on MSDFM-ASTRSB provided by the present invention.

[0067] Figure 2This is a structural diagram of a feature extraction module and an asymmetric soft threshold residual shrinkage block in the MSDFM-ASTRSB rolling bearing fault diagnosis model provided by the present invention.

[0068] Figure 3 This is a structural diagram of the position interactive self-attention mechanism module in the MSDFM-ASTRSB rolling bearing fault diagnosis model provided by the present invention.

[0069] Figure 4 The visualization diagrams of the classification layers of different methods provided by the present invention when the number of samples is 10; wherein (a) is a visualization diagram of the classification layer of the SC-MSCNN method; (b) is a visualization diagram of the classification layer of the MRANet method; (c) is a visualization diagram of the classification layer of the MTF-ResNet method; (d) is a visualization diagram of the classification layer of the MSCNN method; (e) is a visualization diagram of the classification layer of the MDCNN method; and (f) is a visualization diagram of the classification layer of the MSDFM-ASTRSB method.

[0070] Figure 5 The confusion matrix diagrams of different methods provided by the present invention when the number of samples is 100; among them, (a) is the confusion matrix diagram of the SC-MSCNN method; (b) is the confusion matrix diagram of the MRANet method; (c) is MTF-ResNet; (d) is the confusion matrix diagram of the MSCNN method; (e) is the confusion matrix diagram of the MDCNN method; (f) is the confusion matrix diagram of the MSDFM-ASTRSB method. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] refer to Figure 1 As shown, the embodiment of the present invention discloses a rolling bearing fault diagnosis method based on feature fusion under a small sample, comprising the following steps:

[0073] S1: Use multiple sensors to collect one-dimensional vibration signals of rolling bearings under various fault conditions.

[0074] S2: The one-dimensional vibration signal under a single fault state collected by a single sensor is converted into a corresponding two-dimensional time-frequency diagram through continuous wavelet transform.

[0075] S3: Fuse the two-dimensional time-frequency images of multiple sensors under the same fault state to obtain a multi-sensor data fusion image under a single fault state.

[0076] S4: Repeat step S3, and use the multi-sensor data fusion images under all fault states as a data set, and divide the data set into a training set, a verification set, and a test set.

[0077] S5: Based on the asymmetric soft threshold residual shrinkage block and position interactive self-attention mechanism, a MSDFM-ASTRSB rolling bearing fault diagnosis model is constructed, and the MSDFM-ASTRSB rolling bearing fault diagnosis model is trained with the training set obtained in step S4.

[0078] S6: Use the trained MSDFM-ASTRSB rolling bearing fault diagnosis model to complete the rolling bearing fault diagnosis.

[0079] The present invention is further described below.

[0080] In step S1, multiple sensors are used to collect one-dimensional vibration signals of the rolling bearing under various fault states. The multiple sensors can be different types of sensors such as a drive end acceleration sensor, a fan end acceleration sensor and a base acceleration sensor. Each sensor collects one-dimensional vibration signals of the rolling bearing under various fault states. In this step, the various fault states refer to the inner ring, outer ring, rolling element and mixed faults as well as the normal state.

[0081] In step S2, the one-dimensional vibration signal collected by a single sensor under a single fault state is converted into a corresponding two-dimensional time-frequency diagram through continuous wavelet transform. Specifically, this step includes the following steps:

[0082] S21. First, a one-dimensional vibration signal collected by a single sensor under a single fault state is normalized. The normalization operation may specifically include using the following normalization formula:

[0083]

[0084] in, is the i-th data in the normalized one-dimensional vibration signal, X i is the i-th original data in the one-dimensional vibration signal before normalization, X min is the minimum value of the one-dimensional vibration signal before normalization, X max is the maximum value of the one-dimensional vibration signal before normalization.

[0085] S22. Then, the normalized one-dimensional vibration signal is divided using an overlapping sampling method to obtain a one-dimensional sample data set of a single sensor under a single fault state; specifically, in this step, a sliding window with a size of 1028 and a step size of 64 can be used to perform overlapping sampling on the normalized one-dimensional vibration signal to obtain one-dimensional sample data of a single sensor under a single fault state.

[0086] Since the vibration signals of the rolling bearing under different fault states collected by the sensor in the present invention are a series of one-dimensional data, when overlapping sampling, the sliding window size 1028 means that each window contains 1028 data points as features, and the step size 64 represents the distance the sliding window moves each time. Using the overlapping sampling method for division can obtain more detailed feature information and reduce information loss caused by window boundaries.

[0087] S23, performing continuous wavelet transform on the one-dimensional sample data set obtained in step S22, obtaining wavelet coefficients of the input one-dimensional sample data, and mapping the wavelet coefficients into corresponding two-dimensional time-frequency diagrams. In this step, continuous wavelet transform is performed on the one-dimensional sample data set after overlapping sampling, and the wavelet coefficients are mapped into corresponding two-dimensional time-frequency diagrams, thereby effectively retaining the time-frequency characteristics of the original signal. Specifically, this process may include the following steps:

[0088] S231, select cmor3-3 wavelet basis function in the composite Morlet wavelet function, analyze the time-frequency characteristics of the one-dimensional sample data set, the expression of the cmor3-3 wavelet basis function is:

[0089]

[0090] Among them, σ represents the bandwidth parameter of the wavelet, π represents pi, t represents the sampling time, i represents the imaginary unit, f represents the center frequency of the wavelet, and exp(·) represents the exponential function.

[0091] S232, perform continuous wavelet transform on the one-dimensional sample data set, obtain the wavelet coefficients of the input one-dimensional sample data, and map the wavelet coefficients of the input one-dimensional sample data into a two-dimensional time-frequency graph; wherein the specific expression of continuous wavelet transform is:

[0092]

[0093] Among them, CWT(a,b) represents the wavelet transform of the input one-dimensional sample data x(t), x(t) represents the input one-dimensional sample data, a represents the scale parameter, b represents the translation parameter, ψ * represents the conjugate of the wavelet function.

[0094] Step S3, after obtaining the two-dimensional time-frequency graphs under a single fault state collected by all single sensors, the two-dimensional time-frequency graphs of multiple sensors under the same fault state are fused to obtain a multi-sensor data fusion image under a single fault state.

[0095] Specifically, in step S3, firstly, the pixel matrix corresponding to the two-dimensional time-frequency diagram of all single sensors under the same fault state is obtained; then, the pixel matrices corresponding to the two-dimensional time-frequency diagrams of multiple single sensors under the same fault state are fused in the height direction to obtain a pixel fusion matrix; finally, the pixel fusion matrix is ​​mapped into a multi-sensor fusion two-dimensional time-frequency diagram under a single fault state.

[0096] Step S4 is used to obtain multi-sensor data fusion images under all fault states and use them as a data set, and divide the data set into a training set, a verification set and a test set.

[0097] Step S5 is used to construct an MSDFM-ASTRSB rolling bearing fault diagnosis model based on an asymmetric soft threshold residual shrinkage block and a position interactive self-attention mechanism, and use the obtained training set to train the constructed MSDFM-ASTRSB rolling bearing fault diagnosis model.

[0098] The MSDFM-ASTRSB rolling bearing fault diagnosis model constructed by the present invention comprises a feature extraction module, an asymmetric soft threshold residual shrinkage block, a position interactive self-attention mechanism module and a Softmax classifier; wherein the feature extraction module is used to perform feature extraction on a multi-sensor data fusion image in an input training set, and obtain shallow feature information corresponding to each multi-sensor data fusion image; the asymmetric soft threshold residual shrinkage block is used to perform further feature extraction on the shallow feature information of each multi-sensor data fusion image, and output deep feature information; the position interactive self-attention mechanism module is used to further process the deep feature information, and obtain the output information of the position interactive self-attention mechanism module; the Softmax classifier is used to map the output information of the position interactive self-attention mechanism module, and obtain the fault diagnosis result of the rolling bearing.

[0099] like Figure 2 As shown in the figure, for each multi-sensor data fusion image in the training set, its size is expressed as C×H×W=3×128×128; where C represents the number of channels of the input feature map, H represents the length of the input feature map, and W represents the width of the input feature map. The feature extraction module in the MSDFM-ASTRSB rolling bearing fault diagnosis model uses two convolutional layers (the convolutional layer is Figure 2 Conv2D structure on the left) and a pooling layer ( Figure 2The Pooling structure in the above example is used to extract shallow features of the input information. In the embodiment of the present invention, the convolution kernel sizes of the two convolutional layers are both 3×3, the step sizes are both 1, the number of channels is both 16, and the shallow feature information obtained is represented by x, and its size is C1×H1×W1=16×64×64.

[0100] The asymmetric soft threshold residual shrinkage block is used to further extract the shallow feature information of each multi-sensor data fusion image and output the deep feature information. Figure 2 As shown, the asymmetric soft threshold residual shrinkage block mainly includes an asymmetric soft threshold block, an asymmetric soft threshold slope block and an asymmetric soft threshold function.

[0101] In the asymmetric soft threshold block, the shallow feature information x is first divided into positive feature information x + and negative feature information x - In this embodiment, the features greater than or equal to 0 in the shallow features are positive feature information x + , features less than 0 are negative feature information x - ; Then align the feature information x + Processing is performed to obtain the positive threshold, and the negative feature information is x - Processing is performed to obtain a negative threshold.

[0102] Specifically, for positive feature information x + , firstly perform a global average pooling operation to obtain a forward soft vector k1; then input the obtained forward soft vector k1 into two fully connected layers for feature combination and conversion; then input the feature combination and conversion results into the Sigmoid function to obtain the positive dynamic adjustable factor h1; finally, multiply the obtained forward soft vector k1 by the positive dynamic adjustable factor h1 to obtain the positive threshold ω1; the mathematical expression of this process is:

[0103] k1=GPA(x + )

[0104] h1=F sig (F fc (F fc (k1)))

[0105] ω1=k1·h1

[0106] Among them, GPA(·) represents the global average pooling operation, F fc (·) represents the fully connected layer operation function, F sig (·) represents the Sigmoid function.

[0107] Similarly, for negative feature information x -, perform global average pooling operation to obtain negative soft vector k2; input the obtained negative soft vector k2 into two fully connected layers for feature combination and conversion; input the feature combination and conversion results into the Sigmoid function to obtain the negative dynamic adjustable factor h2; multiply the obtained negative soft vector k2 by the negative dynamic adjustable factor h2 to obtain the negative threshold ω2; the mathematical expression of this process is:

[0108] k2=GPA(x - )

[0109] h2=F sig (F fc (F fc (k2)))

[0110] ω2=k2·h2

[0111] Among them, k2 represents the negative soft vector, h2 represents the negative dynamic adjustable factor, and ω2 represents the negative threshold.

[0112] The asymmetric soft threshold slope block is used to obtain the adjustable factor between the positive and negative thresholds, such as Figure 2 As shown in the figure, in the asymmetric soft threshold slope block, the shallow feature information x is first subjected to a global average pooling operation, and then passes through two fully connected layers and a Sigmoid layer to obtain an adjustable factor α (α is usually a very small positive number) between the positive and negative thresholds. The mathematical expression of this process is:

[0113] α=F sig (F fc (F fc (GPA(x))))

[0114] In this process, F fc (·) represents the fully connected layer operation function, F sig (·) represents the Sigmoid function

[0115] After obtaining the positive threshold ω1, the negative threshold ω2 and the adjustable factor α between the positive and negative thresholds, an asymmetric soft threshold function is constructed, and the characteristic information output by the asymmetric soft threshold function is obtained. The mathematical expression of the asymmetric soft threshold function is:

[0116]

[0117] In the formula, y represents the feature information output by the asymmetric soft threshold function; ω1 represents the positive threshold; ω2 represents the negative threshold; α represents the adjustable factor between the positive and negative thresholds; and x represents the shallow feature information of the input.

[0118] The asymmetric soft threshold function constructed in the embodiment of the present invention not only has good continuity, but also uses different scaling factors in different threshold ranges, thereby reducing feature redundancy while avoiding the loss of some useful information.

[0119] The feature information y output by the asymmetric soft threshold function is added to the original information of the corresponding multi-sensor data fusion image to obtain the deep feature information of the multi-sensor data fusion image, which is the output of the asymmetric soft threshold residual shrinkage block. In this embodiment, the deep feature information is represented by x1, and its size is represented by C2×H2×W2. In this step, the feature information y output by the symmetric soft threshold function and the original information of the multi-sensor data fusion image are both represented in matrix form.

[0120] The position interactive self-attention mechanism module in the MSDFM-ASTRSB rolling bearing fault diagnosis model is used to further process the deep feature information x1 to obtain the output information of the position interactive self-attention mechanism module. This process mainly includes three steps: obtaining position information, calculating the interactive self-attention score, and obtaining the output of the position interactive self-attention mechanism module.

[0121] Combine the following Figure 3 The data processing process of the position interaction self-attention mechanism proposed in the present invention is described in detail. First, the position information is obtained, including steps S51-S54. The specific steps are as follows:

[0122] S51. First, perform global average pooling operations in H and W dimensions on the deep feature information x1 output by the asymmetric soft threshold residual shrinkage block to obtain a horizontal vector representation x2 corresponding to the H dimension and a vertical vector representation x3 corresponding to the W dimension; their sizes are C2×1×W2 and C2×H2×1, respectively.

[0123] S52. Then, two layers of convolution operations are performed on the horizontal vector representation x2 and the vertical vector representation x3 to scale and restore the number of channels of the vectors in sequence. Specifically, after the first convolution layer, the size of x2 is (C2 / 4)×1×W2, and the size of x3 is (C2 / 4)×H2×1. After the second convolution layer, the size of x2 is restored to C2×1×W2, and the size of x3 is restored to C2×H2×1. The information exchange between different channels is enhanced through scaling and recovery capabilities.

[0124] S53, respectively pass the horizontal vector representation and the vertical vector representation after the channel number is restored through the Sigmoid function to obtain the horizontal position vector x4 corresponding to the horizontal vector representation x2 and the vertical position vector x5 corresponding to the vertical vector representation x3;

[0125] S54, multiplying the horizontal position vector x4 by the deep feature information x1 to obtain the horizontal position information l of the deep feature information. The mathematical formula of this process is: l = F sig (Conv(Conv(F H-pooling (x1))))·x1.

[0126] Similarly, the longitudinal position vector x5 is multiplied by the deep feature information x1 to obtain the longitudinal position information z of the deep feature information, which is mathematically expressed as z=F sig (Conv(Conv(F W-pooling (x1))))·x1.

[0127] In mathematical expressions, F H-pooling (·) represents the global average pooling in H dimension, F W-pooling (·) represents the global average pooling in W dimension, Conv(·) represents the convolution operation, and F sig (·) represents the Sigmoid function, and · represents the multiplication operation.

[0128] After the horizontal position information l and the vertical position information z are obtained through steps S51-S54, the interactive self-attention score is calculated according to the obtained horizontal position information l and the vertical position information z. The specific steps are as follows:

[0129] S55: Perform a linear transformation on the lateral position information l obtained in step S54 to obtain a query matrix Q corresponding to the lateral position information l. l , key matrix K l Sum value matrix V l Similarly, the longitudinal position information z obtained in step S54 is linearly transformed to obtain the query matrix Q corresponding to the longitudinal position information z z , key matrix K z Sum value matrix V z ;

[0130] S56, interactive calculation of lateral position information l Self-attention score: The query matrix Q of the vertical position information z z , and the key matrix K of the lateral position information l l The transposed matrix of Perform a dot multiplication operation to obtain the correlation information between the vectors (first correlation information), and pass the correlation information through the Softmax function to obtain a correlation matrix (first correlation matrix); then convert the value matrix V of the lateral position information l into l Multiplying by the correlation matrix (first correlation matrix) obtains the self-attention score of the horizontal position information l; the mathematical formula of this process is expressed as: Where d l Denotes the key matrix K l The dimension of .

[0131] Interactively calculate the self-attention score of the longitudinal position information z: the horizontal position information l The query matrix Q l , and the key matrix K of the longitudinal position information z z The transposed matrix of is multiplied by a dot operation to obtain the correlation information between the vectors (the second correlation information), and the correlation information is passed through the Softmax function to obtain the correlation matrix (the second correlation matrix); then the value matrix V of the longitudinal position information z is converted into z Multiplying it with the correlation matrix (the second correlation matrix) yields the self-attention score of the longitudinal position information z; the mathematical formula for this process is: Where d z Denotes the key matrix K z The dimension of .

[0132] Through steps S55-S56, the self-attention score of the horizontal position information l is obtained. l And the self-attention score of the longitudinal position information z z After that, according to Attention l and Attention z Obtain the output information of the position interaction self-attention mechanism module. Specifically, this step includes: S57, converting the self-attention score of the horizontal position information into l Attention is performed with the self-attention score of the longitudinal position information z Add them together and perform residual connection with the input deep feature information x1 to obtain the output information of the position interaction self-attention mechanism. This process is mathematically described as: Attention = Attention l +Attention z +x1. Here Attention l and Attention z They respectively represent the horizontal and vertical position information of the input information of the interactive self-attention mechanism.

[0133] The MSDFM-ASTRSB rolling bearing fault diagnosis model finally uses the Softmax classifier to map the output information Attention of the position interaction self-attention mechanism module to obtain the fault diagnosis result of the rolling bearing. Specifically, after obtaining Attention, the high-dimensional features of the output information of the position interaction self-attention mechanism are firstly globally averaged pooled to convert the multi-channel two-dimensional feature map into a one-dimensional feature vector. Then the multi-channel one-dimensional vector after global average pooling is expanded into a flat one-dimensional vector (also called feature map flattening operation, assuming that the output size of the feature map after global average pooling is 1×1×c, then it is a c-dimensional vector after expansion). Then, the fully connected layer is used to reduce the dimension of the flattened one-dimensional features, and finally the Softmax classifier is used to map the reduced feature information to finally obtain the fault diagnosis result of the rolling bearing.

[0134] In order to verify the effectiveness of the present invention, the present invention uses the Case Western Reserve University bearing dataset and the roller gear dataset to perform rolling bearing fault diagnosis verification under limited samples.

[0135] Verification Example 1:

[0136] Verification Example 1 uses the bearing data set CWRU of Case Western Reserve University in the United States. The experimental device of CWRU mainly includes a torque sensor, a motor, a bearing and other electronic equipment. The model of the tested bearing is SKF6205, and the sampling frequency is 12kHz. According to the different speeds and loads of the motor, data under four working conditions are collected. In this verification example, the data obtained under the condition of motor load 0HP and motor speed 1797r / min are marked as data set A, the data obtained under the condition of motor load 1HP and motor speed 1772r / min are marked as data set B, the data obtained under the condition of motor load 2HP and motor speed 1750r / min are marked as data set C, and the data obtained under the condition of motor load 3HP and motor speed 1750r / min are marked as data set D. According to the location of the fault and the size of the fault, this verification example selects five fault states under each working condition for experiment, and the specific faults and labels are shown in Table 1.

[0137] Table 1 Specific faults and corresponding labels in the CWRU dataset

[0138] Label Fault size Fault location 0 0.007 Rolling element 1 0.007 Inner Circle 2 0.014 Inner Circle 3 0.014 Outer ring 4 0.021 Outer ring

[0139] Each fault sample generates 200 multi-channel data fusion images, of which 50 are training sets, 50 are validation sets, and 100 are test sets. In order to verify the fault diagnosis ability of the ASTRSB-PISAM fault diagnosis model under small sample conditions, 5, 10, 25 and 50 samples are randomly selected from the training set of each fault to construct 4 different training sets of samples. The number of validation sets corresponding to the 4 different training sets is the same as the number of samples in the training set, and the number of test set samples is 100. Table 2 shows the detailed division of the 4 different training set samples in this embodiment.

[0140] Table 2 Detailed division of 4 different training set samples

[0141] Dataset Motor speed (r / min) / load (HP) Number of training sets / validation sets / test sets for each fault Dataset A 1797 / 0 5 / 5 / 100,10 / 10 / 100,25 / 25 / 100,50 / 50 / 100 Dataset B 1772 / 1 5 / 5 / 100,10 / 10 / 100,25 / 25 / 100,50 / 50 / 100 Dataset C 1750 / 2 5 / 5 / 100,10 / 10 / 100,25 / 25 / 100,50 / 50 / 100 Dataset D 1730 / 3 5 / 5 / 100,10 / 10 / 100,25 / 25 / 100,50 / 50 / 100

[0142] Suppose there are n sensor data, namely X1, X2, ···, X n , data processing is mainly divided into the following steps:

[0143] Step 1: Data normalization. In order to prevent the influence of some singular data in the data set on fault diagnosis, the n sensor data are normalized. The normalized mathematical expression is as follows:

[0144]

[0145] in, is the normalized data, X i is the original data, X min For X i The minimum value of X max For X i The maximum value of .

[0146] Step 2: Data segmentation: The experimental data was segmented using the overlapping sampling method, with a sliding window size of 1028 and a step size of 64.

[0147] Step 3: Data conversion image: The one-dimensional data is converted into a two-dimensional fused image based on the multi-sensor data fusion module.

[0148] In order to better verify the advanced nature of the present invention, this verification example is compared with five methods: SC-MSCNN, MRANet, MTF-ResNet, MSCNN and MDCNN. SC-MSCNN converts the original signal into a two-dimensional image through the MTF method, and then designs a self-calibrated coordinate attention mechanism and a multi-scale convolutional neural network to solve the problem of rolling bearing fault diagnosis under small sample conditions. MRANet converts the original signal into a time-frequency diagram through short-time Fourier transform, and then uses multi-branch dilated convolution and improved residual blocks to extract shallow and deep features of fault information, and uses a recalibration mechanism to adjust feature weights, enhance local pulse excitation, and increase the interpretability of the network. MTF-ResNet uses MTF to convert vibration signals into two-dimensional images, retains the time correlation of the original signal, and constructs a deep residual network for fault identification. MSCNN generates two-dimensional images through CWT, designs efficient convolution blocks and residual structures, and then constructs a multi-scale convolutional network. The effectiveness of the proposed method is verified in two data sets. MDCNN converts vibration signals into two-dimensional images through MTF, and then builds a fault diagnosis model based on multi-dimensional attention mechanism and Elu linear unit to solve the fault classification problem under small sample conditions.

[0149] This verification example selects data under 0HP for verification, and the results are shown in Table 3.

[0150] Table 3 Fault diagnosis accuracy of different methods under small samples As shown in Table 3, when there are only 5 samples of each type of fault, the average fault diagnosis accuracy of the proposed method reaches 99.67%, while the average fault diagnosis accuracy of the five comparison methods is 78.86%, 53.67%, 76.73%, 78.94% and 51.38% respectively. This shows that when the number of faults is very small, the proposed method can still extract useful information from the samples and has a strong feature extraction capability. When the number of samples of each type of fault reaches 50, the fault diagnosis accuracy of SC-MSCNN, MTF-ResNet and MDCNN is 98.87%, 97.23% and 88.18% respectively, and the fault diagnosis accuracy of MRANet, MSCNN and the proposed method of the present invention all reaches 100%, indicating that the proposed method of the present invention has excellent fault diagnosis performance in a small sample environment. However, it is worth noting that since the proposed method uses an asymmetric soft threshold module and a position interactive self-attention module, the computational complexity of the network is increased. Therefore, except for MRANet, the running time of the proposed method of the present invention is higher than that of the other four methods.

[0151] In order to more intuitively demonstrate the fault diagnosis capability of the method proposed in the present invention, this embodiment uses T-SNE technology to visualize the classification layers of the six methods when Sample=10. The results are as follows: Figure 4 As shown. Figure 4 It can be clearly seen that in the visualization diagrams of the classification layers of the five comparison methods, the five types of faults are more or less misclassified, especially in MDCNN, where the five types of faults are almost mixed together. In the visualization diagram of the proposed method, the five types of faults are clustered and separated from each other, and there is no misclassification of faults, indicating that the proposed method has good fault classification ability.

[0152] Figure 4 To verify the classification layer visualization diagrams of the present invention and five comparative methods when Sample=10 in Example 1, (a) is the classification layer visualization diagram of the SC-MSCNN method; (b) is the classification layer visualization diagram of the MRANet method; (c) is the classification layer visualization diagram of the MTF-ResNet method; (d) is the classification layer visualization diagram of the MSCNN method; (e) is the classification layer visualization diagram of the MDCNN method; and (f) is the classification layer visualization diagram of the MSDFM-ASTRSB method of the present invention.

[0153] Verification Example 2

[0154] Verification Example 2 uses a roller gear data set. The data used in this verification example comes from a roller gear fault simulation test platform, which is mainly composed of a three-phase AC asynchronous motor, a load device, bearings of different fault types, a frequency converter, a sensor, a DHDAS dynamic signal acquisition and analysis system, etc. Among them, different types of bearings are fixed on the main shaft, and the main shaft is connected to the coupling, and finally the bearing is positioned on the bearing lock ring. The bearing model used in the experiment is a 3 / 4-inch deep groove ball bearing, and the fault states are inner ring, outer ring, rolling element, mixed fault and normal state. The motor frequency is adjusted by the frequency converter to 20Hz, 30Hz and 40Hz respectively. The loading device is used at a fixed motor frequency to collect data of 5 different working conditions with loads of 0HP, 1HP, 2HP, 3HP and 4HP respectively through a single control variable method. The sample data acquisition frequency is 25600Hz, and the acceleration data in the x-axis and y-axis directions are collected using a shear acceleration sensor. The fault states selected in this experiment are inner ring fault, outer ring fault, rolling element fault and mixed fault. In this embodiment, the dual sensor fusion data of the x-axis and y-axis of the working conditions of 1HP, 2HP and 3HP are selected. The rest of the settings are the same as the CWRU data set.

[0155] In order to verify the generalization performance of the method proposed in the present invention under small sample conditions, this verification example selects data under 1HP, 2HP and 3HP for experiments, and the results are shown in Table 4.

[0156] Table 4 Fault diagnosis accuracy under different working conditions

[0157]

[0158] As shown in Table 7, after using the roller gear dataset, the accuracy of the proposed method slightly decreased. When the number of samples of each type of fault was 5, the average accuracy of the proposed method under the three working conditions was 97.67%, 98.88% and 97.16%, respectively. When the number of samples of each type of fault was 50, the average accuracy of the proposed method under the three working conditions was 100%, 99.98% and 100%, respectively. The experimental results show that the proposed method has excellent generalization ability under small sample conditions.

[0159] In order to further illustrate the superiority of the method proposed in the present invention, this experiment selected data under 2HP conditions for verification and compared it with 5 advanced methods. The experimental results are shown in Table 5.

[0160] Table 5 Fault diagnosis accuracy of different methods

[0161]

[0162] As shown in Table 4, after using the roller gear dataset, the fault diagnosis accuracy of MRANet dropped most significantly, which shows that the generalization ability of this method is poor. When Sample=5, the average fault diagnosis accuracy of SC-MSCNN, MTF-ResNet, MSCNN and MDCNN are 78.89%, 69.83%, 96.75% and 71.52% respectively, while the average accuracy of the method proposed in this invention reaches 98.88%, which shows that the feature extraction ability of the method proposed in this invention is stronger under the condition of small sample and has higher fault diagnosis accuracy.

[0163] In order to more intuitively demonstrate the fault diagnosis performance of the method proposed in the present invention, this embodiment selects 5 samples of each type of fault and conducts confusion matrix experiments on 6 methods. The results are as follows: Figure 5 shown.

[0164] Figure 5 Confusion matrix diagrams of different methods when Sample=5 in Example 2, wherein (a) is the confusion matrix diagram of the SC-MSCNN method; (b) is the confusion matrix diagram of the MRANet method; (c) is the MTF-ResNet; (d) is the confusion matrix diagram of the MSCNN method; (e) is the confusion matrix diagram of the MDCNN method; and (f) is the confusion matrix diagram of the MSDFM-ASTRSB method of the present invention.

[0165] Depend on Figure 5It can be seen that when the number of samples of each type of fault is 5, SC-MSCNN, MRANet, MTF-ResNet and MDCNN have serious misclassification phenomena, and their fault diagnosis accuracy rates are 77.50%, 50.25%, 67.00% and 77.25% respectively. This shows that the fault diagnosis capabilities of the four methods are limited in a small sample environment. Only 10 test samples in MSCNN were misclassified, and the diagnostic accuracy rate reached 97.25%, which is a significant improvement over the CWRU dataset under the same conditions, indicating that the method has a strong generalization ability. The accuracy of the method proposed in the present invention is 99.00%, which is significantly higher than the comparison method, indicating that the method proposed in the present invention not only has a high fault diagnosis accuracy rate in a small sample environment, but also has excellent generalization ability.

[0166] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0167] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A rolling bearing fault diagnosis method based on MSDFM-ASTRSB under small sample conditions, characterized in that: The following steps are involved: S1: Use multiple sensors to collect one-dimensional vibration signals of rolling bearings under various fault conditions; S2: The one-dimensional vibration signal collected by a single sensor under a single fault state is converted into a corresponding two-dimensional time-frequency diagram through continuous wavelet transform; S3: Fusing the two-dimensional time-frequency images of multiple sensors under the same fault state to obtain a multi-sensor data fusion image under a single fault state; S4: repeat step S3, and use the multi-sensor data fusion images in all fault states as a data set, and divide the data set into a training set, a verification set, and a test set; S5: Based on the asymmetric soft threshold residual shrinkage block and the position interaction self-attention mechanism, a MSDFM-ASTRSB rolling bearing fault diagnosis model is constructed, and the MSDFM-ASTRSB rolling bearing fault diagnosis model is trained with the training set obtained in step S4; S6: Use the trained MSDFM-ASTRSB rolling bearing fault diagnosis model to complete the rolling bearing fault diagnosis.

2. The rolling bearing fault diagnosis method based on MSDFM-ASTRSB under a small sample according to claim 1 is characterized in that: Step S2 specifically includes the following steps: S21, normalizing a one-dimensional vibration signal collected by a single sensor under a single fault state; S22, dividing the normalized one-dimensional vibration signal by using an overlapping sampling method to obtain a one-dimensional sample data set of a single sensor under a single fault state; S23, performing continuous wavelet transform on the one-dimensional sample data set obtained in step S22 to obtain wavelet coefficients of the input one-dimensional sample data, and mapping the wavelet coefficients into corresponding two-dimensional time-frequency diagrams.

3. The rolling bearing fault diagnosis method based on MSDFM-ASTRSB under small sample according to claim 2 is characterized in that: Step S21 normalizes the one-dimensional vibration signal collected by a single sensor under a single fault state, specifically including using the following normalization formula: in, is the i-th data in the normalized one-dimensional vibration signal, X i is the i-th original data in the one-dimensional vibration signal before normalization, X min is the minimum value of the one-dimensional vibration signal before normalization, X max is the maximum value of the one-dimensional vibration signal before normalization.

4. The rolling bearing fault diagnosis method based on MSDFM-ASTRSB under small sample according to claim 2 is characterized in that: In step S22, the normalized one-dimensional vibration signal is divided using an overlapping sampling method, which specifically includes the following steps: The normalized one-dimensional vibration signal is overlapped and sampled using a sliding window with a size of 1028 and a step size of 64 to obtain a one-dimensional sample data set of a single sensor under a single fault state.

5. The rolling bearing fault diagnosis method based on MSDFM-ASTRSB under small sample according to claim 2 is characterized in that: Step S23 specifically includes the following steps: S231, select cmor3-3 wavelet basis function in the composite Morlet wavelet function, analyze the time-frequency characteristics of the one-dimensional sample data set, the expression of the cmor3-3 wavelet basis function is: Among them, σ represents the bandwidth parameter of the wavelet, π represents pi, t represents the sampling time, i represents the imaginary unit, f represents the center frequency of the wavelet, and exp(·) represents the exponential function; S232, perform continuous wavelet transform on the one-dimensional sample data set, obtain the wavelet coefficients of the input one-dimensional sample data, and map the wavelet coefficients of the input one-dimensional sample data into a two-dimensional time-frequency graph; wherein the specific expression of continuous wavelet transform is: Among them, CWT(a,b) represents the wavelet transform of the input one-dimensional sample data x(t), x(t) represents the input one-dimensional sample data, a represents the scale parameter, b represents the translation parameter, ψ * represents the conjugate of the wavelet function.

6. The rolling bearing fault diagnosis method based on MSDFM-ASTRSB under small sample conditions according to claim 1 is characterized in that: The two-dimensional time-frequency diagrams of multiple sensors under the same fault state are fused, which specifically includes the following steps: Obtain the pixel matrix corresponding to the two-dimensional time-frequency diagram of all single sensors under the same fault state; The pixel matrices corresponding to the two-dimensional time-frequency graphs of multiple single sensors under the same fault state are fused in the height direction to obtain a pixel fusion matrix; The pixel fusion matrix is ​​mapped into a multi-sensor fusion two-dimensional time-frequency diagram under a single fault state.

7. The rolling bearing fault diagnosis method based on MSDFM-ASTRSB under small sample conditions according to claim 1 is characterized in that: The MSDFM-ASTRSB rolling bearing fault diagnosis model constructed in step S4 includes a feature extraction module, an asymmetric soft threshold residual shrinkage block, a position interaction self-attention mechanism module and a Softmax classifier; The feature extraction module is used to extract features from the multi-sensor data fusion image in the input training set, and obtain shallow feature information corresponding to each multi-sensor data fusion image; The asymmetric soft threshold residual shrinkage block is used to further extract the shallow feature information of each multi-sensor data fusion image and output the deep feature information; The position interaction self-attention mechanism module is used to further process the deep feature information to obtain the output information of the position interaction self-attention mechanism module; The Softmax classifier is used to map the output information of the position interaction self-attention mechanism module to obtain the fault diagnosis result of the rolling bearing.

8. The rolling bearing fault diagnosis method based on MSDFM-ASTRSB under small sample conditions according to claim 7 is characterized in that: The asymmetric soft threshold residual shrinkage block is used to further extract the shallow feature information of each multi-sensor data fusion image and output the deep feature information, which specifically includes the following steps: Divide shallow feature information into positive feature information and negative feature information; For positive feature information, a global average pooling operation is performed to obtain a positive soft vector; the obtained positive soft vector is input into two fully connected layers for feature combination and conversion; the feature combination and conversion results are input into the Sigmoid function to obtain a positive dynamic adjustable factor; the obtained positive soft vector is multiplied by the positive dynamic adjustable factor to obtain a positive threshold; For negative feature information, a global average pooling operation is performed to obtain a negative soft vector; the obtained negative soft vector is input into two fully connected layers for feature combination and conversion; the feature combination and conversion results are input into the Sigmoid function to obtain a negative dynamic adjustable factor; the obtained negative soft vector is multiplied by the negative dynamic adjustable factor to obtain a negative threshold; The shallow feature information is globally averaged and then passed through two fully connected layers and a Sigmoid layer to obtain an adjustable factor between the positive and negative thresholds. An asymmetric soft threshold function is constructed according to the positive threshold, the negative threshold and the adjustable factor between the positive and negative thresholds, and the characteristic information output by the asymmetric soft threshold function is obtained. The mathematical expression of the asymmetric soft threshold function is: In the formula, y represents the feature information output by the asymmetric soft threshold function; ω1 represents the positive threshold; ω2 represents the negative threshold; α represents the adjustable factor between the positive and negative thresholds; x represents the shallow feature information of the input; The feature information output by the asymmetric soft threshold function is added to the original information of the corresponding multi-sensor data fusion image to obtain the deep feature information of the multi-sensor data fusion image.

9. The rolling bearing fault diagnosis method based on MSDFM-ASTRSB under small sample according to claim 7 is characterized in that: The position interaction self-attention mechanism module further processes the deep feature information to obtain the output information of the position interaction self-attention mechanism module, which specifically includes the following steps: S51, performing global average pooling operations in H dimensions and W dimensions on the deep feature information output by the asymmetric soft threshold residual shrinkage block, respectively, to obtain a horizontal vector representation corresponding to the H dimension and a vertical vector representation corresponding to the W dimension; S52, performing two-layer convolution operations on both the horizontal vector representation and the vertical vector representation, so as to sequentially scale and restore the number of channels of the vectors; S53, respectively passing the horizontal vector representation and the vertical vector representation after the channel number is restored through the Sigmoid function to obtain a horizontal position vector corresponding to the horizontal vector representation and a vertical position vector corresponding to the vertical vector representation; S54, multiplying the lateral position vector by the deep feature information to obtain lateral position information of the deep feature information; Multiplying the longitudinal position vector by the deep feature information to obtain the longitudinal position information of the deep feature information; S55, performing a linear transformation on the lateral position information obtained in step S54 to obtain a query matrix, a key matrix, and a value matrix corresponding to the lateral position information; Performing a linear transformation on the longitudinal position information obtained in step S54 to obtain a query matrix, a key matrix and a value matrix corresponding to the longitudinal position information; S56, performing a point multiplication operation on the query matrix of the longitudinal position information and the transposed matrix of the key matrix of the lateral position information to obtain first correlation information, and performing a Softmax function on the first correlation information to obtain a first correlation matrix; then multiplying the value matrix of the lateral position information by the first correlation matrix to obtain a self-attention score of the lateral position information; Perform a point multiplication operation on the query matrix of the horizontal position information and the transposed matrix of the key matrix of the vertical position information to obtain the second correlation information, and obtain the second correlation matrix by the Softmax function of the second correlation information; then multiply the value matrix of the vertical position information by the second correlation matrix to obtain the self-attention score of the vertical position information; S57. Add the self-attention score of the horizontal position information and the self-attention score of the vertical position information, and perform a residual connection with the input deep feature information to obtain the output information of the position interaction self-attention mechanism.

10. A rolling bearing fault diagnosis device based on MSDFM-ASTRSB under small sample, characterized in that: It includes a computer system, which can implement the rolling bearing fault diagnosis method based on MSDFM-ASTRSB under a small sample as described in any one of claims 1-9 when executed.

Images