A bearing fault diagnosis method based on simulated acoustic vibration signal driving
By constructing a bearing simulation acoustic and vibration signal dataset and combining it with ICEEMDAN-PCA noise reduction technology, a vibration and acoustic CNN diagnostic model was trained, which solved the problems of difficult sample acquisition and high misjudgment rate in bearing fault diagnosis and achieved high-precision bearing fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ERDOS YINGPANHAO COAL CO LTD
- Filing Date
- 2025-05-20
- Publication Date
- 2026-06-30
AI Technical Summary
Existing technologies for bearing fault diagnosis suffer from problems such as difficulty in obtaining fault samples, high costs, limited effectiveness of traditional noise reduction methods, susceptibility of single sensors to environmental noise, and poor adaptability of deep learning models in industrial settings, leading to high misjudgment rates.
By constructing a bearing simulation acoustic and vibration signal dataset, performing data augmentation processing, and combining it with ICEEMDAN-PCA noise reduction technology, a vibration and acoustic CNN diagnostic model is trained. Transfer learning and DS decision-level fusion are then employed to comprehensively utilize vibration and acoustic signals for fault diagnosis.
It improves the authenticity of fault sample acquisition and noise reduction effect, reduces the false judgment rate, improves the diagnostic accuracy, adapts to noise interference in industrial environment, and realizes high-precision bearing fault diagnosis.
Smart Images
Figure CN120579017B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mechanical equipment fault diagnosis technology, specifically to a bearing fault diagnosis method based on simulated acoustic and vibration signals. Background Technology
[0002] As a core component of rotating machinery, bearing failure can lead to equipment downtime or even safety accidents. When it comes to fault diagnosis of large bearings, the cost of fault testing is high due to the high cost of bearings. For example, the cost of a single large bearing fault test for the QJ338 main ventilation fan bearing in a mine exceeds 30,000 yuan, and it is difficult to obtain fault samples.
[0003] Since vibration and sound signals in industrial settings are easily polluted by environmental noise, traditional noise reduction methods such as wavelet thresholding only improve the signal-to-noise ratio of bearing vibration signals by 4.2 to 5.8 dB, and the characteristic frequency retention rate is <85%.
[0004] Deep learning models trained on experimental data have poor adaptability to simulation signals. The ResNet model trained on CWRU bearing data showed a drop in accuracy from 98.6% to 72.3% in industrial field data testing.
[0005] Single vibration or acoustic sensors are easily affected by changes in sensor installation location and operating conditions, with a false judgment rate exceeding 12%.
[0006] The above problems urgently need to be solved. To address this, the present invention proposes a bearing fault diagnosis method based on simulated acoustic and vibration signals. Summary of the Invention
[0007] The technical problem to be solved by this invention is to generate bearing acoustic vibration signals through high-precision simulation technology, thereby solving the problem of insufficient fault samples and greatly reducing the misjudgment rate, thus providing a bearing fault diagnosis method based on simulated acoustic vibration signals.
[0008] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:
[0009] S1: Construct a dataset of simulated bearing vibration signals and simulated acoustic signals, and perform data augmentation processing;
[0010] S2: Collect the actual vibration and acoustic signals of the experimental bearing and the target bearing to be diagnosed. After preprocessing the actual vibration and acoustic signals, perform noise reduction processing using ICEEMDAN-PCA.
[0011] S3: Train the vibration CNN diagnostic model and the acoustic CNN diagnostic model respectively using the simulated vibration signal and simulated acoustic signal after data augmentation processing;
[0012] S4: Use the actual vibration and acoustic signals of the denoised experimental bearing to perform transfer learning on the vibration CNN diagnostic model and the acoustic CNN diagnostic model;
[0013] S5: Input the noise-reduced actual vibration signal and acoustic signal of the target bearing to be diagnosed into the vibration CNN diagnostic model and acoustic CNN diagnostic model after transfer learning, respectively, to obtain the vibration diagnosis result and acoustic diagnosis result. Then, perform DS decision-level fusion on the vibration diagnosis result and acoustic diagnosis result to output the final fault type.
[0014] Furthermore, in step S1, the specific process of constructing the bearing simulation vibration signal dataset is as follows:
[0015] S101: Establish three-dimensional models of normal bearings and faulty bearings with different fault types and fault sizes. The fault types include inner ring faults, outer ring faults, and rolling element faults. The fault sizes include 0.5mm faults, 1mm faults, and 2mm faults.
[0016] S102: The original simulated vibration signals of each simulated bearing model were obtained by transient dynamic simulation, with the simulated bearing speeds set to 400 r / min, 600 r / min, and 800 r / min, respectively. The peak frequency of the signal envelope spectrum was quantitatively compared with the theoretical fault characteristic frequency.
[0017] S103: Perform time shifting, amplitude scaling, and noise injection processing on the original simulated vibration signal to obtain hybrid enhanced data. The hybrid enhanced data and the original data are then compared.
[0018] Furthermore, in step S1, the specific process of constructing the bearing simulation acoustic signal dataset is as follows:
[0019] S111: Construct an acoustic-structure interaction model of the bearing housing, set the acoustic field boundary conditions, and place probes in the acoustic field to obtain simulated acoustic signals;
[0020] S112: After performing time-frequency domain conversion on the original simulated vibration signal of the bearing in step S102, it is used as an excitation input to the acoustic-structure coupling model of the bearing housing to calculate the three-dimensional sound field distribution and obtain the corresponding original simulated acoustic signal.
[0021] S113: Verify the consistency between the peak frequency of the envelope spectrum of the original simulated acoustic signal and the theoretical fault characteristic frequency;
[0022] S114: Perform time shifting, amplitude scaling, and noise injection processing on the original simulated acoustic signal to obtain hybrid enhanced data. The hybrid enhanced data and the original data are then compared.
[0023] Furthermore, in step S2, the actual vibration and acoustic signals of the experimental bearing and the target bearing to be diagnosed are acquired as follows: A bearing test bench is set up, acoustic and vibration sensors are installed, and normal bearings and bearings with inner ring faults, outer ring faults, and rolling element faults with a fault size of 1 mm are selected as experimental bearings. The sampling frequency is set to 32 kHz, and the actual vibration and acoustic signals of the experimental bearings are acquired at a speed of 600 r / min. Multiple groups of bearings to be diagnosed, including inner ring faults, outer ring faults, and rolling element faults with known fault types, are selected. The sampling frequency is set to 32 kHz, and the actual vibration and acoustic signals of the bearings to be diagnosed are acquired at different speeds.
[0024] Furthermore, in step S2, the mean values of the actual vibration signal and the acoustic signal are calculated separately and zero-mean processing is performed to eliminate sensor bias error, and the signal is bandpass filtered to filter out high-frequency noise and low-frequency components unrelated to bearing failure.
[0025] Furthermore, in step S2, the specific steps of ICEEMDAN-PCA joint noise reduction are as follows:
[0026] S21: The actual acquired signal is decomposed using ICEEMDAN, Gaussian white noise is added and residual components are calculated, and IMF components of each order are extracted from them. The actual acquired signal includes actual vibration signal and acoustic signal.
[0027] S22: The IMF components obtained by ICEEMDAN decomposition are sorted using principal component analysis, and the IMF components with a cumulative variance contribution rate greater than a set value are selected as the main feature components. The signal is then reconstructed using the main feature components.
[0028] Furthermore, in step S3, both the vibration CNN diagnostic model and the acoustic CNN diagnostic model include:
[0029] Input layer: Receives input vibration or acoustic signals;
[0030] Convolutional layers: Extract local features of the signal through multiple convolutional kernels. The first convolutional layer contains 32 3×3 convolutional kernels with a stride of 1; the second convolutional layer contains 64 3×3 convolutional kernels with a stride of 1; and the third convolutional layer contains 128 3×3 convolutional kernels with a stride of 1.
[0031] Batch normalization layer: Normalizes the output of the convolutional layer;
[0032] Activation layer: The ReLU function is used to perform a nonlinear transformation on the normalization result;
[0033] Max pooling layer: 3×3 window, stride of 2, reduces signal dimension while preserving corresponding features;
[0034] Fully connected layer: Integrates features output from the pooling layer;
[0035] Random deactivation layer: Neurons are randomly discarded at a preset ratio after the fully connected layer to suppress overfitting;
[0036] Classification layer: Outputs the probability distribution of each fault type;
[0037] Output layer: Determines fault type labels based on probability distribution.
[0038] Furthermore, in step S4, the specific processing procedure is as follows:
[0039] S41: Freeze the parameters of the convolutional layer, batch normalization layer, and pooling layer of the vibration CNN diagnostic model and the acoustic CNN diagnostic model;
[0040] S42: Unfreeze and retrain the fully connected layer and the classification layer, and input the time-frequency feature map of the actual signal of the experimental bearing;
[0041] S43: A dynamic learning strategy is adopted. The initial learning rate is set to 1 / 10 of the training phase in step S3. The multi-task joint loss of the validation set is used as the decay basis. When the validation loss does not decrease for 3 consecutive epochs, the learning rate is decayed to 1 / 5 of the original value.
[0042] S44: Set an early stop mechanism. When the verification accuracy fluctuates less than the set value for 10 consecutive epochs, the training is terminated, thereby obtaining the vibration CNN diagnostic model and the acoustic CNN diagnostic model after transfer learning.
[0043] Furthermore, in step S5, the specific process of DS decision-level fusion is as follows:
[0044] S51: Define a fault type identification framework, assuming that the space Θ includes four fault types: normal, inner ring fault, outer ring fault and rolling element fault, where N is normal, I is inner ring fault, O is outer ring fault and R is rolling element fault;
[0045] For each fault type θ∈Θ, the trust assignment function of the vibration CNN diagnostic model is defined as:
[0046]
[0047] The trust assignment function for the acoustic CNN diagnostic model is defined as follows:
[0048]
[0049] in, ∑ represents the unnormalized output values of the vibration CNN diagnostic model and the acoustic CNN diagnostic model for the fault type θ, respectively.k∈Θ This represents the summation of all fault types k in the spatial set Θ = {N, I, O, R};
[0050] S52: Let The fusion hypothesis to be determined. These are subsets of the diagnostic results from the vibration CNN diagnostic model and the acoustic CNN diagnostic model, respectively. Each subset includes the corresponding diagnostic results, and the confidence level of the diagnostic result subset is defined as follows:
[0051]
[0052] By fusing the confidence levels of the two types of diagnostic results using the DS combination rule, the fault type with the highest fused confidence level is determined as the final diagnostic result.
[0053] Furthermore, in step S52, the DS combination rule satisfies:
[0054]
[0055] w a =1-w v
[0056]
[0057] Where, m fusion (A) represents the integration of trust levels, w v w is the vibration signal weighting coefficient. a SNR is the acoustic signal weighting coefficient, and SNR is the signal-to-noise ratio. v The signal-to-noise ratio (SNR) of the vibration signal. a P represents the acoustic signal-to-noise ratio. s For the target signal power, P n This represents the background noise power.
[0058] The present invention has the following advantages over the prior art:
[0059] 1. The simulation data has high fidelity, with envelope spectrum feature frequency matching degree >95%. By using simulation to obtain bearing fault sample data, the experimental requirements for fault samples of large bearings are reduced, greatly reducing the cost of obtaining bearing fault samples.
[0060] 2. While preserving the characteristic frequencies of the actual collected signals, the signal-to-noise ratio is significantly improved, industrial environmental noise interference is reduced, and the diagnostic accuracy is improved to a certain extent.
[0061] 3. The method of comprehensive diagnosis using bearing acoustic signals and vibration signals avoids abnormal diagnostic results due to fluctuations in a single signal. It combines the advantages of both acoustic signal fault diagnosis and vibration signal fault diagnosis, greatly improving the diagnostic accuracy. Attached Figure Description
[0062] Figure 1 This is a flowchart illustrating the bearing fault diagnosis method driven by simulated acoustic and vibration signals in an embodiment of the present invention.
[0063] Figure 2 This is the vibration signal simulation model in the embodiments of the present invention;
[0064] Figure 3 This is the envelope spectrum of the simulated vibration signal of the outer ring fault at a speed of 600 r / min in this embodiment of the invention;
[0065] Figure 4 This is the acoustic signal simulation model in the embodiments of the present invention;
[0066] Figure 5 This refers to the low-frequency component of the simulated acoustic signal for an outer ring fault at a speed of 600 r / min in this embodiment of the invention.
[0067] Figure 6 This is a physical diagram of a bearing test bench for collecting actual bearing signals in an embodiment of the present invention, wherein 1 is a vibration sensor, 2 is an acoustic sensor, 3 is a speed controller, and 4 is a data acquisition card;
[0068] Figure 7 This is an IMF component diagram of the actual ICEEMDAN decomposition of the outer ring fault signal in an embodiment of the present invention;
[0069] Figure 8 This is a comparison diagram of the actual signal of the outer ring fault before and after ICEEMDAN-PCA noise reduction in an embodiment of the present invention;
[0070] Figure 9 This is a schematic diagram of the structure of the CNN diagnostic model in an embodiment of the present invention. Detailed Implementation
[0071] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.
[0072] like Figure 1As shown, this embodiment provides a technical solution: a bearing fault diagnosis method based on simulated acoustic and vibration signals. It employs ICEEMDAN-PCA joint noise reduction technology to improve the signal-to-noise ratio while preserving fault characteristic frequencies. A dynamic learning rate transfer strategy is designed to enable rapid adaptation of the simulation model to industrial scenarios. This strategy increases the model's convergence speed by 2.3 times (reducing the number of iterations from 120 to 52) during mixed data training, shortening the training time from 4.2 hours to 2.7 hours, a 37% reduction. The method integrates vibration and acoustic dual-modal diagnostic results and reduces the false positive rate based on signal-to-noise ratio-weighted DS evidence theory. Specifically, it includes the following steps:
[0073] Step 1: Taking the 6012 bearing as an example, build a transient dynamic model of the bearing to obtain the bearing simulation vibration signal; input the obtained simulation vibration signal into the bearing housing acoustic-structure coupling model to obtain the bearing simulation sound signal.
[0074] Step 1.1: Build a transient dynamic model of the bearing to obtain the simulated vibration signal of the bearing, specifically including:
[0075] Construct 3D models of the 6012 bearing for normal bearing conditions, inner ring faults, outer ring faults, and rolling element faults. Fault dimensions include 0.5mm faults, 1mm faults, and 2mm faults.
[0076] like Figure 2 As shown, an ANSYS transient dynamic model was built for a 3D bearing model with different fault types and sizes. Mesh generation was performed, and the sampling frequency was set to 32kHz, with rotational speed gradients of 400 r / min, 600 r / min, and 800 r / min, covering the typical operating conditions of the bearing. The simulation duration was 0.5s, obtaining simulated vibration signals under different fault conditions.
[0077] like Figure 3 The figure shows the envelope spectrum of the simulated vibration signal for an outer ring fault at a speed of 600 r / min. The characteristic frequency of the outer ring fault is calculated according to the formula for calculating the characteristic frequency of the outer ring fault.
[0078]
[0079] Where n is the number of rolling elements, d is the diameter of the rolling elements, and D m Let α be the pitch circle diameter, α be the contact angle, and f be the contact angle. r The rotational speed is the frequency.
[0080] Substituting the parameters of bearing 6012, n=14, d=10mm, D m =77.5mm, α=0°, f r=10Hz, the characteristic frequency of the outer ring fault of bearing 6012 is calculated to be about 60.97Hz. The characteristic frequency and its harmonics of the bearing outer ring fault can be clearly found on the envelope spectrum of the outer ring fault, which verifies the correctness of the simulation signal.
[0081] Step 1.2: Input the obtained simulated vibration signal into the bearing housing acoustic-structure interaction model to obtain the simulated bearing acoustic signal, specifically including:
[0082] like Figure 4 As shown, a COMSOL bearing housing acoustic-structure coupling model was constructed, and a perfectly matched layer was set on the periphery of the sound field as an ideal sound wave absorber to isolate sound wave reflection interference.
[0083] Since the excitation of the COMSOL acoustic-structure interaction model is a frequency domain signal, the simulated vibration signal is subjected to FFT transformation.
[0084] The simulated vibration acceleration signal after FFT transformation is multiplied by the equivalent mass of the bearing to simulate the excitation force. The force is then applied to the bearing housing for simulation. Probes are set in the sound field to acquire the simulated acoustic signal.
[0085] like Figure 5 As shown, the simulated acoustic signal of the outer ring fault at a speed of 600 r / min can also clearly identify the characteristic frequency and its harmonics of the outer ring fault, verifying the correctness of the simulation signal.
[0086] Step 2: Collect the actual vibration and acoustic signals of the experimental bearing and the target bearing to be diagnosed, specifically including:
[0087] like Figure 6 As shown, a bearing test bench was built, and acoustic and vibration sensors were installed. Normal bearings and bearings with inner ring, outer ring, and rolling element faults (1 mm fault size) were selected as test bearings. The sampling frequency was set to 32 kHz, and the actual vibration and acoustic signals of the test bearings were collected at a speed of 600 r / min. Multiple groups of bearings with known fault types, including inner ring, outer ring, and rolling element faults, were selected. The sampling frequency was set to 32 kHz, and the actual vibration and acoustic signals of the bearings to be diagnosed were collected at different speeds.
[0088] Step 3: Preprocess the acquired signals by calculating the mean and zero-mean normalization to eliminate sensor bias errors, and then performing bandpass filtering to remove high-frequency noise and low-frequency components unrelated to bearing failure. Finally, perform ICEEMDAN-PCA combined noise reduction on the preprocessed signals, specifically including:
[0089] like Figure 7As shown, the preprocessed actual signal is decomposed using ICEEMDAN, Gaussian white noise is added, residual components are calculated, and intrinsic mode functions (IMF components) are extracted.
[0090] Table 1. Contribution of each IMF signal to the variance of the entire data matrix
[0091]
[0092]
[0093] Principal component analysis (PCA) was used to screen IMF components with a cumulative variance contribution rate exceeding 95% for signal reconstruction. When the number of principal components was three, the cumulative variance contribution rate reached 96.41%, indicating that the top three principal components were highly representative of the decomposed sequence and preserved complete data information. Therefore, IMF1, IMF3, and IMF2 were selected to reconstruct the signal to ensure noise reduction while preserving as much complete data information as possible.
[0094] like Figure 8 As shown, a comparison diagram of the signal before and after noise reduction of the outer ring fault signal of bearing 6012 is presented. Noise reduction is achieved by removing irrelevant noise while retaining the main characteristics of the signal.
[0095] Step 4: Build and train the vibration CNN diagnostic model and acoustic CNN diagnostic model for fault diagnosis. Specific steps include:
[0096] like Figure 9 As shown, a CNN diagnostic model with a suitable structure is built.
[0097] Data augmentation is performed on simulated vibration and acoustic signals. The original simulated vibration signal is time-shifted by ±100ms, amplitude-scaled by 0.8 to 1.2, and noise-injected by 6% to 18%. The augmented data is then mixed with the original data.
[0098] A vibration CNN diagnostic model was trained using data-augmented simulated vibration signals, and an acoustic CNN diagnostic model was trained using data-augmented simulated acoustic signals.
[0099] The actual vibration and acoustic signals of the collected experimental bearings were used for transfer learning on the diagnostic model. The signals were divided into 1024 points / sample using a sliding window, with a batch size of 64. After training, the model's diagnostic accuracy for the target bearing data improved from 92.4% to 98.2%. Specific details include:
[0100] Freeze the parameters of convolutional layers, batch normalization layers, and pooling layers in the vibration CNN diagnostic model and the acoustic CNN diagnostic model.
[0101] Unfreeze and retrain the fully connected layer and classification layer, and input the time-frequency feature map of the actual signal from the experimental bearing.
[0102] A dynamic learning strategy was adopted, with the initial learning rate set to 1 / 10 of that used in the pre-training phase with augmented data. The multi-task joint loss of the validation set was used as the decay criterion. When the validation loss did not decrease for three consecutive epochs, the learning rate was decayed to 1 / 5 of the original value. The dynamic learning strategy improved the convergence speed of the model on industrial field data by 2.3 times and reduced the training time by 37% compared with the fixed learning rate strategy.
[0103] An early stop mechanism is set up to terminate training when the validation accuracy fluctuates by less than ±0.3% for 10 consecutive epochs.
[0104] Step 5: Perform fault diagnosis on the noise-reduced acquired target bearing signal, including:
[0105] The noise-reduced vibration signal of the target bearing to be diagnosed is input into the trained vibration CNN model for fault diagnosis.
[0106] The denoised acoustic signal of the target bearing to be diagnosed is input into the trained acoustic CNN model for fault diagnosis.
[0107] The diagnostic results from the vibration CNN model and the acoustic CNN model are fused at the DS decision level to obtain the final diagnostic result. The specific content of the DS decision-level fusion includes:
[0108] S51: Define a fault type identification framework, assuming that the space Θ includes four fault types: normal, inner ring fault, outer ring fault and rolling element fault, where N is normal, I is inner ring fault, O is outer ring fault and R is rolling element fault;
[0109] For each fault type θ∈Θ, the trust assignment function of the vibration CNN diagnostic model is defined as:
[0110]
[0111] The trust assignment function for the acoustic CNN diagnostic model is defined as follows:
[0112]
[0113] in, ∑ represents the unnormalized output values of the vibration CNN diagnostic model and the acoustic CNN diagnostic model for the fault type θ, respectively. k∈Θ The summation is performed on all fault types k in the spatial set Θ = {N, I, O, R}.
[0114] It should be noted that in the above formula, θ and k are both elements taken from the fault type set Θ = {N, I, O, R}, but their meanings are slightly different. θ represents the target fault type currently being calculated for trust level. k represents the intermediate variable used to iterate through all fault types during the normalization process. For ease of understanding, an example is given below:
[0115] Assume the unnormalized output values of the vibration CNN diagnostic model for the four types of faults are: At this point, if the trust assignment value m corresponding to the outer ring fault O is required... v If (O), then let θ = O, meaning we are asking for the trust level of class O. The summation variable k iterates through N, I, O, and R in turn. Substituting into the formula, we get:
[0116]
[0117] Thus, θ = O indicates that the current calculation is for the confidence level of the outer ring faults, and k ∈ Θ indicates that the summation of all fault types in the denominator is normalized.
[0118] S52: Let The fusion hypothesis to be determined. These are subsets of the diagnostic results from the vibration CNN diagnostic model and the acoustic CNN diagnostic model, respectively. Each subset includes the corresponding diagnostic results, and the confidence level of the diagnostic result subset is defined as follows:
[0119]
[0120] By fusing the confidence levels of the two types of diagnostic results using the DS combination rule, the fault type with the highest fused confidence level is determined as the final diagnostic result.
[0121] The DS combination rule satisfies:
[0122]
[0123] w a =1-w v
[0124]
[0125] Where, m fusion (A) represents the integration of trust levels, w v w is the vibration signal weighting coefficient. a SNR is the acoustic signal weighting coefficient, and SNR is the signal-to-noise ratio. v The signal-to-noise ratio (SNR) of the vibration signal. a P represents the acoustic signal-to-noise ratio. s For the target signal power, P n This represents the background noise power.
[0126] Taking the outer ring fault test bearing signal fault diagnosis as an example, the vibration signal signal-to-noise ratio (SNR) v =18dB, Acoustic signal-to-noise ratio (SNR) a =12dB, weighted to obtain w v =0.60, w a =0.40, the overall confidence level of the outer ring fault after fusion is 0.88. When performing fault diagnosis on the data of the target bearing to be diagnosed, the accuracy rate of single vibration signal diagnosis is 92.5%, and that of acoustic signal is 85.3%. The overall accuracy rate after fusion is 96.8%, and the misjudgment rate is reduced from 7.5% for single vibration signal to 3.2%.
[0127] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A bearing fault diagnosis method based on simulated acoustic vibration signal driving, characterized in that, Includes the following steps: S1: Construct a dataset of simulated bearing vibration signals and simulated acoustic signals, and perform data augmentation processing; S2: Collect the actual vibration and acoustic signals of the experimental bearing and the target bearing to be diagnosed. After preprocessing the actual vibration and acoustic signals, perform noise reduction processing using ICEEMDAN-PCA. S3: Train the vibration CNN diagnostic model and the acoustic CNN diagnostic model respectively using the simulated vibration signal and simulated acoustic signal after data augmentation processing; S4: Use the actual vibration and acoustic signals of the denoised experimental bearing to perform transfer learning on the vibration CNN diagnostic model and the acoustic CNN diagnostic model; S5: Input the noise-reduced actual vibration signal and acoustic signal of the target bearing to be diagnosed into the vibration CNN diagnostic model and acoustic CNN diagnostic model after transfer learning, respectively, to obtain the vibration diagnosis result and acoustic diagnosis result. Then, perform DS decision-level fusion on the vibration diagnosis result and acoustic diagnosis result to output the final fault type. In step S4, the specific processing procedure is as follows: S41: Freeze the parameters of the convolutional layer, batch normalization layer, and pooling layer of the vibration CNN diagnostic model and the acoustic CNN diagnostic model; S42: Unfreeze and retrain the fully connected layer and the classification layer, and input the time-frequency feature map of the actual signal of the experimental bearing; S43: A dynamic learning strategy is adopted. The initial learning rate is set to 1 / 10 of the training phase in step S3. The multi-task joint loss of the validation set is used as the decay basis. When the validation loss does not decrease for 3 consecutive epochs, the learning rate is decayed to 1 / 5 of the original value. S44: Set an early stop mechanism. When the verification accuracy fluctuates less than the set value for 10 consecutive epochs, the training will be terminated, and the vibration CNN diagnostic model and acoustic CNN diagnostic model after transfer learning will be obtained. In step S5, the specific process of DS decision-level fusion is as follows: S51: Define a fault type identification framework, assuming that the space Θ includes four fault types: normal, inner ring fault, outer ring fault and rolling element fault, where N is normal, I is inner ring fault, O is outer ring fault and R is rolling element fault; For each fault type θ∈Θ, the trust assignment function of the vibration CNN diagnostic model is defined as: ； The trust assignment function for the acoustic CNN diagnostic model is defined as follows: ； in, , The vibration CNN diagnostic model and the acoustic CNN diagnostic model are used to diagnose fault types, respectively. The unnormalized output value, This represents all fault types in the spatial set Θ={N,I,O,R}. Perform summation; S52: Let The fusion hypothesis to be determined. , These are subsets of the diagnostic results from the vibration CNN diagnostic model and the acoustic CNN diagnostic model, respectively. Each subset includes the corresponding diagnostic results, and the confidence level of the diagnostic result subset is defined as follows: ； ； By fusing the confidence levels of the two types of diagnostic results using the DS combination rule, the fault type with the highest fused confidence level is determined as the final diagnostic result. In step S52, the DS combination rule satisfies: ； ； ； ； in, The vibration signal weighting coefficient, These are the acoustic signal weighting coefficients. Signal-to-noise ratio (SNR) The signal-to-noise ratio of the vibration signal. The signal-to-noise ratio of the acoustic signal. For the target signal power, Background noise power; In step S1, the specific process of constructing the bearing simulation vibration signal dataset is as follows: S101: Establish three-dimensional models of normal bearings and faulty bearings with different fault types and fault sizes. The fault types include inner ring faults, outer ring faults, and rolling element faults. The fault sizes include 0.5mm faults, 1mm faults, and 2mm faults. S102: The original simulated vibration signals of each simulated bearing model were obtained using transient dynamic simulation. The simulated bearing speeds were set to 400 r / min, 600 r / min, and 800 r / min, respectively. The peak frequency of the signal envelope spectrum was quantitatively compared with the theoretical fault characteristic frequency, and the matching degree of the envelope spectrum characteristic frequency was >95%. S103: Perform time shifting, amplitude scaling and noise injection processing on the original simulated vibration signal to obtain hybrid enhanced data. The hybrid enhanced data and the original data are compared. In step S1, the specific process of constructing the bearing simulation acoustic signal dataset is as follows: S111: Construct an acoustic-structure interaction model of the bearing housing, set the acoustic field boundary conditions, and place probes in the acoustic field to obtain simulated acoustic signals; S112: After performing time-frequency domain conversion on the original simulated vibration signal of the bearing in step S102, it is used as an excitation input to the acoustic-structure coupling model of the bearing housing to calculate the three-dimensional sound field distribution and obtain the corresponding original simulated acoustic signal. S113: Verify the consistency between the peak frequency of the envelope spectrum of the original simulated acoustic signal and the theoretical fault characteristic frequency, with an envelope spectrum characteristic frequency matching degree >95%; S114: Perform time shifting, amplitude scaling, and noise injection processing on the original simulated acoustic signal to obtain hybrid enhanced data. The hybrid enhanced data and the original data are then compared.
2. The bearing fault diagnosis method based on simulated acoustic vibration signal drive according to claim 1, characterized in that, In step S2, the actual vibration and acoustic signals of the experimental bearing and the target bearing to be diagnosed are acquired as follows: A bearing test bench is set up, acoustic and vibration sensors are installed, and normal bearings and bearings with inner ring faults, outer ring faults, and rolling element faults with a fault size of 1 mm are selected as experimental bearings. The sampling frequency is set to 32 kHz, and the actual vibration and acoustic signals of the experimental bearings are acquired at a speed of 600 r / min. Multiple groups of bearings to be diagnosed, including inner ring faults, outer ring faults, and rolling element faults with known fault types, are selected. The sampling frequency is set to 32 kHz, and the actual vibration and acoustic signals of the bearings to be diagnosed are acquired at different speeds.
3. The bearing fault diagnosis method based on simulated acoustic vibration signal drive according to claim 1, characterized in that, In step S2, the mean values of the actual vibration signal and the acoustic signal are calculated and zero-mean processing is performed to eliminate sensor bias error. The signal is then bandpass filtered to remove high-frequency noise and low-frequency components unrelated to bearing failure.
4. The bearing fault diagnosis method based on simulated acoustic vibration signal drive according to claim 1, characterized in that, In step S2, the specific steps of ICEEMDAN-PCA joint noise reduction are as follows: S21: The actual acquired signal is decomposed using ICEEMDAN, Gaussian white noise is added and residual components are calculated, and IMF components of each order are extracted from them. The actual acquired signal includes actual vibration signal and acoustic signal. S22: The IMF components obtained by ICEEMDAN decomposition are sorted using principal component analysis, and the IMF components with a cumulative variance contribution rate greater than a set value are selected as the main feature components. The signal is then reconstructed using the main feature components.
5. The bearing fault diagnosis method based on simulated acoustic vibration signal drive according to claim 4, characterized in that, In step S3, both the vibration CNN diagnostic model and the acoustic CNN diagnostic model include: Input layer: Receives input vibration or acoustic signals; Convolutional layers: Extract local features of the signal through multiple convolutional kernels. The first convolutional layer contains 32 3×3 convolutional kernels with a stride of 1; the second convolutional layer contains 64 3×3 convolutional kernels with a stride of 1; and the third convolutional layer contains 128 3×3 convolutional kernels with a stride of 1. Batch normalization layer: Normalizes the output of the convolutional layer; Activation layer: The ReLU function is used to perform a nonlinear transformation on the normalization result; Max pooling layer: 3×3 window, stride of 2, reduces signal dimension while preserving corresponding features; Fully connected layer: Integrates features output from the pooling layer; Random deactivation layer: Neurons are randomly discarded at a preset ratio after the fully connected layer to suppress overfitting; Classification layer: Outputs the probability distribution of each fault type; Output layer: Determines fault type labels based on probability distribution.