Bearing fault diagnosis method and system based on double attention mechanism enhanced hierarchical decision network

Through the dual attention-guided tree-level decision-making network, the problem of accurate diagnosis of bearing fault detection in unbalanced data and strong noise environments is solved, and the reliable identification of bearing fault location and size is achieved, which improves the credibility of diagnosis and Interpretability.

CN119004204AActive Publication Date: 2024-11-22ZHEJIANG NORMAL UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202411056336.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-22
Estimated Expiration
2044-08-02

AI Technical Summary

Technical Problem

The existing bearing fault detection methods are difficult to achieve accurate diagnosis under unbalanced data and strong noise environments, and ignore the effective reasoning of the intermediate process, reduce the credibility of the diagnosis results, and mix the fault location and size as the same level. , deviates from the conventional thinking process of human beings.

Method used

Using a double attention-guided tree-level decision-making network, the information closely related to the fault information in the bearing fault signal is enhanced and interference information is weakened by integrating the ternary attention mechanism and multi-head convolutional attention mechanism into the CNN model. Design tree heuristic hierarchical decision-making network step by step decision failure location and size.

Benefits of technology

Accurate identification of bearing health status in an unbalanced and strong noise environment improves the accuracy and interpretability of fault diagnosis, conforms to the conventional cognitive process of humans, and is different from the traditional hybrid decision-making method of fault location and size at the same level.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004204A_ABST
    Figure CN119004204A_ABST
Patent Text Reader

Abstract

The invention discloses a bearing fault diagnosis method and system based on a double attention mechanism enhanced hierarchical decision network. The method comprises the following steps: collecting bearing vibration signals in different health states; constructing a hierarchical multi-class fault diagnosis model based on the bearing vibration signal; and determining the fault position and size of the bearing by using the hierarchical multi-class fault diagnosis model. According to the method, the health state of the bearing can be accurately identified in an unbalanced strong noise environment, and the decision-making process of the sample state type has interpretability. Meanwhile, the decision-making thought of positioning the fault position first and then quantifying the fault size is different from the previous practice of fault position and size same-level mixed decision making, and the method better conforms to the conventional cognition of human beings.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of bearing fault detection, and in particular to a bearing fault diagnosis method and system based on a dual attention mechanism reinforced hierarchical decision network. Background Art

[0002] In the existing detection of bearing faults, the fault diagnosis work of bearing health status classification is carried out under a relatively balanced data set; however, in the real fault diagnosis work, on the one hand, the data of bearing faults of different categories are unevenly distributed, making the training and testing of the model more complicated. On the other hand, the inevitable noise environment may further interfere with the feature extraction and diagnosis process of bearing signals. Therefore, how to achieve good bearing fault diagnosis in the environment of unbalanced data and strong noise interference is still a very challenging task. In terms of related technologies, although CNNs assisted by attention mechanisms have shown strong capabilities in feature extraction and achieved encouraging results in bearing fault diagnosis tasks, they have two main shortcomings: (1) they only focus on input signals and output results, but ignore the effective reasoning of the intermediate process, which reduces the credibility of the diagnosis results; (2) the bearing fault location and size are mixed and classified as the same level, which deviates from the normal thinking process of humans. Summary of the invention

[0003] In order to solve the technical problems in the above background, the present invention proposes a hierarchical multi-category fault diagnosis model of a double-attention guided tree-shaped hierarchical decision network, which aims to realize the fault diagnosis of bearing health status classification in an unbalanced strong noise environment.

[0004] To achieve the above object, the present invention proposes a bearing fault diagnosis method based on a dual attention mechanism to strengthen the hierarchical decision network, the steps comprising:

[0005] Collect bearing vibration signals under different health conditions;

[0006] Based on the bearing vibration signal, a hierarchical multi-category fault diagnosis model is constructed;

[0007] The hierarchical multi-class fault diagnosis model is used to determine the fault location and size of the bearing.

[0008] Preferably, the hierarchical multi-category fault diagnosis model includes a dual attention guidance mechanism and a tree-inspired hierarchical decision network;

[0009] The dual attention guidance mechanism is used to enhance the information in the bearing fault signal that is closely related to the fault information and weaken the interference information that is not closely related to the fault information;

[0010] The tree-inspired hierarchical decision network is used to determine the location and size of bearing faults step by step.

[0011] Preferably, the method for constructing the dual attention guidance mechanism includes: integrating the ternary attention mechanism and the multi-head convolutional attention mechanism into a CNN model to form the dual attention guidance mechanism;

[0012] Among them, the ternary attention mechanism introduces the convolutional block attention module through the concept of cross-dimensional interaction, making the interaction between channels and spatial dimensions more compact and comprehensive;

[0013] Construct a multi-head convolution self-attention mechanism based on single-head convolution attention.

[0014] Preferably, the tree-inspired hierarchical decision network is designed based on the basic logic of the fault diagnosis task, including: a fault type layer and a fault size layer;

[0015] The fault type layer is used to determine the fault type of the input sample;

[0016] The fault size layer is used to determine the fault size of input samples.

[0017] Preferably, the dual attention guidance mechanism is used as the backbone network of the hierarchical multi-category fault diagnosis model; at the same time, a tree-inspired hierarchical decision network with a two-layer structure is constructed using two fully connected layers, and the weight information generated by the fully connected layer in the backbone network is inherited by the threshold of the seed node of the tree-inspired hierarchical decision network, and the threshold of the leaf node is further determined by the embedded decision rule of the seed node and the leaf node.

[0018] The present invention also provides a bearing fault diagnosis system based on a dual attention mechanism reinforced hierarchical decision network, the system is used to implement the above method, including: an acquisition module, a construction module and a detection module;

[0019] The acquisition module is used to collect bearing vibration signals in different health states;

[0020] The building module is used to build a hierarchical multi-category fault diagnosis model based on the bearing vibration signal;

[0021] The detection module is used to determine the fault location and size of the bearing using the hierarchical multi-category fault diagnosis model.

[0022] Preferably, the hierarchical multi-category fault diagnosis model includes a dual attention guidance mechanism and a tree-inspired hierarchical decision network;

[0023] The dual attention guidance mechanism is used to enhance the information in the bearing fault signal that is closely related to the fault information and weaken the interference information that is not closely related to the fault information;

[0024] The tree-inspired hierarchical decision network is used to determine the location and size of bearing faults step by step.

[0025] Preferably, the workflow of the building module includes: integrating the ternary attention mechanism and the multi-head convolutional attention mechanism into a CNN model to form the dual attention guidance mechanism;

[0026] Among them, the ternary attention mechanism introduces the convolutional block attention module through the concept of cross-dimensional interaction, making the interaction between channels and spatial dimensions more compact and comprehensive;

[0027] Construct a multi-head convolution self-attention mechanism based on single-head convolution attention.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention can not only accurately identify the health status of the bearing in an unbalanced and noisy environment, but also make the decision process of the sample status type explainable. At the same time, the decision-making idea of ​​the present invention of "first locating the fault position, then quantifying the fault size" is different from the previous "mixed decision-making of fault position and size at the same level", which is more in line with human conventional cognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0031] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention;

[0032] Figure 2 It is a schematic diagram of the input part of the model structure of an embodiment of the present invention;

[0033] Figure 3 A schematic diagram of a dual attention guidance mechanism of a model structure according to an embodiment of the present invention;

[0034] Figure 4 A schematic diagram of a tree-inspired hierarchical decision layer of a model structure according to an embodiment of the present invention;

[0035] Figure 5 Schematic diagram of a ternary attention mechanism according to an embodiment of the present invention;

[0036] Figure 6 A detailed structural diagram of a tree-inspired hierarchical decision network according to an embodiment of the present invention;

[0037] Figure 7Schematic diagram of average classification recognition rate in an embodiment of the present invention; wherein, Figure 7 (a) represents the recognition rate corresponding to dataset A; Figure 7 (b) represents the corresponding recognition rate of data set B;

[0038] Figure 8 Schematic diagram of the reasoning decision process of an embodiment of the present invention; Figure 8 (a) Corresponding outer ring fault signal; Figure 8 (b) Corresponding inner race fault signal. DETAILED DESCRIPTION

[0039] 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.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] Embodiment 1

[0042] like Figure 1 As shown, it is a schematic diagram of the method flow of the present invention, and the steps include:

[0043] S1. Collect bearing vibration signals under different health conditions.

[0044] In this embodiment, the vibration signal samples collected are vibration signals of unbalanced bearing samples with various health states under a noise background collected from a fault test bench using a dedicated acceleration sensor.

[0045] S2. Based on the bearing vibration signal, a hierarchical multi-category fault diagnosis model is constructed.

[0046] In order to realize the fault diagnosis of bearing health status classification in an unbalanced and noisy environment, this embodiment proposes a hierarchical multi-category fault diagnosis model of a dual attention guided tree-inspired grade decision network (DATGDN), such as Figure 2 , 3 As shown in Figure 4, a hierarchical multi-class fault diagnosis model is used to diagnose the bearing vibration signal and determine the fault location and size. This not only reduces the complexity associated with the diagnostic task, but also enhances the accuracy of fault diagnosis.

[0047] The hierarchical multi-class fault diagnosis model includes a dual-attention guidance mechanism and a tree-inspired hierarchical decision network. Its construction method is as follows: first, the ternary attention mechanism and the multi-head convolutional attention mechanism are integrated into a new CNN model to form a dual-attention guidance mechanism to enhance the information closely related to the fault information in the bearing fault signal and weaken the interference information that is not closely related to the fault information; secondly, a hierarchical decision network with a two-layer tree structure is designed to form a tree-inspired hierarchical decision layer, which is used to decide the location and size of the bearing fault step by step; finally, the tree-inspired hierarchical decision layer is arranged after the dual-attention guided CNN model to seamlessly integrate the hierarchical multi-class fault diagnosis model.

[0048] like Figure 5 As shown, this is the ternary attention mechanism architecture of this embodiment. In order to more clearly describe the proposed ternary attention mechanism, this embodiment first introduces cross-dimensional interaction and Z-pooling operations.

[0049] Cross-dimensional interaction

[0050] Cross-dimensional interaction involves addressing the lack of interdependence between channel and spatial dimensions in traditional methods. Typically, traditional methods use singular weights when computing channel attention, resulting in a significant loss of spatial features in signal processing. This results in a lack of tight correlation between the channel and spatial dimensions. To overcome this problem, the Convolutional Block Attention Module (CBAM) was introduced, which uses spatial attention as a complementary module to channel attention. Specifically, CBAM makes up for the lack of where to focus within the channel by introducing spatial attention, allowing the model to have a more comprehensive understanding of where to focus within the channel. At the same time, channel attention still plays the role of specifying which channel to focus on, thereby maintaining the focus on channel information.

[0051] Z-pool

[0052] The task of Z-pool is to reduce the zeroth dimension of the tensor to two by connecting the features of global average pooling (GMP) and global maximum pooling (GAP) in dimension. The purpose of this operation is to improve the efficiency of subsequent calculations by reducing the depth while maintaining the comprehensive representation of the original tensor. The introduction of the Z-pool layer enables the model to process input data more efficiently while still maintaining a comprehensive grasp of the overall features. Its mathematical representation is shown in Equation 1. The Z-Pool operation generates a tensor of shape (2×H×W) from a tensor of shape (C×H×W).

[0053] Z-pool(χ)=[GMP 0d (χ),GAP 0d (x)] (1)

[0054] Where 0d represents the 0th dimension of the tensor on which the GMP and GAP operations are performed; x represents the input tensor.

[0055] like Figure 5 As shown, the ternary attention for a given input tensor x∈R C×H×W , which are passed to the three branches of the proposed ternary attention module respectively.

[0056] In the first branch, an interaction is constructed between the height dimension and the channel dimension. To achieve this function, first, the input tensor x is rotated 90° counterclockwise along the H axis to produce a rotated tensor x with a shape of (W×H×C) 0H ; Secondly, x 0H Further processing through Z-Pool operation results in x with shape size (2×H×C) 0HZ ; Again, x 0HZ A convolution operation is performed with a convolution kernel of size k×k, and batch normalization is performed to generate an intermediate output tensor of dimension (1×H×C); next, a sigmoid activation layer (σ) is performed on the intermediate output tensor to generate attention weights and apply them to x 0H ; Finally, rotate 90° clockwise along the H axis to produce an attention weight tensor with the same shape as the original input tensor x.

[0057] Similarly, in the second branch, the original signal is rotated 90° counterclockwise along the W axis to produce a rotation tensor x with a shape of (H×C×W) 1W ; Secondly, perform Z-pool operation on the tensor to obtain x with shape size (2×C×W) 1WZ ; Again, x 1WZ A convolution operation is performed with a convolution kernel of size k×k, and batch normalization is performed to generate an intermediate output tensor of dimension (1×H×C); next, a sigmoid activation layer (σ) is performed on the intermediate output tensor to generate attention weights and apply them to x 1W ; Finally, rotate 90° clockwise along the W axis to produce an attention weight tensor with the same shape as the original input tensor x.

[0058] For the third branch, first directly convert the input tensor x with a shape of (C×H×W) directly through the Z-pool operation to generate a tensor x with a shape of (2×H×W) 2Z ; Secondly, x 2Z A convolution operation is performed with a convolution kernel of size k×k, and batch normalization is performed to produce an intermediate output tensor of dimension (1×H×C); finally, a sigmoid activation layer (σ) is performed on the intermediate output tensor to generate an attention weight tensor of shape size (1×H×W).

[0059] Finally, the tensors of shape (C×H×W) generated by the three branches are merged and then averaged to obtain the refined attention tensor y. The calculation process is expressed as formula 2:

[0060]

[0061] Where ψ1, ψ2 and ψ3 all represent standard convolution operations.

[0062] Formula 2 can be simplified to Formula 3:

[0063]

[0064] Where ω1, ω2 and ω3 represent the three cross-dimensional attention weights calculated in the three branches of the ternary attention respectively; y1, y2 and y3 represent the attention weight tensors generated by the first branch, the second branch and the third branch respectively.

[0065] In short, the ternary attention mechanism introduces CBAM through the concept of cross-dimensional interaction, making the interaction between channel and spatial dimensions closer and more comprehensive. Furthermore, the ternary attention can simultaneously consider the information of channel, space, and channel-space cross-dimensions, improving the model's perception of key features.

[0066] This embodiment also proposes Multi-head Convolutional Self-Attention (MCSA) based on Single Convolutional Attention (SCA). It aims to reduce the cost of memory usage and computation during training or inference while maintaining the interactivity and diversity between multiple heads. MCSA overcomes the memory and computational limitations of standard MSA to more effectively cope with the processing requirements of large-scale data and high-dimensional data. This allows MCSA to handle complex tasks while maintaining the advantages of the multi-head self-attention mechanism. The specific method is as follows:

[0067] First, the one-dimensional original signal is converted into a two-dimensional input tensor using an embedded coding method. where n and d m Represent the spatial dimension and channel dimension of x1 respectively.

[0068] Second, similar to MSA, MCSA first uses a set of projections to obtain the query. At the same time, in order to effectively compress memory, the two-dimensional input tensor is reshaped into a three-dimensional tensor along the spatial dimension

[0069] Again, the three-dimensional input tensor Perform depth-wise separable convolution (DWConv) and layer normalization techniques to obtain a new three-dimensional tensor In a depth-wise separable convolution operation, the three-dimensional input tensor The new tensor is obtained by reducing its height and width dimensions by scaling factor s (depending on the size of the feature map). Among them, the convolution kernel size, step size and padding are s+1, s and s / 2 respectively.

[0070] Next, the new 3D input tensor is reshaped into a new 2D input tensor Two sets of fully connected feature mapping are performed to obtain Key and Value, where n = (h / s) × (w / s).

[0071] Furthermore, Formula 5 is used instead of Formula 4 to calculate the Query (Q), Key (K) and Value (V) in the attention function, and the Query and Key are multiplied and normalized, and then input into the convolution operation with Softmax and perform instance normalization (IN) operation.

[0072]

[0073] Where Conv(·) represents a standard 1*1 convolution operation, which is used to construct the interaction of information between different heads in multi-head attention; d k Represents the channel dimension of the input data; CSA represents the convolutional self-attention mechanism; SA represents the self-attention mechanism; T represents transpose.

[0074] Finally, the output values ​​of each head are concatenated and linearly projected to form the final output.

[0075] MCSA(Q,K,V)=Concat(head i )W o (6)

[0076] Where W o represents the weight matrix generated by the linear projection operation; MCSA represents the multi-head convolution self-attention mechanism; Concat represents the connection operation; head i Indicates the detection head.

[0077] After the multi-head convolutional attention mechanism is completed, it is integrated into the CNN model together with the ternary attention mechanism to form a convolutional neural network called Double Attention-guided Convolutional Neural Network (DACNN), and the DACNN is used as the backbone network of the overall multi-category fault diagnosis model. In the constructed DACNN, the two types of attention are applied as two branches to enhance the CNN respectively, ensuring that the bearing fault features extracted by the two attention mechanism enhanced CNN networks do not interfere with each other. Finally, the fault features extracted by the two attention enhanced CNNs are aggregated. This operation ensures that the fault features extracted by each individual attention enhanced CNN retain their specific meaning and ensures that these features reinforce each other rather than weaken each other.

[0078] There is an inherent logical relationship between fault location and fault size severity, which is not reflected in most deep learning models. This embodiment designs a novel tree-inspired hierarchical decision network, such as Figure 6 As shown, the diagnostic logic between the fault location and the fault size is constructed. The tree-inspired hierarchical decision network is designed based on the basic logic of the fault diagnosis task. In this embodiment, a two-layer hierarchical fault diagnosis architecture is deployed, corresponding to the fault type layer and the fault size layer of the bearing. The first layer focuses on determining the fault type of the input sample to obtain the corresponding superclass attribute, while the second layer focuses on determining the fault size of the input sample to obtain the corresponding subclass attribute.

[0079] In order to more clearly understand the distribution of embedded features generated by this model, a two-layer tree-inspired hierarchical decision network was built using two fully connected layers. In this process, the weight information generated by the fully connected layer in the backbone network is inherited by the threshold of the seed node of the tree-inspired hierarchical decision network, and the threshold of the leaf node is further determined by the embedding decision rule between the seed node and the leaf node. The weight of the seed node directly inherits the probability distribution of the pre-trained fully connected layer. This operation ensures that the recognition ability of the subclass is similar to that of the pre-trained hierarchical multi-class fault diagnosis model.

[0080] The bearing vibration signal with known state is input into the multi-category fault diagnosis model for training to complete the construction of the hierarchical multi-category fault diagnosis model.

[0081] S3. Determine the fault location and size of the bearing using a hierarchical multi-class fault diagnosis model.

[0082] The bearing vibration signal in an unknown state is input into the trained multi-category fault diagnosis model to test the health status of the bearing.

[0083] Embodiment 2

[0084] The advantages of the present invention compared with the prior art will be described in detail below in conjunction with this embodiment.

[0085] This embodiment uses two sets of unbalanced bearing data sets (data sets A and B) under a strong noise background to verify the diagnostic effect of the present invention.

[0086] Among them, the fault data of data set A comes from the precision metrology laboratory of the Department of Mechanical Engineering of SantLongowal Institute of Engineering and Technology in India. The vibration signal of the test bench is measured by an accelerometer installed on the top of the bearing seat. All signals are obtained under the acquisition conditions of a sampling frequency of 70000Hz, a shaft speed of 2050rad, and a vertical load of 200N. The obtained signals include 1 group of normal signals, 5 groups of inner ring fault signals, 4 groups of rolling element fault signals, and 5 groups of outer ring fault signals, totaling 15 state signals. The signal of each state contains 150 samples, each sample contains 2048 data points, totaling 2250 samples, and the training-test ratio is 7:3, that is, 1575 training samples and 675 test samples. The data information used in this embodiment is listed in Table 1, and the corresponding Normal, IR1, IR2, IR3, IR4, IR5, BE1, BE2, BE3, BE4, OR1, OR2, OR3, OR4, and OR5 are marked as 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, and 14, respectively.

[0087] Table 1

[0088]

[0089] Dataset B contains three sub-datasets (a, b, and c), where:

[0090] Dataset a: Bearing data from Beijing University of Technology. The test bearing used in the experiment is SKF6010 radially mounted on a 1200rpm speed shaft. In addition, acceleration sensors are installed in the horizontal, vertical and axial directions of the bearing. This chapter uses 10 types of bearing health status data sets collected in the vertical direction for testing. The 10 types of bearing health status data sets include 5 groups of inner ring faults and 5 groups of outer ring fault data. The fault types of the inner and outer rings are all peeling faults. The 5 groups of fault sizes correspond to 1mm, 3mm, 4mm, 5mm and 6mm respectively. During the acquisition process, the sampling frequency was set to 65536Hz, and the sampling time of each sample was 0.25s. Therefore, the signal length of each sample is 16384.

[0091] Dataset b: Case Western Reserve University bearing data, using EDM technology to machine single-point faults with fault diameters of 0.007 inches, 0.014 inches, and 0.021 inches on rolling bearings. Under different speed and load conditions (1730rpm&3HP, 1750rpm&2HP, 1772rpm&1HP, 1797rpm&0HP), the drive end bearing vibration signals of 10 states were collected at a sampling frequency of 12kHz, including 1 group of normal signals, 3 groups of inner ring fault signals, 3 groups of outer ring fault signals, and 3 groups of rolling element fault signals.

[0092] Dataset c: Hunan University bevel gear bearing data, the helical gearbox bearing data comes from the fault experiment team of Hunan University, all signals are collected under the conditions of a speed of 900r / min, a vertical load of 6N and a sampling frequency of 8192Hz. The inner ring, outer ring and cage faults of the bevel gear bearing have different fault sizes (2mm, 4mm and 6mm), and 9 sets of fault data are collected for the inner ring, outer ring and cage faults of the three sizes of bearings, plus a set of bearing health status data, a total of 10 sets of data are used for the bearing fault classification task.

[0093] In the three experimental data subsets, there are 10 categories of health status in each dataset. The specific labels of each category are shown in Table 2.

[0094] Table 2

[0095]

[0096] The following examples illustrate the meaning of the abbreviations of each data set in Table 2: (a) For data set A, IR1 represents the inner ring fault sample with a fault size of 1 mm; OR6 represents the outer ring fault sample with a fault size of 6 mm. (b) For data set B, RO07 represents the rolling element fault sample with a fault size of 0.007 inches; IR14 represents the inner ring fault sample with a fault size of 0.014 inches; OR21 represents the outer ring fault sample with a fault size of 0.0021 inches. (c) For data set C, CA02 represents the cage fault sample with a fault size of 2 mm; IR04 represents the inner ring fault sample with a fault size of 4 mm; OR06 represents the outer ring fault sample with a fault size of 6 mm; "Health" represents the bearing sample in a healthy state.

[0097] For each of the three data subsets mentioned above, each sample contains 2048 data points, each category of samples contains 1000 samples, and a total of 10000 samples are used for training and testing.

[0098] In order to carry out bearing fault diagnosis experiments with different URs (unbalance ratios) under strong noise, this embodiment adds noises with different signal-to-noise ratios (SNRs) to noise-free samples (the data length of a single sample is 2048) as needed, where the calculation formula of SNR is as follows:

[0099]

[0100] Where P signal Represents the power of the original signal; P noise Represents the power of the noise added to the original signal.

[0101] Table 3

[0102]

[0103] The experimental sample information of bearings with different UR under different noise environments is shown in Table 3.

[0104] After the dataset is constructed, parameters are set.

[0105] The experimental results are significantly affected by the batch size and the weight parameter r. Relevant experiments were conducted on the dataset B with UR = 2:1 and SNR = 8 to select these two key parameters. In this experiment, the cross entropy loss function was used to measure the error between the real sample and the predicted sample, and the stochastic gradient descent algorithm was used to update the model gradient change.

[0106] The results show that when r is the same, when Batchsize exceeds 16, the overall recognition rate of the proposed model will be significantly reduced. When Batchsize is set to 8 or 16, the recognition rate of the proposed model achieved relatively good results, especially when Batchsize is 16, the proposed model showed the best recognition rate. Compared with Batchsize, the effect of parameter r on the experimental results is relatively small but equally significant. The experimental results show that when Batchsize = 16, the proposed model can obtain the best overall recognition rate of 98.73% when r = 0.2. When r = 0.2 and Batchsize = 16, only a small number of rolling element fault samples with a fault size of 1.1mm and inner ring fault samples with a fault size of 1.01mm in the model of the present invention are misclassified as rolling element fault samples with a fault size of 2.03mm and healthy samples without bearing faults. Based on the above analysis, Batchsize is set to 16 and parameter r is set to 0.2 in subsequent experiments.

[0107] In this example, the effectiveness of the proposed DATGDN was first evaluated using two sets of unbalanced bearing data sets under different noise environments. Subsequently, the method of the present invention was compared with other advanced methods to further verify the superiority of the proposed model.

[0108] like Figure 7 As shown in (a), the fault recognition rate of the model of the present invention continues to decrease with the increase of UR in the two groups of bearing experimental data, and steadily increases with the increase of SNR value. This performance is consistent with the actual industrial performance, the objective law, and the expectation. Figure 7 As shown in (b), the model of the present invention achieved a recognition rate ranging from a minimum of 92.21% to a maximum of 99.45% under various conditions, which indicates that the proposed model can effectively extract different distribution features of the original signal for diagnosis in various interference environments, and the diagnosis results basically meet the needs of practical engineering applications.

[0109] In addition, the model of the present invention not only has a high recognition rate, but also has interpretability in the reasoning and decision-making process. Taking two samples (one outer ring fault sample and one inner ring fault sample) in data set B under UR = 10:1 as an example, the hierarchical classification decision process of the outer ring fault sample and the inner ring fault sample is as follows: Figure 8 As shown. Figure 8 (a) It can be seen that: a sample with an outer ring fault size of 4mm is first located as an outer ring fault with a probability of 96.82%, and then the fault size is further determined to be 4mm with a probability of 95.26%. Figure 8 (b) It can be seen that an inner race fault sample with a fault size of 4 mm was initially located as an inner race fault with a probability of 98.65%, and then the fault size was further determined to be 4 mm with a probability of 96.77%.

[0110] In order to demonstrate the superiority of the model of the present invention, this embodiment sets the following algorithm to compare with the proposed method.

[0111] (1) DT: a basic shallow machine learning algorithm. The DT used in this embodiment is the CART algorithm in scikit-learn.

[0112] (2) CNN: The CNN constructed in this embodiment is a 1D-CNN including three convolutional layers for extracting bearing fault features, followed by a fully connected layer with a softmax layer for bearing fault classification.

[0113] (3) CNN-DT: CNN-DT is a combination of (1) and (2).

[0114] (4) DCTN: A network model consisting of a 1D-CNN backbone network with three convolutional layers followed by a tree-inspired hierarchical decision network.

[0115] (5) CBAM-CNN: A network model formed by embedding a CBAM module into the third convolutional layer of a 1D-CNN model with three convolutional layers.

[0116] (6)TA-CNN: A network model formed by embedding a ternary attention module into the third convolutional layer of a 1D-CNN model with three convolutional layers.

[0117] (7)TATN: A network model consisting of a backbone network TA-CNN with three convolutional layers followed by a tree-inspired hierarchical decision network.

[0118] After testing, the following conclusions were drawn:

[0119] (1) Overall, the proposed model is consistently superior to other comparison methods under various URs and SNRs, highlighting its obvious advantages in bearing fault diagnosis.

[0120] (2) The diagnostic results of DT (<60%) and CNN (>60%) on the two sets of data sets clearly demonstrate that deep learning algorithms are superior to shallow machine learning in bearing fault diagnosis.

[0121] (3) The diagnosis results of CNN, CBAM-CNN and TA-CNN show that the attention mechanism can further improve the performance of CNN in bearing fault diagnosis tasks. In addition, since the ternary attention is an extended model of CBAM, the overall performance of TA-CNN is better than that of CBAM-CNN.

[0122] (4) The diagnostic results of DCTN are significantly improved compared with those of CNN, which highlights that the tree-inspired hierarchical decision network constructed by the deep learning network significantly enhances the feature extraction and state discrimination process of CNN.

[0123] (5) Compared with TATN, the proposed model has more excellent diagnostic performance on the noisy unbalanced bearing dataset. This is because DATGDN combines dual attention on the basis of TATN, which effectively enhances the feature extraction ability of the model.

[0124] Embodiment 3

[0125] This embodiment also provides a bearing fault diagnosis system based on a dual attention mechanism to enhance the hierarchical decision network, including: an acquisition module, a construction module and a detection module; the acquisition module is used to collect bearing vibration signals under different health states; the construction module is used to construct a hierarchical multi-category fault diagnosis model based on the bearing vibration signal; the detection module is used to use the hierarchical multi-category fault diagnosis model to determine the fault location and size of the bearing.

[0126] The hierarchical multi-class fault diagnosis model includes a dual attention guidance mechanism and a tree-inspired hierarchical decision network; the dual attention guidance mechanism is used to enhance the information closely related to the fault information in the bearing fault signal and weaken the interference information that is not closely related to the fault information; the tree-inspired hierarchical decision network is used to make decisions on the location and size of the bearing fault step by step. The workflow of the construction module includes: integrating the ternary attention mechanism and the multi-head convolutional attention mechanism into a CNN model to form a dual attention guidance mechanism; among them, the ternary attention mechanism introduces the convolutional block attention module through the concept of cross-dimensional interaction, making the interaction between channels and spatial dimensions closer and more comprehensive; constructing a multi-head convolutional self-attention mechanism based on the single-head convolutional attention.

[0127] This embodiment designs a tree-inspired hierarchical decision network based on the basic logic of the fault diagnosis task, including: a fault type layer and a fault size layer; the fault type layer is used to determine the fault type of the input sample; the fault size layer is used to determine the fault size of the input sample. The dual attention guidance mechanism is used as the backbone network of the hierarchical multi-category fault diagnosis model; at the same time, a tree-inspired hierarchical decision network with a two-layer structure is built using two fully connected layers, and the weight information generated by the fully connected layer in the backbone network is inherited by the threshold of the seed node of the tree-inspired hierarchical decision network, and the threshold of the leaf node is further determined by the embedded decision rule between the seed node and the leaf node.

[0128] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A bearing fault diagnosis method based on a dual attention mechanism to strengthen the hierarchical decision network, characterized in that the steps include: Collect bearing vibration signals under different health conditions; Based on the bearing vibration signal, a hierarchical multi-category fault diagnosis model is constructed; The hierarchical multi-class fault diagnosis model is used to determine the fault location and size of the bearing.

2. The bearing fault diagnosis method based on the dual attention mechanism enhanced hierarchical decision network according to claim 1 is characterized in that: The hierarchical multi-class fault diagnosis model includes a dual attention guidance mechanism and a tree-inspired hierarchical decision network; The dual attention guidance mechanism is used to enhance the information in the bearing fault signal that is closely related to the fault information and weaken the interference information that is not closely related to the fault information; The tree-inspired hierarchical decision network is used to determine the location and size of bearing faults step by step.

3. The bearing fault diagnosis method based on the dual attention mechanism enhanced hierarchical decision network according to claim 2 is characterized in that: The method for constructing the dual attention guidance mechanism includes: integrating the ternary attention mechanism and the multi-head convolutional attention mechanism into a CNN model to form the dual attention guidance mechanism; Among them, the ternary attention mechanism introduces the convolutional block attention module through the concept of cross-dimensional interaction, making the interaction between channels and spatial dimensions more compact and comprehensive; Construct a multi-head convolution self-attention mechanism based on single-head convolution attention.

4. The bearing fault diagnosis method based on dual attention mechanism enhanced hierarchical decision network according to claim 2 is characterized in that: The tree-inspired hierarchical decision network is designed based on the basic logic of the fault diagnosis task, including: a fault type layer and a fault size layer; The fault type layer is used to determine the fault type of the input sample; The fault size layer is used to determine the fault size of input samples.

5. The bearing fault diagnosis method based on dual attention mechanism enhanced hierarchical decision network according to claim 2 is characterized in that: The dual attention guidance mechanism is used as the backbone network of the hierarchical multi-category fault diagnosis model; at the same time, a tree-inspired hierarchical decision network with a two-layer structure is constructed using two fully connected layers. The weight information generated by the fully connected layer in the backbone network is inherited by the threshold of the seed node of the tree-inspired hierarchical decision network, and the threshold of the leaf node is further determined by the embedded decision rule of the seed node and the leaf node.

6. A bearing fault diagnosis system based on a dual attention mechanism to strengthen a hierarchical decision network, the system being used to implement the method described in any one of claims 1 to 5, characterized in that: include: Acquisition module, construction module and detection module; The acquisition module is used to collect bearing vibration signals in different health states; The building module is used to build a hierarchical multi-category fault diagnosis model based on the bearing vibration signal; The detection module is used to determine the fault location and size of the bearing using the hierarchical multi-category fault diagnosis model.

7. The bearing fault diagnosis system based on the dual attention mechanism enhanced hierarchical decision network according to claim 6 is characterized in that: The hierarchical multi-class fault diagnosis model includes a dual attention guidance mechanism and a tree-inspired hierarchical decision network; The dual attention guidance mechanism is used to enhance the information in the bearing fault signal that is closely related to the fault information and weaken the interference information that is not closely related to the fault information; The tree-inspired hierarchical decision network is used to determine the location and size of bearing faults step by step.

8. The bearing fault diagnosis system based on the dual attention mechanism enhanced hierarchical decision network according to claim 7 is characterized in that: The workflow of the building module includes: integrating the ternary attention mechanism and the multi-head convolutional attention mechanism into a CNN model to form the dual attention guidance mechanism; Among them, the ternary attention mechanism introduces the convolutional block attention module through the concept of cross-dimensional interaction, making the interaction between channels and spatial dimensions more compact and comprehensive; Construct a multi-head convolution self-attention mechanism based on single-head convolution attention.

Citation Information

Patent Citations

  • Bearing fault diagnosis method

    CN116861343A

  • Psychological state prediction method and system based on house tree sketch, and storage medium

    CN117372786A

  • Multistage intelligent bearing fault diagnosis method based on TARTDN

    CN117390554A