New energy vehicle bearing diagnosis method based on synchronous compression wavelet and markov

Abstract

This invention provides a diagnostic method for new energy vehicle bearings based on synchronous compressed wavelet and Markov models, belonging to the field of fault diagnosis technology for new energy vehicle components. It involves collecting raw vibration data of new energy vehicle bearings on a test bench, converting the collected raw vibration signals into high-resolution time-frequency maps using synchronous compressed wavelet transform, constructing a transfer learning model integrating a MobileNetV2 network and a Markov chain, training and fine-tuning the transfer learning model, and inputting the high-resolution time-frequency map generated by the synchronous compressed wavelet transform of the vibration signal of the bearing to be tested into the trained and fine-tuned transfer learning model. The model automatically outputs the corresponding bearing fault type and fault severity level. Based on synchronous compressed wavelet transform, MobileNetV2 network, and Markov chain model, it achieves sufficient extraction of fault features and accurate cross-domain diagnosis, improving the model's generalization ability and diagnostic accuracy under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for new energy vehicle components, specifically to a new energy vehicle bearing diagnosis method based on synchronous compressed wavelet and Markov. Background Technology

[0002] As a core component of the drive and transmission systems, the operating status of bearings in new energy vehicles directly affects the safety and reliability of the entire vehicle. Bearings are prone to failure due to wear, fatigue, and corrosion during long-term operation. Furthermore, new energy vehicles experience complex operating conditions such as changes in vehicle speed, load fluctuations, and differences in ambient temperature during operation, resulting in bearing time-series vibration data exhibiting nonlinearity, non-stationarity, and low signal-to-noise ratio characteristics.

[0003] Existing bearing fault diagnosis methods are mainly divided into two categories: one is based on traditional signal processing methods, such as wavelet transform and Fourier transform, but these methods are insufficient for extracting features from nonlinear and non-stationary vibration data, have low time-frequency resolution, and are difficult to capture subtle fault features; the other is based on deep learning methods, such as CNN and MobileNet, which can achieve automatic feature extraction, but have problems such as requiring a large amount of labeled data for model training, weak generalization ability when diagnosing across working conditions / equipment, and being prone to overfitting. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a new energy vehicle bearing diagnosis method based on synchronous compressed wavelet and Markov model. The method fully extracts fault features and performs accurate cross-domain diagnosis based on synchronous compressed wavelet transform, MobileNetV2 network, and Markov chain model, thereby improving the model's generalization ability and diagnostic accuracy under complex working conditions.

[0005] To achieve the above objectives, the technical solution adopted is: a new energy vehicle bearing diagnosis method based on synchronous compressed wavelet and Markov, including the following steps: Step 1: Vibration signal acquisition: Set up a new energy vehicle bearing performance test bench and collect the original vibration signal data of the bearing under different speed, load and temperature conditions through an accelerometer. Step 2, Time-Frequency Image Conversion: Perform synchronous compressed wavelet transform on the acquired original vibration signal to convert the one-dimensional time domain signal into a high-resolution time-frequency image; Step 3: Construct a transfer learning model that integrates MobileNetV2 network and Markov chain: Use MobileNetV2 network as the backbone to extract spatial features of high-resolution time-frequency map, and introduce Markov chain model to model the state transition probability of spatial features, extract temporal features, and then fuse spatial features and temporal features and output the execution result after passing through fully connected layer and Softmax classification. Step 4, Model Training and Fine-tuning: First, the transfer learning model is pre-trained using a publicly available bearing fault dataset to initialize the parameters of the feature extractor; then, the model is fine-tuned using the new energy vehicle bearing test fault dataset constructed in this invention to optimize the Softmax classification output parameters. Step 5, Fault Diagnosis and Output: Input the high-resolution time-frequency graph generated by synchronous compressed wavelet transform of the vibration signal of the bearing to be tested into the transfer learning model trained and fine-tuned in Step 4. The model will automatically output the fault type and fault severity level of the corresponding bearing.

[0006] The synchronous compressed wavelet transform described in step 2 specifically includes: 1) Select Morlet wavelet as the basis function to perform continuous wavelet transform, generate wavelet coefficients, obtain the initial time-frequency matrix from the wavelet coefficients, and use the initial time-frequency matrix to compress and redistribute to obtain the time-frequency coefficients. 2) Calculate the instantaneous frequency of the wavelet coefficients and use this as a basis to regroup and compress the time-frequency coefficients, thus completing the compression of the time-frequency graph; 3) Normalize the compressed time-frequency image and crop the pixels to a uniform size to obtain a high-resolution time-frequency image.

[0007] The fine-tuning steps described in step 4 specifically include: 1) Using a transfer learning strategy, the parameters of the first 60%-70% of the complete linear bottleneck layers of the MobileNetV2 network are frozen, and only the fully connected layers, Softmax classification output, and Markov chain model are trained. 2) Set the initial learning rate to 0.001 and use the Adam optimizer and learning rate decay strategy; 3) Introduce the cross-entropy loss function as the training objective and introduce the early stopping method.

[0008] The fault types mentioned in step 5 include inner ring faults, outer ring faults, and rolling element faults.

[0009] The fault severity levels described in step 5 are specifically divided as follows: Minor fault: The characteristic amplitude is less than 1.5 times the normal average but higher than the threshold; Moderate fault: The characteristic amplitude is between 1.5 and 3 times the average value under normal conditions; Critical fault: Characteristic amplitude is more than 3 times the average value under normal conditions.

[0010] The aforementioned new energy vehicle bearings are deep groove ball bearings used in the drive motors or reducers of new energy vehicles, with bearing models ranging from the 6200 to the 6310 series.

[0011] The beneficial effects of this invention are: 1. Feature Extraction Level: Compared with traditional wavelet transform, Synchronous Compressed Wavelet Transform (SST) improves the resolution and feature aggregation of time-frequency plots by compressing the frequency dimension. It can effectively capture the subtle fault features in the nonlinear and non-stationary vibration data of new energy vehicle bearings, and solve the problem of insufficient feature extraction in traditional signal processing.

[0012] 2. Model Diagnostic Level: A Markov chain model is integrated into the MobileNetV2 network to uncover the temporal variation patterns of time-frequency features. This enables the fusion extraction of spatial and temporal features, improving feature discriminability, enhancing the representation of weak fault features, and improving diagnostic robustness under varying operating conditions. A transfer learning scheme is designed using a publicly available dataset (CWRU / XJTU-SY) as the source domain and a real-world dataset of new energy vehicle bearings as the target domain. The parameters of the first 60%-70% of the convolutional layers in MobileNetV2 are frozen, and only the top-level classifier and Markov module are trained. This addresses the industry pain point of limited labeled fault samples and weak generalization ability across operating conditions / equipment in new energy vehicle bearings, while simultaneously reducing training costs. This effectively solves the problems of limited labeled fault data and weak diagnostic generalization ability across operating conditions / equipment in new energy vehicle bearings.

[0013] 3. Practical application: This method has high diagnostic accuracy and fast computation efficiency. The lightweight structure of MobileNetV2 is compatible with the hardware resources of the vehicle terminal of new energy vehicles, which can realize online real-time diagnosis and early warning of bearing faults, thereby improving the operational safety and reliability of new energy vehicles.

[0014] 4. Not only can it identify the fault types of the inner ring / outer ring / rolling element, but it also quantifies the fault severity into three levels: minor / moderate / severe (based on the ratio of characteristic amplitude to the average value under normal conditions) for the first time, providing more refined diagnostic results that meet the actual engineering needs of vehicle online monitoring and early warning. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a comparison diagram of the inverted residual structure of the present invention and the traditional residual structure; Figure 3 This is a schematic diagram of the MobileNetV2 network core module structure of the present invention. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Example 1

[0017] like Figure 1 As shown, a new energy vehicle bearing diagnostic method based on synchronous compressed wavelet and Markov model includes the following steps: S1 Vibration Signal Acquisition This study focuses on a deep groove ball bearing (model 6206) at the drive motor end of a new energy vehicle. A piezoelectric accelerometer was used on a bearing test bench with a sampling frequency of fs = 10000 Hz. Vibration signals were collected under different operating conditions for four bearing states: normal, inner ring fault, outer ring fault, and rolling element fault. 1000 samples were collected for each state, with each sample containing 100 points. Test conditions were set with motor speeds of 3000 r / min, 12000 r / min, and 20000 r / min, and loads of 0 N·m, 5 N·m, and 10 N·m, covering typical operating conditions of new energy vehicles. For example, under a load of 0 N·m and a speed of 3000 r / min, a 10 kHz sampling frequency was used for 10 seconds, resulting in 100,000 points, divided into 1000 sample groups, each containing 100 points.

[0018] S2, Time-Frequency Image Conversion The original vibration signal is preprocessed by detrending and normalization to eliminate DC components and amplitude differences. A continuous wavelet transform is performed using the Morlet wavelet basis to generate wavelet coefficients. An initial time-frequency matrix is ​​obtained from these coefficients, and the time-frequency coefficients are then compressed and redistributed using this matrix. The instantaneous frequencies of the wavelet coefficients are calculated, and based on these frequencies, the time-frequency coefficients are regrouped and compressed to complete the compression of the time-frequency map. This concentrates energy at the true fault characteristic frequencies, suppresses noise and cross-term interference, and generates a compressed time-frequency domain image.

[0019] The compressed time-frequency domain image was uniformly scaled to 2179×1638 pixels and used in RGB three-channel format as the input image for subsequent models, i.e., high-resolution time-frequency image.

[0020] S3. Construct a transfer learning model that integrates MobileNetV2 network and Markov chain. The feature extraction network is based on the lightweight MobileNetV2 network, and its inverted residual structure is preserved. Figure 2 ) and linear bottleneck layer ( Figure 3 This reduces the number of model parameters and computational load, meeting the real-time inference requirements of in-vehicle terminals.

[0021] Figure 2 The text explains that the inverted residual structure of the MobileNetV2 network differs from the traditional residual structure in that it first increases the dimensionality and then decreases it. The left side shows the traditional residual structure, and the right side shows the inverted residual structure.

[0022] Figure 3 The mid-linear bottleneck layer consists of a dilated layer, a deep convolutional layer, and a projection layer.

[0023] The spatial sequence output by MobileNetV2 is input into the Markov chain module to establish the feature state transition probability matrix, model the correlation of features at adjacent time points, output the temporal features, and then fuse the temporal features with the spatial features to enhance the temporal dependence of fault features and the ability to express weak features, thereby improving the diagnostic robustness under varying operating conditions.

[0024] The overall structure of the model is as follows Figure 1 As shown: Signal input → Synchronous compressed wavelet time-frequency plot → MobileNetV2 feature extraction → Spatial features → Markov state transition modeling → Temporal features; Spatial features are fused with temporal features → fully connected layer → Softmax classification output.

[0025] The overall model parameter size is controlled within 5MB, and it supports quantization deployment on edge devices. It supports 32-bit floating-point to 8-bit integer model quantization, is compatible with automotive MCU / ARM embedded platforms, and has a single-sample inference time of less than 15ms. It enables online real-time diagnosis of bearing faults in new energy vehicles, solving the problems of high hardware dependence and difficulty in vehicle deployment of traditional deep learning models.

[0026] S4. Transfer Learning Model Training and Fine-tuning Employing a source-target domain transfer learning strategy: 1) Source domain: The MobileNetV2-Markov chain model is pre-trained using a public bearing fault dataset (such as CWRU / XJTU-SY) to learn general fault features; 2) Target domain: Using the measured dataset of new energy vehicle bearings collected by this invention, the parameters of the first 8 complete linear bottleneck layers of MobileNetV2 are frozen, and the subsequent networks (fully connected layers, Softmax classification output) and Markov chain modules are trained to complete the fine-tuning.

[0027] Training parameter settings: Batch size 32, initial learning rate 1×10⁻⁶ -3 The Adam optimizer is used, with cross-entropy loss as the loss function. Overfitting is prevented by early stopping and L2 regularization.

[0028] S5. Fault Diagnosis and Output The vibration signal of the bearing under test is used to generate a high-resolution time-frequency map in step S2. This map is then input into the trained transfer learning model. After forward inference, the model outputs diagnostic categories, including: normal state, inner ring pitting fault, outer ring spalling fault, rolling element wear fault, etc.

[0029] In this embodiment, the model achieves an average diagnostic accuracy of no less than 98.2% under varying speed and load conditions, and a single-sample inference time of less than 15ms, meeting the requirements for online fault diagnosis of bearings in new energy vehicles.

[0030] A fault diagnosis system for bearings in new energy vehicles based on synchronous compressed wavelet time-frequency graphs and MobileNetV2-Markov transfer learning includes: 1) Signal acquisition unit, used to acquire raw vibration signal data of new energy vehicle bearings in real time during operation; 2) The time-frequency diagram generation unit is electrically connected to the signal acquisition unit and is used to perform synchronous compressed wavelet transform to generate a high-resolution time-frequency diagram; 3) Fault diagnosis model unit, with a built-in trained MobileNetV2 network-Markov chain transfer learning model, used for feature extraction and classification reasoning of high-resolution time-frequency maps; 4) Results display unit, electrically connected to the diagnostic model unit, is used to visualize the fault type, level, and recommended maintenance information of the output bearing.

[0031] The signal acquisition unit includes: 1) Piezoelectric accelerometer with a sensitivity of not less than 100mV / g and a frequency response range of 1Hz-10kHz; 2) Data acquisition card, sampling frequency set to 10kHz-50kHz, supports multi-channel synchronous acquisition.

[0032] The fault diagnosis model unit also includes: 1) Model quantization module, used to convert model weights from 32-bit floating-point to 8-bit integer to reduce inference latency; 2) Edge deployment interface, supporting communication and docking with vehicle MCU or ARM embedded platform.

[0033] A computer-readable storage medium storing a computer program that, when executed by a processor, implements a new energy vehicle bearing fault diagnosis method based on synchronous compressed wavelet time-frequency graph and MobileNetV2 network-Markov chain transfer learning.

[0034] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned method for diagnosing bearings in new energy vehicles based on synchronous compressed wavelets and Markov models.

[0035] An in-vehicle edge diagnostic terminal includes the aforementioned computer-readable storage medium and the aforementioned computer device. The terminal device is integrated inside the electronic control unit (ECU) of a new energy vehicle and is used to realize real-time online monitoring and early warning of bearing faults.

Claims

1. A new energy vehicle bearing diagnostic method based on synchronous compressed wavelet and Markov, characterized in that, Includes the following steps: Step 1: Vibration signal acquisition: Set up a new energy vehicle bearing performance test bench and collect the original vibration signal data of the bearing under different speed, load and temperature conditions through an accelerometer. Step 2, Time-Frequency Image Conversion: Perform synchronous compressed wavelet transform on the acquired original vibration signal to convert the one-dimensional time domain signal into a high-resolution time-frequency image; Step 3: Construct a transfer learning model that integrates MobileNetV2 network and Markov chain: Use MobileNetV2 network as the backbone to extract spatial features of high-resolution time-frequency map, and introduce Markov chain model to model the state transition probability of spatial features, extract temporal features, and then fuse spatial features and temporal features and output the execution result after passing through fully connected layer and Softmax classification. Step 4, Model Training and Fine-tuning: First, the transfer learning model is pre-trained using a publicly available bearing fault dataset to initialize the parameters of the feature extractor; then, the model is fine-tuned using the new energy vehicle bearing test fault dataset constructed in this invention to optimize the Softmax classification output parameters. Step 5, Fault Diagnosis and Output: Input the high-resolution time-frequency graph generated by synchronous compressed wavelet transform of the vibration signal of the bearing to be tested into the transfer learning model trained and fine-tuned in Step 4. The model will automatically output the fault type and fault severity level of the corresponding bearing.

2. The new energy vehicle bearing diagnosis method based on synchronous compressed wavelet and Markov as described in claim 1, characterized in that, The synchronous compressed wavelet transform described in step 2 specifically includes: 1) Select Morlet wavelet as the basis function to perform continuous wavelet transform, generate wavelet coefficients, obtain the initial time-frequency matrix from the wavelet coefficients, and use the initial time-frequency matrix to compress and redistribute to obtain the time-frequency coefficients. 2) Calculate the instantaneous frequency of the wavelet coefficients and use this as a basis to regroup and compress the time-frequency coefficients, thus completing the compression of the time-frequency graph; 3) Normalize the compressed time-frequency image and crop the pixels to a uniform size to obtain a high-resolution time-frequency image.

3. The new energy vehicle bearing diagnosis method based on synchronous compressed wavelet and Markov as described in claim 1, characterized in that, The fine-tuning steps described in step 4 specifically include: 1) Using a transfer learning strategy, the parameters of the first 60%-70% of the complete linear bottleneck layers of the MobileNetV2 network are frozen, and only the fully connected layers, Softmax classification output, and Markov chain model are trained. 2) Set the initial learning rate to 0.001 and use the Adam optimizer and learning rate decay strategy; 3) Introduce the cross-entropy loss function as the training objective and introduce the early stopping method.

4. The new energy vehicle bearing diagnosis method based on synchronous compressed wavelet and Markov as described in claim 1, characterized in that, The fault types mentioned in step 5 include inner ring faults, outer ring faults, and rolling element faults.

5. The new energy vehicle bearing diagnosis method based on synchronous compressed wavelet and Markov as described in claim 1, characterized in that, The fault severity levels described in step 5 are specifically divided as follows: Minor fault: The characteristic amplitude is less than 1.5 times the normal average but higher than the threshold; Moderate fault: The characteristic amplitude is between 1.5 and 3 times the average value under normal conditions; Critical fault: Characteristic amplitude is more than 3 times the average value under normal conditions.

6. The new energy vehicle bearing diagnosis method based on synchronous compressed wavelet and Markov as described in claim 1, characterized in that, The aforementioned new energy vehicle bearings are deep groove ball bearings used in the drive motors or reducers of new energy vehicles, with bearing models ranging from the 6200 to the 6310 series.

Images