A bearing fault feature extraction method, system, medium and device

Abstract

The present disclosure provides a bearing fault feature extraction method, which introduces a Gauss-Bernoulli restricted Boltzmann machine model, solves the problem that the input vector of the traditional restricted Boltzmann machine is restricted by the Bernoulli binary distribution and has poor reconstruction and fitting effect for non-binomial distributed data; uses the cosine loss function as the loss function, retains the advantage of the Softmax loss function in expanding the inter-class difference, and reduces the sensitivity to different signal strengths; combines the attention mechanism adaptively to more effectively extract features that are effective in describing the bearing status.

Description

Technical Field

[0001] The present disclosure belongs to the technical field of feature extraction, and in particular relates to a bearing fault feature extraction method based on an improved deep belief network. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] With the development of machine learning, fault feature extraction methods based on machine learning models have become a research hotspot, such as support vector machine (SVM), k-nearest neighbor (KNN), fuzzy reasoning, etc. However, in the case of high-dimensional big data, when shallow learning models are applied to gearbox fault diagnosis, fault diagnosis lacks diagnostic generalization ability, and its accuracy depends on the extraction quality of fault features in big data. As a new method in the field of machine learning.

[0004] When a rolling bearing fails, it generates unsteady vibration signal characteristics. Furthermore, time domain and frequency domain features are not suitable as bearing fault diagnosis feature parameters due to their insensitivity to non-stationary signals and unstable feature performance. Traditional feature extraction cannot effectively extract hidden bearing fault features, resulting in a low fault recognition rate. Since rolling bearings can fail at different locations, the degree of confusion in the vibration signals caused by different faults varies. Current feature extraction extracts feature values ​​from the time and frequency domains without careful screening and in-depth mining. After obtaining a comprehensive feature set, it is crucial to mine the inherent information to accurately identify bearing faults under different loads. Some scholars have increased the number of deep learning model layers and used large-scale parameter training to learn bearing fault features, which has also led to problems such as long training time and difficulty in convergence. Summary of the Invention

[0005] In order to solve the above problems, the present disclosure proposes a bearing fault feature extraction method, system, medium and equipment, which utilizes an improved deep belief network integrated with an attention mechanism to extract bearing features, introduces a Gaussian-Bernoulli restricted Boltzmann machine model, and combines the attention mechanism to adaptively extract features that are effective in describing the bearing status.

[0006] According to some embodiments, the present disclosure adopts the following technical solutions:

[0007] A bearing fault feature extraction method, the specific training process includes the following steps:

[0008] Step 1: Collect vibration signals from different positions of the rolling bearing, use the vibration signals as input data, and divide the data into a training data set and a test data set;

[0009] Step 2: Extract the time domain characteristic indicators and frequency domain characteristic indicators of the input data;

[0010] Step 3: Deeply mine the two dimensional features of the time domain feature and the frequency domain feature to form two deep feature sets;

[0011] Step 4: Adaptively and dynamically weight the two deep feature sets using the attention mechanism to form a feature set, which is input into the fully connected layer to perform preliminary fault identification on the bearing;

[0012] Step 5: Use the improved loss function to train the model to achieve accurate fault identification of bearings in the training dataset;

[0013] Step 6: Process the vibration signals in the test data set according to steps 2 to 4, input them into the model for identification, and output the bearing fault identification results.

[0014] According to other embodiments, the present disclosure adopts the following technical solutions:

[0015] A bearing fault feature extraction system, comprising:

[0016] Signal acquisition sensor: used to collect vibration signals at different positions of the rolling bearing;

[0017] Data processing module: divides the collected vibration signals into training data sets and test data sets;

[0018] Feature extraction module: used to extract time domain feature indicators and frequency domain feature indicators of input data;

[0019] Feature recognition module: This module is used to adaptively and dynamically weight two deep feature sets using the attention mechanism to form a feature set, which is then input into the fully connected layer to perform preliminary fault identification on the bearing.

[0020] Model training module: used to train the model to achieve accurate fault identification of bearings.

[0021] According to other embodiments, the present disclosure also adopts the following technical solutions:

[0022] A computer-readable storage medium is used to store a computer program, which can run the steps of a bearing fault feature extraction method when executed.

[0023] An electronic device comprises a processor, a memory and a computer program stored in the memory. When the device is running, the processor can run the computer program stored in the memory to execute the steps of a bearing fault feature extraction method.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] (1) Based on the use of time domain and frequency domain features, this method uses the attention mechanism to fuse deep features, effectively extracts effective features that affect the fault and improves the accuracy of bearing fault diagnosis under different loads.

[0026] (2) The problem of poor fitting effect of traditional data reconstruction is solved by introducing the Gauss-Bernoulli restricted Boltzmann machine model.

[0027] (3) By improving the loss function of the DBN model, the model's sensitivity to the amplitude value of the fault signal is reduced, thereby reducing the network's overfitting burden and improving the generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings, which constitute a part of the present disclosure, are used to provide a further understanding of the present disclosure. The exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation to the present disclosure.

[0029] Figure 1 This is the training flow chart of the attention combined with deep belief network model disclosed in this paper;

[0030] Figure 2 Schematic diagram of the cosine loss function disclosed herein;

[0031] Figure 3 This is a flowchart of bearing fault feature extraction using an improved deep belief network that integrates an attention mechanism. DETAILED DESCRIPTION

[0032] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0033] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present disclosure belongs.

[0034] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0035] Those skilled in the art will appreciate that embodiments of the present disclosure may be provided as methods, systems, or computer program products. Thus, the present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0036] Example 1

[0037] The present disclosure provides a method for extracting bearing fault features. Figure 3 As shown in the figure, the specific recognition training process is as follows:

[0038] Step 1: Collect vibration signals from different positions of the rolling bearing, use the vibration signals as input data, and divide the data into a training data set and a test data set;

[0039] Step 2: Extract the time domain characteristic indicators and frequency domain characteristic indicators of the input data;

[0040] The extracted time domain feature indicators are classified according to whether they have dimensions or not. Dimensional indicators include mean, root mean square value, kurtosis, etc. Dimensionless indicators include wave index, margin index, etc., totaling 11 features.

[0041] Then, the frequency domain features are extracted through fast Fourier transform. The frequency domain feature indicators can reflect the changes in spectrum distribution, vibration energy and frequency band position. Common feature indicators include center of gravity frequency, frequency variance and mean square frequency, totaling 13 frequency domain features.

[0042] Step 3: Deeply mine the two dimensional features of the time domain feature and the frequency domain feature to form two deep feature sets;

[0043] Specifically, the improved DBN greedy learning method is used to deeply mine the two dimensional features respectively to form two deep feature sets P1 and P2.

[0044] Step 4: Adaptively and dynamically weight the two deep feature sets using the attention mechanism to form a feature set, which is input into the fully connected layer to perform preliminary fault identification on the bearing;

[0045] Specifically, the attention mechanism is used to adaptively and dynamically weight P1 and P2 to form a feature set P, which is then input into the fully connected layer to complete the bearing fault identification.

[0046] Step 5: Use the improved loss function to train the model to achieve accurate fault identification of bearings in the training dataset;

[0047] Step 6: Process the vibration signals in the test data set according to steps 2 to 4, input them into the model for identification, and output the bearing fault identification results.

[0048] This paper proposes a bearing fault feature extraction method using an improved deep belief network (DBN) that incorporates an attention mechanism. This method incorporates a Gaussian-Bernoulli restricted Boltzmann machine (RBM) model, addressing the problem that traditional RBMs, whose input vectors are restricted to a Bernoulli binary distribution and exhibit poor reconstruction and fitting performance for non-binomial data, utilizes a cosine loss function, retaining the Softmax loss function's advantage of amplifying inter-class variance while reducing sensitivity to varying signal strengths. Furthermore, an attention mechanism is incorporated to adaptively prioritize features that effectively describe bearing condition. Furthermore, the method utilizes both the time and frequency domains to facilitate fault diagnosis. Experimental results demonstrate that this method effectively improves the model's adaptive feature extraction capabilities and fault diagnosis accuracy, demonstrating good generalization.

[0049] GB-RBM:

[0050] A DBN is composed of multiple stacked restricted Boltzmann machines (RBMs). An RBM is an undirected probabilistic graphical model with a bipartite graph structure, consisting of a visible layer v consisting of m neurons and a hidden layer h consisting of n neurons. The nodes in each layer are independent and disconnected from each other.

[0051] Since the visible layer and hidden layer of the traditional restricted Boltzmann machine are both binary units, that is, v i and h j The values ​​of are all 0 or 1, and the vibration signal is a set of continuous time series data. The binary unit model is not conducive to the operation of continuous data.

[0052] Therefore, the present disclosure introduces continuous values ​​of independent Gaussian distribution into the traditional RBM model to process continuous data, and limits the output value of the structural unit of the RBM to a continuous value between 0 and 1.

[0053] The Gaussian Bernoulli Restricted Boltzmann Machine (GB-RBM) was proposed by Krizhevsky and Hinton. GB-RBM assumes that the visible layer variables are Gaussian distributed and the hidden layer variables are Bernoulli distributed. Its energy function is defined as:

[0054] (1)

[0055] In formula (1): a iis the visible unit bias; b j is the bias of the hidden unit; σ is the standard deviation of the Gaussian noise of the visible unit; w ij is the connection weight between the visible unit and the hidden unit.

[0056] Among them, the conditional probabilities of the visible layer and the hidden layer are:

[0057]

[0058] N(μ,σ i ) is a Gaussian function with mean μ and standard deviation σ.

[0059] Like traditional RBMs, GB-RBMs are pre-trained using the contrastive divergence (CD) algorithm. This algorithm uses the KL distance between the estimated and true probability distributions as a metric and performs k-step Gibbs sampling on each batch of training samples to generate samples with the highest probability.

[0060] The stochastic gradient ascent method is used to find the maximum value of the likelihood function, and the GB-RBM network parameter update rule is obtained as follows:

[0061]

[0062] Where: η is the learning rate; data and<·> model are the expected value of the real data and the network output value respectively

[0063] After the RBM layer-by-layer unsupervised pre-training is complete, the learned model parameters are used as the initial values ​​for supervised learning, providing prior knowledge of the input data. During the global fine-tuning phase, the categories determined by the Softmax classifier are compared with the category information of the training data, and recognition errors are counted. The DBN network is then trained in reverse using the backpropagation algorithm to fine-tune the initial connection weights.

[0064] This paper introduces a Gaussian-Bernoulli restricted Boltzmann machine model, which solves the problem that the input vector of the traditional restricted Boltzmann machine is limited to the Bernoulli binary distribution and has poor reconstruction and fitting effect for non-binomial distributed data. It also uses the greedy learning advantages and data feature highlighting capabilities of the RBM structure in DBN to deeply mine the features in the time domain and frequency domain, respectively, and learn the deep features that are effective for bearing faults.

[0065] Fusion of Attention Mechanism

[0066] After obtaining multi-dimensional deep feature maps, these features need to be combined to obtain more comprehensive fault information. Currently, some researchers often use a serial connection method when combining multi-network outputs and multi-domain features. This method ignores the fact that different features contribute to the different degrees of fault diagnosis results.

[0067] Therefore, the present invention adds an attention mechanism after DBN deep mining. It imitates humans when observing different things, automatically shifting their attention to the parts they are interested in, and reducing their attention to irrelevant information, thereby improving the ability to obtain information.

[0068] The attention mechanism is used to combine the time domain and frequency domain features obtained after DBN to form the global feature of bearing fault diagnosis.

[0069]

[0070] Among them, hi are the depth features of time domain and frequency domain respectively; a i are their corresponding importance weights; c is the global feature obtained by attention.

[0071] In step 5, it is proposed to use the improved loss function to train the model to achieve accurate fault identification of bearings in the training data set. The improvement process of the loss function is as follows:

[0072] The cosine loss function is based on the Softmax loss function, retaining its advantage of expanding the difference between classes, but reducing its sensitivity to different signal strengths, and paying more attention to the difference in the direction of the vector, such as Figure 2 shown.

[0073] For the same fault signals q1, q2, the fault label is p1. When Softmax is used as the loss function, the final loss function Ls is shown in formula (10).

[0074]

[0075] Where B and N are the number of training samples and fault categories in a batch, respectively; x is the hidden layer output; W is the weight matrix; and θ is the angle between them. Equation (10) shows that the Softmax loss is related to the signal amplitude. Using 1-cosine similarity as the loss function, the final loss function Lc is shown in Equation (11).

[0076]

[0077] From formula (11), we can see that ‖x‖ 2It can be eliminated, so the cosine loss is independent of the signal amplitude. Therefore, using the cosine loss as the loss function for gearbox fault diagnosis can convert the loss from Euclidean space to angular space, thereby eliminating the influence of signal amplitude and reducing the network fitting burden.

[0078] Example 2

[0079] The present invention discloses a bearing fault feature extraction method with the following specific processes:

[0080] Vibration signals at different positions of the rolling bearing are collected and acquired. Let the vibration signal be s(t). The time domain and frequency domain features of s(t) are extracted to form a feature set, which is then input into the improved DBN respectively. After combining the deep features of each dimension using the attention mechanism, they are mapped to different fault types.

[0081] Step 1: Take the vibration signal as input and extract time domain feature indicators. Time domain feature indicators can be divided into two categories based on whether they are dimensioned or not. These include dimensioned indicators such as mean, root mean square value, and kurtosis, and dimensionless indicators such as wave characteristics and margin indicators, totaling 11 features. Frequency domain features are then extracted using fast Fourier transform. Frequency domain feature indicators can reflect the spectral distribution, vibration energy, and changes in frequency band position. Common feature indicators include center of gravity frequency, frequency variance, and mean square frequency, totaling 13 frequency domain features.

[0082] Step 2: Use DBN’s greedy learning method to deeply mine the two dimensional features to form two deep feature sets P1 and P2;

[0083] Step 3: Use the attention mechanism to adaptively and dynamically weight P1 and P2 to form a feature set P, and input it into the fully connected layer to complete the bearing fault identification;

[0084] Step 4: Use the improved loss function to train the model and accurately identify bearings in the training set.

[0085] Step 5: Process the vibration signals in the test set according to steps 1 to 3 to identify the bearing status and finally output the bearing fault identification result.

[0086] Example 3

[0087] The present disclosure provides a bearing fault feature extraction system, which is characterized by comprising:

[0088] Signal acquisition sensor: used to collect vibration signals at different positions of the rolling bearing;

[0089] Data processing module: divides the collected vibration signals into training data sets and test data sets;

[0090] Feature extraction module: used to extract time domain feature indicators and frequency domain feature indicators of input data;

[0091] Feature recognition module: This module is used to adaptively and dynamically weight two deep feature sets using the attention mechanism to form a feature set, which is then input into the fully connected layer to perform preliminary fault identification on the bearing.

[0092] Model training module: used to train the model to achieve accurate fault identification of bearings.

[0093] The modules involved in the above system perform the following method steps:

[0094] Step 1: Collect vibration signals from different positions of the rolling bearing, use the vibration signals as input data, and divide the data into a training data set and a test data set;

[0095] Step 2: Extract the time domain characteristic indicators and frequency domain characteristic indicators of the input data;

[0096] The extracted time domain feature indicators are classified according to whether they have dimensions or not. Dimensional indicators include mean, root mean square value, kurtosis, etc. Dimensionless indicators include wave index, margin index, etc., totaling 11 features.

[0097] Then, the frequency domain features are extracted through fast Fourier transform. The frequency domain feature indicators can reflect the changes in spectrum distribution, vibration energy and frequency band position. Common feature indicators include center of gravity frequency, frequency variance and mean square frequency, totaling 13 frequency domain features.

[0098] Step 3: Deeply mine the two dimensional features of the time domain feature and the frequency domain feature to form two deep feature sets;

[0099] Specifically, the improved DBN greedy learning method is used to deeply mine the two dimensional features respectively to form two deep feature sets P1 and P2.

[0100] Step 4: Adaptively and dynamically weight the two deep feature sets using the attention mechanism to form a feature set, which is input into the fully connected layer to perform preliminary fault identification on the bearing;

[0101] Specifically, the attention mechanism is used to adaptively and dynamically weight P1 and P2 to form a feature set P, which is then input into the fully connected layer to complete the bearing fault identification.

[0102] Step 5: Use the improved loss function to train the model to achieve accurate fault identification of bearings in the training dataset;

[0103] Step 6: Process the vibration signals in the test data set according to steps 2 to 4, input them into the model for identification, and output the bearing fault identification results.

[0104] Example 4

[0105] This embodiment provides a computer-readable storage medium for storing a computer program, which, when executed, can execute the steps of the above-mentioned bearing fault feature extraction method.

[0106] Example 5

[0107] This embodiment provides an electronic device, including a processor, a memory, and a computer program stored in the memory. When the device is running, the processor can run the computer program stored in the memory to execute the steps of the above-mentioned bearing fault feature extraction method.

[0108] This invention discloses a bearing fault feature extraction method using an improved DBN that incorporates an attention mechanism. Given the high noise levels in rolling bearings, conventional feature extraction methods hinder fault extraction and classification. This method proposes a bearing fault feature extraction method using an improved deep belief network (DBN) that incorporates an attention mechanism. This method incorporates a Gaussian-Bernoulli restricted Boltzmann machine (GBM) model to address the problem that the input vector of a conventional restricted Boltzmann machine is restricted to a Bernoulli binary distribution and exhibits poor reconstruction and fitting performance for non-binomial data. It utilizes a cosine loss function, retaining the Softmax loss function's advantage of amplifying inter-class differences while reducing sensitivity to varying signal strengths. Furthermore, the attention mechanism is incorporated to adaptively prioritize features that effectively describe the bearing condition. Furthermore, the method utilizes both the time and frequency domains to facilitate fault diagnosis. Experimental results demonstrate that this method effectively improves the model's adaptive feature extraction capabilities and fault diagnosis accuracy, demonstrating good generalization.

[0109] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0110] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0111] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0112] The foregoing description is merely a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present disclosure shall be included within the scope of protection of the present disclosure.

[0113] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Those skilled in the art should understand that, based on the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A bearing fault feature extraction method, characterized in that: The specific training process includes the following steps: Step 1: Collect vibration signals from different positions of the rolling bearing, use the vibration signals as input data, and divide the data into a training data set and a test data set; Step 2: Extract the time domain characteristic indicators and frequency domain characteristic indicators of the input data; Step 3: Deeply mine the two dimensional features of the time domain feature and the frequency domain feature to form two deep feature sets; The improved DBN greedy learning method is used to deeply mine the two dimensional features to form two deep feature sets P1 and P2. The attention mechanism is used to adaptively and dynamically weight P1 and P2 to form a feature set P, which is then input into the fully connected layer to complete the bearing fault identification. DBN is formed by stacking multiple restricted Boltzmann machines (RBMs). Independent Gaussian distributed continuous values ​​are introduced into the RBM model to process continuous data, and the output values ​​of the RBM structural units are limited to continuous values ​​between 0 and 1. GB-RBM assumes that the visible layer variables are Gaussian distributed and the hidden layer variables are Bernoulli distributed. Its energy function is defined as: a i is the visible unit bias; b j is the bias of the hidden unit; σ is the standard deviation of the Gaussian noise of the visible unit; w ij is the connection weight between the visible unit and the hidden unit; Step 4: Adaptively and dynamically weight the two deep feature sets using the attention mechanism to form a feature set, which is input into the fully connected layer to perform preliminary fault identification on the bearing; The attention mechanism is used to combine the time domain and frequency domain features obtained after DBN to form the global features of bearing fault diagnosis, which is calculated as: Among them, hi are the depth features of time domain and frequency domain respectively; a i are their corresponding importance weights; c is the global feature obtained by attention; Step 5: Use the improved loss function to train the model to achieve accurate fault identification of bearings in the training dataset; The loss function is improved as follows: For the same fault signals q1, q2, the fault label is p1. When Softmax is used as the loss function, the final loss function Ls is as follows: N is the number of training samples and fault categories in a batch respectively; x is the hidden layer output; W is the weight matrix; θ is the angle between them; Step 6: Process the vibration signals in the test data set according to steps 2 to 4, input them into the model for identification, and output the bearing fault identification results.

2. A bearing fault feature extraction method according to claim 1, characterized in that: The extracted time domain characteristic indicators are classified according to whether they have dimensions or not. Dimensional indicators include mean, root mean square value, and kurtosis, while dimensionless indicators include wave index and margin index.

3. A bearing fault feature extraction method according to claim 1, characterized in that: The frequency domain characteristic indicators are extracted by fast Fourier transform. The frequency domain characteristic indicators reflect the distribution of the spectrum, the magnitude of the vibration energy and the change of the frequency band position, including the center of gravity frequency, frequency variance and mean square frequency.

4. A bearing fault feature extraction system, characterized in that: include: Signal acquisition sensor: used to collect vibration signals at different positions of the rolling bearing; Data processing module: divides the collected vibration signals into training data sets and test data sets; Feature extraction module: used to extract time domain feature indicators and frequency domain feature indicators of input data; The improved DBN greedy learning method is used to deeply mine the two dimensional features to form two deep feature sets P1 and P2. The attention mechanism is used to adaptively and dynamically weight P1 and P2 to form a feature set P, which is then input into the fully connected layer to complete the bearing fault identification. DBN is formed by stacking multiple restricted Boltzmann machines (RBMs). Independent Gaussian distributed continuous values ​​are introduced into the RBM model to process continuous data, and the output values ​​of the RBM structural units are limited to continuous values ​​between 0 and 1. GB-RBM assumes that the visible layer variables are Gaussian distributed and the hidden layer variables are Bernoulli distributed. Its energy function is defined as: a i is the visible unit bias; b j is the bias of the hidden unit; σ is the standard deviation of the Gaussian noise of the visible unit; w ij is the connection weight between the visible unit and the hidden unit; Feature recognition module: This module is used to adaptively and dynamically weight two deep feature sets using the attention mechanism to form a feature set, which is then input into the fully connected layer to perform preliminary fault identification on the bearing. The attention mechanism is used to combine the time domain and frequency domain features obtained after DBN to form the global features of bearing fault diagnosis, which is calculated as: Among them, hi are the depth features of time domain and frequency domain respectively; a i are their corresponding importance weights; c is the global feature obtained by attention; Model training module: used to train the model to achieve accurate fault identification of bearings; The loss function is improved as follows: For the same fault signals q1, q2, the fault label is p1. When Softmax is used as the loss function, the final loss function Ls is as follows: N is the number of training samples and fault categories in a batch; x is the hidden layer output; W is the weight matrix; θ is the angle between them.

5. A computer-readable storage medium for storing a computer program, which, when executed, is capable of executing the steps of any one of the bearing fault feature extraction methods according to claims 1 to 3.

6. An electronic device comprising a processor, a memory and a computer program stored in the memory. When the device is running, the processor is capable of running the computer program stored in the memory to execute the steps of any one of the bearing fault feature extraction methods described in claims 1-3.

Images