A gear multi-sensor collaborative fault diagnosis method

Abstract

The present invention provides a gear multi-sensor collaborative fault diagnosis method, comprising: fusing data from sensors on a gear monitoring platform, emphasizing the mutual relationship and temporal correlation between different time points in the time series data, and generating a multidimensional distance matrix image; constructing a multi-scale adaptive feature fusion convolutional neural network diagnostic model, and accurately identifying gear faults using multi-scale features and pyramid features. In the diagnostic model, a multi-scale adaptive feature fusion convolutional neural network is designed, and the multi-dimensional distance matrix image is input to perform fault diagnosis; a new convolution block is constructed, in which two large-scale depthwise separable convolutions of different sizes are used to obtain spatial information, and its attention branch aggregates position information along two directions; an adaptive fusion module is constructed to adaptively fuse pyramid features, introduce efficient attention and cross-space learning, generate learnable weights based on the importance of features at different stages, and comprehensively utilize deep and shallow information to diagnose gear faults.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating equipment fault diagnosis, and in particular to a gear multi-sensor collaborative fault diagnosis method based on a multi-dimensional distance matrix and a convolutional network. Background Art

[0002] Multi-sensor collaborative fault diagnosis uses multiple sensors to obtain monitoring data, enhances key information that can reflect the operating status of the equipment, effectively integrates multi-source complementary information through data fusion technology, and combines deep learning models to effectively diagnose different types of faults. Existing multi-sensor collaborative fault diagnosis methods for some rotating machinery, including gears, still have the following three problems:

[0003] 1) Existing methods typically develop multi-input models to facilitate data fusion, but lack solutions for achieving fusion during the data-to-image process. Multi-input models increase computational cost, reduce diagnostic efficiency, and suffer from difficulties in fitting, making them incompatible with the requirements of intelligent fault diagnosis in industry.

[0004] 2) Existing methods usually use more traditional deep learning models as diagnostic frameworks. These models have problems such as weak feature extraction capabilities and complex calculations, and are not suitable for direct use in intelligent fault diagnosis in industry.

[0005] 3) Existing methods usually only use abstract features in the deep layer of the network for classification, while ignoring favorable diagnostic features such as texture information in shallow features, which limits the performance of intelligent diagnosis models. Summary of the Invention

[0006] Based on the technical problems raised above, in order to further improve the recognition accuracy of some existing rotating machinery multi-sensor collaborative fault diagnosis methods, a gear multi-sensor collaborative fault diagnosis method based on multi-dimensional distance matrix and convolutional network is proposed.

[0007] The technical means adopted in the present invention are as follows:

[0008] A gear multi-sensor collaborative fault diagnosis method based on a multidimensional distance matrix and a convolutional network, comprising:

[0009] S1. Fusion of data from different sensors of the gear monitoring platform, emphasizing the mutual relationship and temporal correlation between different time points in the time series data, and generating a multi-dimensional distance matrix image;

[0010] S2. Construct a multi-scale adaptive feature fusion convolutional neural network diagnostic model and use multi-scale features and pyramid features to achieve accurate gear fault identification.

[0011] Furthermore, step S1 specifically includes:

[0012] S11. In the data fusion process of multidimensional distance matrix, let A=[a1,a2,...,a n ],B=[b1,b2,...,b n ],C=[c1,c2,...,c n ] are three sets of time series data with length n;

[0013] S12. Combine the three sets of time series data with length n and timestamp t to obtain X = [(a1, b1, c1, 1), (a2, b2, c2, 2), ..., (a n ,b n ,c n ,n)]; Encode X into a two-dimensional matrix M, each element in M ​​is recorded as

[0014]

[0015] In the above formula, x i and x j Represents (a i ,b i ,c i ,i) and (a j ,b j ,c j ,j), It is obtained from the multidimensional standard Euclidean distance matrix as follows:

[0016]

[0017] The above formula is a process of solving the multi-dimensional standard Euclidean distance. The above formula calculates the correlation between any two time points in the original time series data from different sources. The purpose of dividing by the standard deviation is to eliminate the influence of different scales. In the formula, σ a ,σ b ,σ c ,σ t are the standard deviations of A, B, C, and t respectively;

[0018] S13, transform M into an antisymmetric matrix M am And mapped to the range of [0,255], the multidimensional distance matrix (MDM) image is obtained as follows:

[0019]

[0020] Furthermore, the multi-scale adaptive feature fusion convolutional neural network diagnostic model constructed in step S2 includes a new convolution block and an adaptive feature fusion module, specifically including:

[0021] S21. Construct a new convolutional block to extract multi-scale information from the image. The attention branch can aggregate the position information of the image in two directions.

[0022] S22. Construct an adaptive feature fusion module to comprehensively utilize pyramid features at different stages for classification tasks.

[0023] Furthermore, the new convolutional block utilizes two large-scale depthwise separable convolutions of the main branch to obtain spatial information of different scales, utilizes the attention branch to aggregate position information along two directions and adds two one-dimensional convolution layers on the low dimension to alleviate the limitations of the shared convolution kernel mechanism.

[0024] Furthermore, step S21 specifically includes:

[0025] S211, assuming that the intermediate feature of the processing is The intermediate features are divided into two features in the channel dimension and And use two convolutions to extract features as follows:

[0026] Y1=DWconv 7×7 (X1)

[0027] Y2=DWconv 11×11 (X2)

[0028] Among them, DWconv 7×7 and DWconv 11×11 There are two channel-by-channel convolutions of size 7 and 11. The features processed by these two convolutions are Y1 and Y2 respectively; Y1 and Y2 are concatenated into new main branch features in the channel dimension.

[0029] S212. In the attention branch, use the pooling layers (1, W) and (H, 1) to aggregate the features of X in two directions:

[0030]

[0031] in, and are the output features of width w and height h on the cth channel respectively;

[0032] S213, will and Reduce the dimension and obtain low-dimensional features f w and f h :

[0033] f w =σ1(F w1 (z w))

[0034] f h =σ1(F h1 (z h ))

[0035] Among them, F w1 and F h1 There are two convolutional layers of size 1, σ1 represents the activation function;

[0036] S214, use two one-dimensional convolutional layers to process features and map them back to high latitude to obtain the output feature g of the attention branch w and g h :

[0037] g w =σ2(F w2 (DWconv1D3(f w )))

[0038] g h =σ2(F h2 (DWconv1D3(f h )))

[0039] Among them, F w1 and F h1 is a convolutional layer of size 1, DWconv1D3 is a one-dimensional convolutional layer of size 3, and σ2 represents the activation function

[0040] S215. Use the output features of the attention branch to reweight the output features Y of the main branch to obtain the feature R:

[0041]

[0042] S216. Use the multi-layer perceptron to process R and add it to the body X to obtain the output Q of the convolution block:

[0043] Q=F4(σ1(F3(σ1(R))))+X

[0044] Among them, F3 and F4 are two convolutional layers of size 1.

[0045] Furthermore, the adaptive feature fusion module introduces efficient attention and cross-space learning to extract important local and global information of pre-fused features.

[0046] Furthermore, step S22 specifically includes:

[0047] S221, for a given n different stage features S i , i∈n, alignment feature S i The size of the obtained n S is equal in sizei ′:

[0048] S i ′=Avgpool(F5(S i ))

[0049] Among them, Avgpool is the average pooling operation, F5 is a convolutional layer of size 1;

[0050] S222, use global average pooling to change S i ' to 1 and concatenate in the H dimension to get a new tensor

[0051] S223. Utilize efficient attention and cross-space learning to generate a set of learnable weights S′′:

[0052] S″′ L =F6(S″)

[0053] S″′ G =F7(DWconv n×1 (S″))

[0054] S″′=Sigmoid(GAP(S″′ L ))⊙S″′ G +Sigmoid(S″′ G )⊙S″′ L

[0055] Among them, S″′ L and S″′ G are local features and global features, F6 and F7 are two convolutional layers of size 1, DWconv n×1 It is a channel-by-channel convolution, and GAP is a global average pooling operation;

[0056] S224, use the softmax function operation on the H dimension to obtain i weights

[0057]

[0058] S225, will w i Alignment features S at different stages i ′ are multiplied and then added to obtain the output T of the adaptive feature fusion module:

[0059]

[0060] Compared with the prior art, the present invention has the following advantages:

[0061] 1. This invention provides a multi-sensor collaborative gear fault diagnosis method based on a multidimensional distance matrix and a convolutional network. The proposed multidimensional distance matrix method can convert multi-source signals into images. This scheme uses standard Euclidean distance to emphasize the relationship between any two points in time series data while preserving temporal correlation. The multidimensional distance matrix integrates data from multiple sensors, providing a novel information fusion method that simultaneously encodes images.

[0062] 2. This paper provides a multi-sensor collaborative gear fault diagnosis method based on a multidimensional distance matrix and a convolutional network. To enhance the network's ability to acquire multi-scale information, a novel convolutional block is constructed. Two large-scale depthwise separable convolutions of different sizes are used to acquire spatial information. Its attention branch aggregates position information in two directions, and two one-dimensional convolutional layers are introduced in the low-dimensional layer to mitigate the limitations of convolutional parameter sharing and enhance the representation of spatial information.

[0063] 3. This paper provides a multi-sensor collaborative fault diagnosis method for gears based on a multidimensional distance matrix and a convolutional network. It proposes an adaptive feature fusion module that utilizes features from different stages of the network to achieve adaptive fusion of pyramidal features. This module incorporates efficient attention and cross-spatial learning to extract important local and global information about pre-fused features. It generates learnable weights for features at different stages based on their respective importance, effectively leveraging multi-layer features to improve fault diagnosis performance.

[0064] Based on the above reasons, the present invention can be widely promoted in the fields of rotating equipment fault diagnosis and the like. 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 following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0066] Figure 1 Flowchart of the gear diagnosis method of the present invention.

[0067] Figure 2 A multidimensional distance matrix image of multi-source signal encoding of a gear in nine different health states provided by an embodiment of the present invention.

[0068] Figure 3 This is the multi-scale adaptive feature fusion convolutional neural network of the present invention.

[0069] Figure 4 Schematic diagram of the new convolution block of the present invention.

[0070] Figure 5 Schematic diagram of the adaptive feature fusion module of the present invention.

[0071] Figure 6 A rotating equipment test bench is provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described 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 making creative efforts should fall within the scope of protection of the present invention.

[0073] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0074] like Figure 1 As shown, the present invention provides a gear multi-sensor collaborative fault diagnosis method based on a multidimensional distance matrix and a convolutional network, comprising:

[0075] S1. Fusion of data from different sensors of the gear monitoring platform, emphasizing the mutual relationship and temporal correlation between different time points in the time series data, and generating a multi-dimensional distance matrix image;

[0076] S2. Construct a multi-scale adaptive feature fusion convolutional neural network diagnostic model and use multi-scale features and pyramid features to achieve accurate gear fault identification.

[0077] In specific implementation, as a preferred embodiment of the present invention, step S1 specifically includes:

[0078] S11. In the data fusion process of multidimensional distance matrix, let A=[a1,a2,...,a n ],B=[b1,b2,...,b n ],C=[c1,c2,...,cn ] are three sets of time series data with length n;

[0079] S12. Combine the three sets of time series data with length n and timestamp t to obtain X = [(a1, b1, c1, 1), (a2, b2, c2, 2), ..., (a n ,b n ,c n ,n)]; Encode X into a two-dimensional matrix M, each element in M ​​is recorded as

[0080]

[0081] In the above formula, x i and x j Represents (a i ,b i ,c i ,i) and (a j ,b j ,c j ,j), It is obtained from the multidimensional standard Euclidean distance matrix as follows:

[0082]

[0083] The above formula is a process of solving the multi-dimensional standard Euclidean distance. The above formula calculates the correlation between any two time points in the original time series data from different sources. The purpose of dividing by the standard deviation is to eliminate the influence of different scales. In the formula, σ a ,σ b ,σ c ,σ t are the standard deviations of A, B, C, and t respectively;

[0084] S13, transform M into an antisymmetric matrix M am And mapped to the range of [0,255], such as Figure 2 As shown, the multidimensional distance matrix (MDM) image is obtained as follows:

[0085]

[0086] When specifically implemented, as a preferred embodiment of the present invention, Figure 3 As shown, the multi-scale adaptive feature fusion convolutional neural network diagnostic model constructed in step S2 includes a new convolution block and an adaptive feature fusion module, specifically including:

[0087] S21. Construct a new convolutional block to extract multi-scale information from the image. The attention branch can aggregate the position information of the image in two directions.

[0088] S22. Construct an adaptive feature fusion module to comprehensively utilize pyramid features at different stages for classification tasks.

[0089] In specific implementation, as a preferred embodiment of the present invention, the new convolution block uses two large-size depth-separable convolutions of the main branch to obtain spatial information of different scales, uses the attention branch to aggregate position information in two directions and adds two one-dimensional convolution layers on the low dimension to alleviate the limitations of the shared convolution kernel mechanism.

[0090] In specific implementation, as a preferred embodiment of the present invention, step S21, as Figure 4 As shown, specifically including:

[0091] S211, assuming that the intermediate feature of the processing is The intermediate features are divided into two features in the channel dimension and And use two convolutions to extract features as follows:

[0092] Y1=DWconv 7×7 (X1)

[0093] Y2=DWconv 11×11 (X2)

[0094] Among them, DWconv 7×7 and DWconv 11×11 There are two channel-by-channel convolutions of size 7 and 11. The features processed by these two convolutions are Y1 and Y2 respectively; Y1 and Y2 are concatenated into new main branch features in the channel dimension.

[0095] S212. In the attention branch, use the pooling layers (1, W) and (H, 1) to aggregate the features of X in two directions:

[0096]

[0097] in, and are the output features of width w and height h on the cth channel respectively;

[0098] S213, will and Reduce the dimension and obtain low-dimensional features f w and f h :

[0099] f w=σ1(F w1 (z w ))

[0100] f h =σ1(F h1 (z h ))

[0101] Among them, F w1 and F h1 There are two convolutional layers of size 1, σ1 represents the activation function;

[0102] S214, use two one-dimensional convolutional layers to process features and map them back to high latitude to obtain the output feature g of the attention branch w and g h :

[0103] g w =σ2(F w2 (DWconv1D3(f w )))

[0104] g h =σ2(F h2 (DWconv1D3(f h )))

[0105] Among them, F w1 and F h1 is a convolutional layer of size 1, DWconv1D3 is a one-dimensional convolutional layer of size 3, and σ2 represents the activation function

[0106] S215. Use the output features of the attention branch to reweight the output features Y of the main branch to obtain the feature R:

[0107]

[0108] S216. Use the multi-layer perceptron to process R and add it to the body X to obtain the output Q of the convolution block:

[0109] Q=F4(σ1(F3(σ1(R))))+X

[0110] Among them, F3 and F4 are two convolutional layers of size 1.

[0111] In practice, as a preferred embodiment of the present invention, the adaptive feature fusion module incorporates efficient attention and cross-spatial learning to extract important local and global information from pre-fused features. It generates learnable weights for features at different stages based on their importance, making it a highly efficient feature fusion technique.

[0112] In specific implementation, as a preferred embodiment of the present invention, step S22, as Figure 5 As shown, specifically including:

[0113] S221, for a given n different stage features S i , i∈n, alignment feature S i The size of the obtained n S is equal in size i ′:

[0114] S i ′=Avgpool(F5(S i ))

[0115] Among them, Avgpool is the average pooling operation, F5 is a convolutional layer of size 1;

[0116] S222, use global average pooling to change S i ' to 1 and concatenate in the H dimension to get a new tensor

[0117] S223. Utilize efficient attention and cross-space learning to generate a set of learnable weights S′′:

[0118] S″′ L =F6(S″)

[0119] S″′ G =F7(DWconv n×1 (S″))

[0120] S″′=Sigmoid(GAP(S″′ L ))⊙S″′ G +Sigmoid(S″′ G )⊙S″′ L

[0121] Among them, S″′ L and S″′ G are local features and global features, F6 and F7 are two convolutional layers of size 1, DWconv n×1 It is a channel-by-channel convolution, and GAP is a global average pooling operation;

[0122] S224, use the softmax function operation on the H dimension to obtain i weights

[0123]

[0124] S225, will w i Alignment features S at different stages i ′ are multiplied and then added to obtain the output T of the adaptive feature fusion module:

[0125]

[0126] Example

[0127] Experimental data introduction

[0128] By Figure 6 The rotating equipment test bench shown here collects data on gears in various health states to validate the effectiveness of the proposed method. The test bench includes two accelerometers, a microphone, a gearbox, and a data logger. The sampling frequency is set to 20 kHz, and the motor speed is 1430 rpm. The constructed dataset includes one set of healthy data and eight sets of fault data. Failure modes such as cracking, pitting, spalling, and wear are simulated. Each data type is divided into 1000 samples, with a sample length of 1000. The fault modes and labels are shown in Table 1.

[0129] Table 1 Detailed classification information of category labels and samples

[0130]

[0131] Comparative experimental setup

[0132] The purpose of this section is to verify the effectiveness and superiority of the proposed convolutional neural network for multidimensional distance matrix image and multiscale adaptive feature fusion. The experimental verification process was conducted on an Intel Core i5-12400F CPU processor, a single NVIDIA GeForce RTX 3060 graphics processing unit, and the Windows 11 operating system. During the training process for the proposed method and the comparative method, the number of model updates was set to 50 and the number of batches was set to 32. The Adam optimization algorithm was used to update the feature extractor, with an initial learning rate of 0.001, which was gradually decreased by 0.5 times with a step size of 5. To reduce the randomness of the experimental results, each set of comparative experiments was repeated five times during the algorithm verification.

[0133] To further validate the performance of the proposed method, this example introduces several comparison methods. The experimental results are shown in Table 2. MAFFCNN is the proposed multi-scale adaptive feature fusion convolutional neural network (MAFFCNN). ConvNext-tiny, ResNet18, MobileNetV3-small, and MobileNetV3-large are popular CNN models used in many artificial intelligence applications, including fault diagnosis. LICNN and SC-MSCNN are the most recently proposed fault diagnosis models. RGB images are a common method used in multi-sensor fault diagnosis solutions to fuse data from multiple sources.

[0134] Result Analysis

[0135] This example conducted five tests, and the classification results are presented using the average and maximum diagnostic accuracy rates. It should be noted that the self-built gear dataset contains a certain degree of noise during data acquisition, which is more consistent with real-world industrial environments and enhances the rationality of the proposed method. Table 2 shows the statistical results of the diagnostic accuracy of the proposed method and other comparative methods.

[0136] Table 2 Statistics of the accuracy of different diagnostic methods

[0137]

[0138] 1) The multidimensional distance matrix image can achieve good performance in different models, and the diagnostic accuracy is quite stable, which proves that the multidimensional distance matrix image has excellent ability and good generalization performance in fusion encoding of multiple gear fault signals.

[0139] 2) Using a multi-scale adaptive feature fusion convolutional neural network, a gear diagnosis method based on multidimensional distance matrix images achieved an accuracy of 99.6%, surpassing the 99.2% accuracy achieved by traditional RGB images. This demonstrates that multidimensional distance matrix images outperform traditional methods in multi-source information fusion.

[0140] 3) Among various diagnostic models using multidimensional distance matrix images as input, the multi-scale adaptive feature fusion convolutional neural network achieved a diagnostic accuracy of 99.6%, while the diagnostic accuracy of other models ranged from 97.2% to 99.0%. This demonstrates the superiority of the multi-scale adaptive feature fusion convolutional neural network as a gear diagnosis model.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A gear multi-sensor collaborative fault diagnosis method, characterized in that: include: S1. Integrate data from different sensors on the gear monitoring platform, emphasize the interrelationships and temporal correlations between different time points in the time series data, and generate a multidimensional distance matrix image. Specifically, it includes: S11. In the data fusion process of multidimensional distance matrix, let A=[ a 1, a 2, ..., a n ], B=[ b 1, b 2, ..., b n ], C=[ c 1, c 2, ..., c n ] are three sets of time series data with length n; S12. Combine the three sets of time series data with length n and timestamp t to obtain X=[( a 1, b 1, c 1, 1), ( a 2, b 2, c twenty two), ..., ( a n , b n , c n , n )]; Encode X into a two-dimensional matrix M, and each element in M ​​is recorded as : In the above formula, x i and x j Respectively represent ( a i, b i, c i, i ) and( a j, b j, c j, j ), It is obtained from the multidimensional standard Euclidean distance matrix as follows: The above formula is a process of solving the multi-dimensional standard Euclidean distance. It calculates the correlation between any two time points in the original time series data from different sources. The purpose of dividing by the standard deviation is to eliminate the influence of different scales. In the formula, , , , are the standard deviations of A, B, C, and t respectively; S13, transform M into an antisymmetric matrix M am And mapped to the range of [0, 255], the image of the multidimensional distance matrix MDM is obtained as follows: ； S2. Construct a multi-scale adaptive feature fusion convolutional neural network diagnostic model, which uses multi-scale features and pyramid features to achieve accurate gear fault identification. The constructed multi-scale adaptive feature fusion convolutional neural network diagnostic model includes a new convolution block and an adaptive feature fusion module, specifically including: S21. Construct a new convolutional block to extract multi-scale information in the image, and the attention branch aggregates the position information of the image in two directions; S22. Construct an adaptive feature fusion module to comprehensively utilize pyramid features at different stages for classification tasks.

2. A gear multi-sensor collaborative fault diagnosis method according to claim 1, characterized in that: The new convolutional block utilizes two large-scale depthwise separable convolutions of the main branch to obtain spatial information of different scales, utilizes the attention branch to aggregate position information along two directions and adds two one-dimensional convolution layers on the low dimension to alleviate the limitations of the shared convolution kernel mechanism.

3. A gear multi-sensor collaborative fault diagnosis method according to claim 2, characterized in that: Step S21 specifically includes: S211, assuming that the intermediate feature of the processing is , the intermediate features are divided into two features in the channel dimension and , and use two convolutions to extract features as follows: Among them, DWconv 7×7 and DWconv 11×11 There are two channel-by-channel convolutions of size 7 and 11. The features processed by these two convolutions are Y 1 and Y 2. In the channel dimension, Y 1 and Y 2 Splice into a new main branch feature ; S212, on the attention branch, using the pooling layer (1, W )and( H , 1) Aggregation in two directions X Features: in, and are respectively c Channel width w and height h Output features on ; S213, will and Reduce the dimension and obtain low-dimensional features f w and f h : in, F w1 and F h1 are two convolutional layers of size 1, represents the activation function; S214, use two one-dimensional convolutional layers to process features and map them back to high dimensions to obtain the output features of the attention branch and : in, and are two convolutional layers of size 1, DWconv1D3 is a one-dimensional convolutional layer of size 3, represents the activation function; S215, using the output features of the attention branch to the output features of the main branch Y Reweight and get features R : S216, using multi-layer perceptron R After processing and X Add together to get the output of the convolution block Q : in, F 3 and F 4 is two convolutional layers of size 1.

4. The gear multi-sensor collaborative fault diagnosis method according to claim 1, characterized in that: The adaptive feature fusion module introduces efficient attention and cross-space learning to extract important local and global information of pre-fused features.

5. The gear multi-sensor collaborative fault diagnosis method according to claim 4, characterized in that: Step S22 specifically includes: S221、For a given n Characteristics of different stages S i , , align features S i The size of the n indivual : Among them, Avgpool is the average pooling operation, , F 5 is a convolutional layer of size 1; S222, using global average pooling changes to 1 and concatenate along the H dimension to get a new tensor ; S223, using efficient attention and cross-space learning to generate a set of learnable weights : in, and are local features and global features, F 6 and F 7 is two convolutional layers of size 1, DWconv n×1 It is a channel-by-channel convolution, and GAP is a global average pooling operation; S224, using the softmax function operation on the H dimension, obtain i Weight : S225, will w i Alignment features with different stages Multiply and then add to get the output of the adaptive feature fusion module T : 。

Images