Bearing fault diagnosis method based on neural network

Abstract

The invention provides a bearing fault diagnosis method based on a neural network, and the method comprises the steps: obtaining bearing fault signals of a bearing under the conditions of different rotating speeds, loads and noises based on an adaptive sensor array, carrying out the preprocessing of the bearing fault signals, and extracting time domain features; wherein the bearing fault signal comprises a vibration signal and an acoustic emission signal; decomposing the preprocessed bearing fault signal by adopting wavelet transform to obtain a frequency band component; based on Fourier transform, extracting a frequency domain feature of each frequency band component, and constructing a multi-dimensional feature vector in combination with a time domain feature; training a deep convolutional neural network model by using the multi-dimensional feature vector to obtain a bearing fault diagnosis model; and based on the bearing fault diagnosis model, diagnosing a to-be-diagnosed bearing fault to obtain a fault diagnosis result. According to the invention, the bearing fault signal can be effectively processed, the accuracy and reliability of fault diagnosis are improved, and meanwhile, high-quality data support is provided for subsequent fault analysis and prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial equipment condition monitoring and fault diagnosis, and particularly relates to a method for bearing fault diagnosis based on a neural network. Background Art

[0002] Traditional bearing fault diagnosis relies on vibration signal spectrum analysis or expert experience, and has the following problems: low signal sensitivity: it is difficult for a single vibration sensor to capture early weak fault features. Dependence on artificial features: time-frequency domain features (such as wavelet coefficients, kurtosis values) need to be manually extracted, with low efficiency and strong subjectivity. Poor generalization ability: traditional machine learning models (such as SVM) are insufficient in adapting to complex working conditions.

[0003] Although existing neural network solutions (such as CNN, LSTM) can automatically extract features, they have the following defects: data redundancy: the fusion method of multi-source sensor data is not optimized, resulting in excessive computational complexity. Loss of time series features: it is difficult for traditional convolutional networks to capture the long-period time series correlation of fault signals. Sample imbalance: in the actual scenario, the number of normal samples is much larger than that of fault samples, resulting in model overfitting.

[0004] How to construct a diagnostic model with strong robustness and good generalization ability that can adapt to complex and changeable working conditions and different types of bearings is a technical problem to be solved urgently. Summary of the Invention

[0005] The present invention provides a method for bearing fault diagnosis based on a neural network, which solves the problems of low signal processing efficiency and poor model generalization ability in the prior art.

[0006] A method for bearing fault diagnosis based on a neural network, the method comprising:

[0007] Obtaining bearing fault signals under different rotational speeds, loads and noise conditions of a bearing based on an adaptive sensor array, and preprocessing the bearing fault signals to extract time domain features; wherein, the bearing fault signals include vibration signals and acoustic emission signals;

[0008] Decomposing the preprocessed bearing fault signals by using wavelet transform to obtain frequency band components;

[0009] Extracting frequency domain features of each frequency band component based on Fourier transform, and combining the time domain features to construct a multi-dimensional feature vector;

[0010] Training a deep convolutional neural network model by using the multi-dimensional feature vector to obtain a bearing fault diagnosis model;

[0011] Diagnosing the bearing fault to be diagnosed based on the bearing fault diagnosis model to obtain a fault diagnosis result.

[0012] Preferably, the adaptive sensor array is a circularly expandable sensor array. A plurality of groups of sensor nodes are evenly arranged circumferentially on the bearing housing. Each group of sensor nodes includes a triaxial vibration sensor and an acoustic emission sensor; each of the sensor nodes is connected to the bearing housing through an electromagnetic drive slide rail.

[0013] Preferably, the method for preprocessing the bearing fault signal includes:

[0014] Calculating the cross-correlation coefficient of the bearing fault signals collected by adjacent sensors, and using a preset correlation coefficient threshold to divide the cross-correlation coefficient to obtain homologous signals and heterologous signals;

[0015] Performing weighted averaging on the homologous signals to filter random noise and obtain preprocessed homologous signals;

[0016] Performing adaptive noise reduction on the heterologous signals to obtain preprocessed heterologous signals.

[0017] Preferably, the method for obtaining the frequency band components includes:

[0018] Based on the impact characteristics of the bearing fault signal, establishing candidate wavelet bases;

[0019] Calculating the matching degrees of the candidate wavelet bases and the preprocessed bearing fault signal, and using the wavelet bases that meet the preset matching degree threshold as decomposition basis functions;

[0020] Based on the decomposition basis functions, performing multi-scale decomposition on the bearing fault signal to obtain initial frequency band components;

[0021] Using the frequency band entropy analysis method to screen the initial frequency band components to obtain the final frequency band components.

[0022] Preferably, the time domain features include mean, variance, kurtosis, instantaneous frequency, and envelope spectrum; the frequency domain features include spectral peak, spectral centroid, and spectral entropy; the multi-dimensional feature vector is formed by splicing the time domain features and the frequency domain features according to the frequency band dimension.

[0023] Preferably, the bearing fault diagnosis model includes an input layer, a feature extraction layer, and a classification layer;

[0024] The input layer is used to map the input multi-dimensional feature vector to a high-dimensional space to obtain a high-dimensional feature vector;

[0025] The feature extraction layer is used to capture the frequency band correlation in the high-dimensional feature vector based on parallel convolution kernels, and calculate the weights of each frequency band using a frequency band attention module to obtain bearing fault features;

[0026] A classification layer for classifying the bearing fault features to obtain the fault category and confidence level.

[0027] Preferably, in the frequency band attention module, global average pooling is performed on the high-dimensional feature vector to obtain a frequency band weight vector; based on the frequency band weight vector, an attention coefficient is generated through a fully connected layer and a Sigmoid function; the attention coefficient is weighted to the high-dimensional feature vector to obtain the bearing fault features.

[0028] Preferably, when the confidence level is lower than a preset confidence threshold, a multi-model voting mechanism is triggered, and three bearing fault diagnosis models with different initialization parameters are used for independent prediction, and the majority voting method is adopted to obtain the final diagnosis result.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: By using an adaptive sensor array to acquire the vibration signals and acoustic emission signals of the bearing under different working conditions (speed, load, noise), the fault features can be comprehensively captured. In the preprocessing stage, wavelet transform and Fourier transform are combined to extract time-domain and frequency-domain features respectively, and a multi-dimensional feature vector is constructed, effectively retaining the deep features of the fault signals and improving the robustness of the features. The deep convolutional neural network model trained by the multi-dimensional feature vector can effectively handle the fault diagnosis tasks under complex working conditions, and has high generalization ability and adaptability. This method performs excellently in the tests of various fault types (such as inner race fault, outer race fault, rolling element fault), and both the accuracy rate and the recall rate are significantly better than the traditional methods. Description of the Drawings

[0030] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0031] Figure 1 It is a flowchart of a method for bearing fault diagnosis based on a neural network according to an embodiment of the present invention. Detailed Embodiments

[0032] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0033] To make the above objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Embodiment 1

[0035] As Figure 1 shown, a method for bearing fault diagnosis based on a neural network, the method includes:

[0036] S1: Acquire bearing fault signals under different rotational speeds, loads, and noise conditions based on an adaptive sensor array, and preprocess the bearing fault signals to extract time-domain features; wherein, the bearing fault signals include vibration signals and acoustic emission signals; a further implementation is that the adaptive sensor array is a ring-shaped telescopic sensor array, and several groups of sensor nodes are evenly arranged circumferentially on the bearing housing. Each group of sensor nodes includes a triaxial vibration sensor and an acoustic emission sensor; in this embodiment, a micro temperature sensor is also included. Each sensor node is connected to the bearing housing through an electromagnetic drive slide rail and can be dynamically adjusted in position circumferentially from 0 to 360° and radially from ±10 mm. Activate different sensor combinations according to the rotational speed range (0 - 3000 rpm): Low speed (<500 rpm): Enable all 6 groups of sensors to improve signal resolution; Medium and high speed (≥500 rpm): Turn off 2 groups of sensors at diagonal positions to suppress signal crosstalk caused by centrifugal force. When the load exceeds 80% of the rated value, start the radial pressure compensation mode: 1. Monitor the radial force direction of the bearing in real time through a pressure sensor; 2. Control the slide rail to move the sensor node in the opposite direction of the force (maximum offset 5 mm) to reduce the influence of contact surface deformation on the signal. Calculate the signal-to-noise ratio (SNR) of each sensor signal in real time. If the SNR of a certain node continuously drops below 15 dB, turn off this node and enable the backup node.

[0037] In this embodiment, the sensor node is moved in the opposite direction according to the radial force direction, and the offset calculation formula is:

[0038] Δd = k·(F / F max )

[0039] where k = 5 mm is the maximum offset, F is the real-time load, and F max is the rated load.

[0040] In this embodiment, the ring-shaped telescopic sensor array is collaboratively optimized, specifically:

[0041] Define the state space: including rotational speed, load, temperature, and SNR values of each node;

[0042] s t= [Rotation speed, load, temperature, SNR1, SNR2,..., SNR6],

[0043] Rotation speed: Current bearing rotation speed (unit: rpm); Load: Current bearing radial load (unit: % rated load); Temperature: Bearing housing temperature (unit: °C); SNR i : Signal-to-noise ratio of the i-th sensor node (unit: dB).

[0044] Define the action space: Sensor enable / disable, position offset (0 - 5mm, step size 1mm); Specifically, a t = [Enable / Disable 1, Enable / Disable 2,..., Enable / Disable 6, Δd1, Δd2,..., Δd6],

[0045] Enable / Disable i : Enable (1) or disable (0) status of the i-th sensor node; Ad i : Position offset of the i-th sensor node (unit: mm, range: 0 - 5). The enable / disable status is a binary action; The position offset is discretized into 6 levels (0mm, 1mm, 2mm, 3mm, 4mm, 5mm).

[0046] Reward function: R = ω1 * SNR avg + ω2 * (1 / data redundancy) - ω3 * power consumption,

[0047] where the weight coefficients ω1 = 0.6, ω2 = 0.3, ω3 = 0.1, and the data redundancy is calculated by the signal mutual information entropy. SNR avg : Average signal-to-noise ratio of all currently enabled sensors; Total power consumption of currently enabled sensors (unit: mW).

[0048] Adopt the PPO algorithm to update the policy network online, and perform layout optimization every 10 minutes. Specifically, the update process of the policy network includes:

[0049] In the policy network, input a 9-dimensional state vector, and the hidden layer uses a 2-layer fully connected network (number of neurons = 64, activation function = ReLU); Output the action probability distribution (dimension = 12).

[0050] In the value network, input a 9-dimensional state vector, and the hidden layer uses a 2-layer fully connected network (number of neurons = 64, activation function = ReLU); Output the state value (dimension = 1).

[0051] The algorithm process specifically includes:

[0052] Under the current policy, perform layout adjustment every 10 minutes and collect trajectory data:

[0053] Store the trajectory data in the experience replay buffer (Buffer).

[0054] Calculate the advantage function using Generalized Advantage Estimation (GAE), where the discount factor used is specifically set to 0.99.

[0055] Sample a batch of data from the Buffer, calculate the policy loss and the value loss, and complete the update of the policy network parameters.

[0056] A further implementation manner lies in that the method for preprocessing the bearing fault signal includes:

[0057] Calculate the cross-correlation coefficient of the bearing fault signals collected by adjacent sensors, and use a preset correlation coefficient threshold to divide the cross-correlation coefficient to obtain homologous signals and heterologous signals;

[0058] Perform weighted averaging on the homologous signals to filter out random noise and obtain the preprocessed homologous signals;

[0059] Perform adaptive noise reduction on the heterologous signals to obtain the preprocessed heterologous signals.

[0060] S2: Decompose the preprocessed bearing fault signal using wavelet transform to obtain frequency band components; A further implementation manner lies in that the method for obtaining the frequency band components includes:

[0061] Based on the impact characteristics of the bearing fault signal, establish candidate wavelet bases;

[0062] Calculate the matching degree of each candidate wavelet base with the preprocessed bearing fault signal, and use the wavelet base that meets the preset matching degree threshold as the decomposition basis function;

[0063] Based on the decomposition basis function, perform multi-scale decomposition on the bearing fault signal to obtain initial frequency band components; Specifically, perform 6-layer wavelet packet transform (WPT) on the vibration signal to obtain 16 sub-frequency band signals; Perform continuous wavelet transform (CWT) on the acoustic emission signal, set the scale range to 50 - 1000, and the step size to 10 to generate a time-frequency diagram.

[0064] Use the frequency band entropy analysis method to screen the initial frequency band components to obtain the final frequency band components. Specifically, the frequency band entropy (Band Entropy, BE) is used to measure the information complexity of the frequency band components. Perform short-time Fourier transform (STFT) on each frequency band component to obtain a time-frequency matrix; Average the time-frequency matrix along the time dimension to obtain the frequency domain amplitude spectrum; Calculate the normalized probability distribution of the frequency domain amplitude spectrum to obtain the final frequency band entropy. Calculate the average entropy value of all frequency band components. When the average entropy value is greater than the entropy value of the current frequency band component, it is considered that the frequency band component contains more noise or redundant information and is excluded, otherwise it is retained.

[0065] S3: Based on Fourier transform, extract the frequency-domain features of each frequency band component, and combine with the time-domain features to construct a multi-dimensional feature vector. A further implementation is that the time-domain features include mean, variance, kurtosis, instantaneous frequency, and envelope spectrum; the frequency-domain features include spectral peak, spectral centroid, and spectral entropy; the multi-dimensional feature vector is formed by splicing the time-domain features and the frequency-domain features according to the frequency band dimension.

[0066] S4: Use the multi-dimensional feature vector to train a deep convolutional neural network model to obtain a bearing fault diagnosis model. A further implementation is that the bearing fault diagnosis model includes an input layer, a feature extraction layer, and a classification layer;

[0067] The input layer is used to map the input multi-dimensional feature vector to a high-dimensional space to obtain a high-dimensional feature vector. Specifically, the multi-dimensional feature vector (size = 16 frequency bands × 6 features × 2 modes) is used as the input; the feature dimension is mapped to a high-dimensional space (number of channels = 64) through a 1×1 convolutional layer.

[0068] The feature extraction layer is used to capture the frequency band correlation in the high-dimensional feature vector based on parallel convolutional kernels, and use a frequency band attention module to calculate the weights of each frequency band to obtain bearing fault features. Specifically, 3 groups of parallel convolutional kernels (sizes are 1×3, 1×5, 1×7 respectively) are used to capture the frequency band correlation at different scales; each group is followed by a BatchNorm layer and a ReLU activation function after convolution. A further implementation is that in the frequency band attention module, global average pooling is performed on the high-dimensional feature vector to obtain a frequency band weight vector; based on the frequency band weight vector, an attention coefficient is generated through a fully connected layer and a Sigmoid function; the attention coefficient is weighted to the high-dimensional feature vector to obtain bearing fault features.

[0069] The classification layer is used to classify the bearing fault features to obtain the fault category and confidence level.

[0070] A further implementation is that when the confidence level is lower than the preset confidence threshold, a multi-model voting mechanism is triggered, and 3 bearing fault diagnosis models with different initialization parameters are used for independent prediction, and the majority voting method is used to obtain the final diagnosis result.

[0071] In this embodiment, the Focal Loss is adopted to solve the problem of sample imbalance.

[0072] S5: Based on the bearing fault diagnosis model, diagnose the bearing fault to be diagnosed to obtain a fault diagnosis result.

[0073] In this embodiment, under the transfer learning framework, the trained model parameters are transferred to a new bearing model, and a small amount of new data is used for fine-tuning, specifically including adjusting the learning rate and retraining some network layers to improve the generalization ability of the model. Load the weight parameters and bias parameters of the bearing fault diagnosis model, and adjust the model learning rate for a small amount of training data of the target bearing model. Judge the feature distribution of the target data set to determine the range of network layers that need to be retrained. Using the adjusted learning rate, perform parameter fine-tuning on the target network layer, and retain the pre-trained parameters of the remaining network layers. Use the fine-tuned model to train the training data of the new bearing model to obtain updated model parameters. According to the updated model, calculate the accuracy on the validation set to judge whether the preset threshold is reached. If the accuracy meets the preset threshold, save the fine-tuned model parameters. If the accuracy does not reach the preset threshold, further adjust the learning rate or network layer structure and retrain the model.

[0074] In this embodiment, the method further includes: for unknown fault types, convert the vibration signal into an input form suitable for the Isolation Forest algorithm, and use the Isolation Forest algorithm to perform anomaly detection on the new vibration signal to judge whether there is a new type of fault. Specifically,

[0075] Convert the feature vector into an input form suitable for the Isolation Forest algorithm, use the Isolation Forest algorithm to train the converted data to obtain a bearing fault diagnosis model. Use the trained bearing fault diagnosis model to detect new fault signals to judge whether there is an anomaly. If the detection result is abnormal, it is judged as a new type of fault, and the fault characteristics are recorded. If the detection result is normal, continue with real-time monitoring to obtain new fault signals. Update the fault diagnosis model according to the fault characteristics and fault signal data to improve the detection accuracy.

[0076] In this embodiment, the output of the bearing fault diagnosis model includes, in addition to the bearing fault type, the fault severity, time-frequency characteristics, and specific operating conditions data (speed, load, and temperature).

[0077] Adopt a multi-task learning (MTL) framework to simultaneously predict the remaining useful life (RUL) and health status (Health Index, HI) of the bearing. Specifically, the multi-task learning architecture includes an input layer, a shared feature extraction layer, an RUL prediction branch, and an HI prediction branch.

[0078] In the input layer, the input data includes:

[0079] Shared features: high-dimensional feature vectors extracted from a pre-trained deep convolutional neural network (see step S4);

[0080] Operating conditions data: speed, load, and temperature;

[0081] Fault diagnosis result: fault type and severity;

[0082] Shared feature extraction layer: Extract high-dimensional shared features from multi-dimensional feature vectors for use in the RUL prediction and HI prediction branches; Ensure the stability and generalization ability of feature extraction through pre-trained network parameters. Among them, the input is a multi-dimensional feature vector (size = 16 frequency bands × 6 features × 2 modalities), and the output is a high-dimensional shared feature vector (size = 1 × 128).

[0083] In the RUL prediction branch, based on the shared features and operating condition data, predict the remaining useful life (RUL) of the bearing; Provide the confidence interval of the RUL to support uncertainty quantification. The specific structure is: a fully connected layer 1 with the number of neurons = 64 and the activation function = ReLU; a fully connected layer 2 with the number of neurons = 32 and the activation function = ReLU; an output layer with the number of neurons = 1 and the activation function = Sigmoid (output the normalized RUL value). Use the Monte Carlo Dropout (MC Dropout) method to randomly discard neurons (dropout rate = 0.2) during the inference stage, and repeat the prediction 100 times to calculate the mean and standard deviation of the RUL.

[0084] In this embodiment, Monte Carlo Dropout (MC Dropout), as an uncertainty quantification method based on Bayesian approximation, generates multiple prediction results by randomly discarding neurons during the inference stage, so as to calculate the mean and standard deviation of the predicted values. In the present invention, MCDropout is used to quantify the uncertainty of RUL prediction, which specifically includes the following steps: Enable Dropout during the training stage and randomly discard neurons; Keep Dropout activated during the inference stage and repeat the prediction multiple times; Calculate the mean and standard deviation of the predicted values to generate a confidence interval.

[0085] Among them, during the model training stage, Dropout layers (dropout rate p = 0.2) are added after the two fully connected layers of the RUL prediction branch; Enable Dropout during the training stage and randomly discard neurons to prevent overfitting. Use the Huber loss function to optimize the RUL prediction branch. Use the AdamW optimizer (learning rate = 1e-4, weight decay = 1e-5).

[0086] Keep the Dropout layer activated during the inference stage and randomly discard neurons (dropout rate p = 0.2); Each time during inference, the Dropout layer generates different subsets of neurons, resulting in different prediction results. Then repeat the prediction, repeat the prediction N = 100 times for the same input data to obtain a set of RUL prediction values.

[0087] In the HI prediction branch, based on the shared features and the fault diagnosis results, the bearing health index (HI) is predicted; a trend graph of HI is provided to support the visualization of the health status. The specific structure includes: a fully-connected layer 1 with 64 neurons and the activation function ReLU; a fully-connected layer 2 with 32 neurons and the activation function ReLU; an output layer with 1 neuron and the activation function Sigmoid (outputting the HI value). The HI prediction value is recorded every 1 hour, and a trend graph of HI for the past 24 hours is generated; the sliding average method (window size = 5) is used to smooth the trend graph to reduce fluctuations.

[0088] The output layer integrates the RUL prediction and HI prediction results to generate a comprehensive life prediction report; maintenance suggestions and alarm functions are provided. The RUL outputs the remaining service life (unit: hours) and the confidence interval; the HI outputs the health index (range: 0 - 1) and the trend graph; the maintenance suggestions include: if the RUL ≤ 24 hours, an emergency maintenance alarm is triggered; if the HI ≤ 0.3, a preventive maintenance suggestion is triggered.

[0089] Embodiment 2

[0090] The present invention also provides a system for bearing fault diagnosis based on a neural network, applying the method of Embodiment 1, including:

[0091] A data acquisition module, which is used to obtain bearing fault signals under different rotational speeds, loads, and noise conditions of the bearing based on an adaptive sensor array, and preprocess the bearing fault signals to extract time-domain features; wherein, the bearing fault signals include vibration signals and acoustic emission signals;

[0092] A frequency band decomposition module, which is used to decompose the preprocessed bearing fault signals by using wavelet transform to obtain frequency band components;

[0093] A multi-dimensional vector construction module, which is used to extract the frequency-domain features of each frequency band component based on Fourier transform, and combine the time-domain features to construct a multi-dimensional feature vector;

[0094] A model construction module, which is used to train a deep convolutional neural network model by using the multi-dimensional feature vector to obtain a bearing fault diagnosis model;

[0095] A diagnosis module, which is used to diagnose the bearing fault to be diagnosed based on the bearing fault diagnosis model to obtain a fault diagnosis result.

[0096] In the data acquisition module, the adaptive sensor array is a ring-shaped telescopic sensor array, and a plurality of groups of sensor nodes are uniformly arranged circumferentially on the bearing housing. Each group of sensor nodes includes a three-axis vibration sensor and an acoustic emission sensor; each of the sensor nodes is connected to the bearing housing through an electromagnetic drive slide rail.

[0097] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for bearing fault diagnosis based on neural network, characterized in that: The method comprises: Acquire bearing fault signals under different speeds, loads and noise conditions based on an adaptive sensor array, and preprocess the bearing fault signals to extract time domain features; wherein the bearing fault signals include vibration signals and acoustic emission signals; Decomposing the preprocessed bearing fault signal by wavelet transform to obtain frequency band components; Based on Fourier transform, frequency domain features of each frequency band component are extracted, and combined with the time domain features to construct a multi-dimensional feature vector; Using the multidimensional feature vector to train a deep convolutional neural network model to obtain a bearing fault diagnosis model; Based on the bearing fault diagnosis model, the bearing fault to be diagnosed is diagnosed to obtain a fault diagnosis result.

2. The method according to claim 1, characterized in that The adaptive sensor array is an annular retractable sensor array, in which several groups of sensor nodes are evenly arranged around the bearing seat, each group of sensor nodes includes a three-axis vibration sensor and an acoustic emission sensor; each of the sensor nodes is connected to the bearing seat via an electromagnetically driven slide rail.

3. The method according to claim 1, characterized in that The method for preprocessing the bearing fault signal includes: Calculating the mutual correlation coefficient of the bearing fault signals collected by adjacent sensors, and dividing the mutual correlation coefficient by using a preset correlation coefficient threshold to obtain homologous signals and heterologous signals; Performing weighted averaging on the homologous signals and filtering random noise to obtain preprocessed homologous signals; Adaptively reduce noise on the heterogeneous signal to obtain a preprocessed heterogeneous signal.

4. The method according to claim 1, characterized in that: The method for obtaining the frequency band components comprises: Based on the impact characteristics of the bearing fault signal, establishing a candidate wavelet base; Calculate the matching degree between each candidate wavelet base and the preprocessed bearing fault signal, and use the wavelet base that meets the preset matching degree threshold as the decomposition basis function; Based on the decomposition basis function, multi-scale decomposition is performed on the bearing fault signal to obtain initial frequency band components; The initial frequency band components are screened by using a frequency band entropy analysis method to obtain final frequency band components.

5. The method according to claim 1, characterized in that The time domain features include mean, variance, kurtosis, instantaneous frequency and envelope spectrum; the frequency domain features include spectrum peak, spectrum center of gravity and spectrum entropy; the multidimensional feature vector is formed by splicing the time domain features and the frequency domain features according to the frequency band dimension.

6. The method according to claim 1, characterized in that The bearing fault diagnosis model includes an input layer, a feature extraction layer and a classification layer; The input layer is used to map the input multi-dimensional feature vector to a high-dimensional space to obtain a high-dimensional feature vector; The feature extraction layer is used to capture the frequency band correlation in the high-dimensional feature vector based on the parallel convolution kernel, and calculate the weight of each frequency band using the frequency band attention module to obtain the bearing fault feature; The classification layer is used to classify the bearing fault features to obtain the fault category and confidence level.

7. The method according to claim 6, characterized in that In the frequency band attention module, the high-dimensional feature vector is globally averaged pooled to obtain a frequency band weight vector; based on the frequency band weight vector, an attention coefficient is generated through a fully connected layer and a Sigmoid function; the attention coefficient is weighted to the high-dimensional feature vector to obtain the bearing fault feature.

8. The method according to claim 6, characterized in that When the confidence is lower than the preset confidence threshold, the multi-model voting mechanism is triggered, and three bearing fault diagnosis models with different initialization parameters are used for independent prediction, and the final diagnosis result is obtained by majority voting.

Images