New energy automobile reducer vibration signal noise reduction and bearing fault diagnosis method

Abstract

The invention belongs to the technical field of new energy automobile key part fault diagnosis, and particularly relates to a new energy automobile reducer vibration signal noise reduction and bearing fault diagnosis method comprising the following steps: fault signal acquisition: building an experiment platform capable of simulating various complex vehicle-mounted working conditions, a vibration sensor which has anti-interference and adjustable sampling frequency and covers speed reducer operation and fault characteristic frequency is utilized to collect shell vibration signals when a speed reducer operates; in a signal acquisition link, an experimental platform capable of simulating various complex vehicle-mounted working conditions is established, a vibration sensor with high precision and strong anti-interference capability is equipped, and the vibration sensor is reasonably installed at a key part of a speed reducer shell, so that vibration signals reflecting the real running state of a speed reducer bearing can be comprehensively and accurately acquired; in the preprocessing process, various methods such as digital filtering, detrending processing and normalization processing are adopted, noise and interference components in the signals are effectively removed, and the quality and the characteristic definition of the signals are enhanced.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for key components of new energy vehicles, and in particular to methods for noise reduction of vibration signals in new energy vehicle reducers and fault diagnosis of bearings. Background Technology

[0002] With the booming development of the new energy vehicle industry, the performance stability and reliability of its core component—the reducer—plays a crucial role in the safe operation and performance of the entire vehicle. During long-term operation, reducers inevitably experience various faults due to complex working conditions, frequent starts and stops, and wear and tear on components, such as bearing wear and abnormal gear meshing. If these faults are not diagnosed and repaired promptly and accurately, they will not only affect the normal operation of the vehicle and reduce passenger comfort, but may also lead to serious safety accidents, causing significant economic losses and safety risks to users.

[0003] Currently, fault diagnosis of reducers in new energy vehicles mainly relies on traditional diagnostic methods, such as manual judgment, regular maintenance inspections, and techniques based on simple signal analysis. However, these methods have many limitations. Manual judgment is limited by the technical level and work experience of maintenance personnel, is highly subjective, and is prone to misdiagnosis and missed diagnosis; while regular maintenance inspections can prevent faults to some extent, they cannot monitor the operating status of the reducer in real time, making it difficult to detect and handle sudden faults in a timely manner; techniques based on simple signal analysis, such as Fourier transform, can perform spectral analysis of signals, but their ability to process non-stationary and nonlinear vibration signals is limited, making it difficult to accurately extract fault features, resulting in low diagnostic accuracy.

[0004] In the signal acquisition stage, traditional vibration sensors typically use wired transmission to transmit the acquired signals to processing equipment. This transmission method not only requires a large number of cables, increasing system complexity and cost, but is also susceptible to external electromagnetic interference, leading to signal quality degradation and affecting subsequent diagnostic results. Furthermore, the installation location and number of sensors are often determined based on experience, lacking scientific rigor and systematic approach, and may fail to comprehensively and accurately acquire vibration signals reflecting the true operating state of the reducer.

[0005] In signal processing, existing preprocessing methods are relatively simple, typically employing only basic filtering to remove noise. However, actual vibration signals may contain non-fault features such as trend terms in addition to noise interference. If these components are not removed, they will interfere with subsequent feature extraction and fault diagnosis. Furthermore, for complex non-stationary and nonlinear vibration signals, traditional signal processing methods struggle to effectively extract fault features, limiting the performance of diagnostic models and hindering accurate identification of the reducer's fault type and severity. Therefore, we provide a method for noise reduction of vibration signals from new energy vehicle reducers and for bearing fault diagnosis. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a method for noise reduction of vibration signals in new energy vehicle reducers and for diagnosing bearing faults, thereby more accurately resolving the issues raised in the background section.

[0007] This invention is achieved through the following technical solution: The invention proposes a method for noise reduction of vibration signals in the reducer of new energy vehicles and for diagnosing bearing faults, including the following steps: Fault signal acquisition: An experimental platform capable of simulating various complex vehicle operating conditions was built. Vibration sensors with anti-interference capabilities, adjustable sampling frequency, and coverage of the operating and fault characteristic frequencies of the reducer were used to acquire the vibration signal of the reducer housing during operation. Signal preprocessing: The acquired signal is fully processed, and advanced digital filtering algorithms are used to remove high-frequency noise and low-frequency interference. Detrending algorithms are used to eliminate long-term trend terms, and normalization is then performed to adjust the amplitude to an appropriate range. Hybrid Model Construction and Training: A hybrid model based on Stacked Denoising Autoencoder Network (SDAE) and Bidirectional Gated Recurrent Neural Network (BiGRU) is constructed. The number of nodes in each layer of SDAE is optimized according to signal characteristics and diagnostic requirements. After full connection, its encoder is used to train noisy samples to obtain robust features, and the output is used as input to BiGRU to extract temporal features. Mini-batch gradient descent combined with an improved Adam optimization algorithm is used for training, and dynamic Dropout is used to prevent overfitting. Algorithm accuracy verification and optimization: The diagnostic results of this algorithm are compared with those of various other deep network algorithms using different test samples. The model and algorithm parameters are dynamically adjusted and optimized based on the evaluation of accuracy, recall, F1 score, etc. Operating status identification and decision-making: The optimized Softmax classifier is used to identify the bearing operating status, and maintenance decisions are made based on the diagnostic results.

[0008] Furthermore, a vibration signal noise reduction and bearing fault diagnosis system for a new energy vehicle reducer includes: Signal acquisition module: includes hardware equipment for building an experimental platform, capable of accurately simulating various different vehicle operating conditions, as well as a high-precision, high-sensitivity vibration sensor. The vibration sensor has strong anti-interference ability, adjustable sampling frequency, and the sampling frequency range covers the operating frequency and fault characteristic frequency of the new energy vehicle reducer. It is used to collect the vibration signal of the housing of the new energy vehicle reducer during operation. Preprocessing module: It is equipped with a dedicated signal processing unit that can perform comprehensive and detailed preprocessing operations on the vibration signals acquired by the signal acquisition module. The preprocessing operations include, but are not limited to, using advanced digital filtering algorithms to remove high-frequency noise and low-frequency interference from the signal, using detrending algorithms to eliminate long-term trend terms from the signal, and performing normalization processing to adjust the amplitude range of the signal to a pre-set appropriate range. Hybrid Model Construction and Training Module: This module is capable of constructing a hybrid model based on Stacked Denoising Autoencoder Network (SDAE) and Bidirectional Gated Recurrent Neural Network (BiGRU). The number of nodes in the input, hidden, and output layers of the SDAE network structure can be optimized according to the characteristics of the collected vibration signals and the requirements of fault diagnosis. Each layer is fully connected. This module can use the encoder of SDAE to train samples with added noise of specific types and intensities to obtain robust features. The output of SDAE is used as the input of BiGRU to extract temporal features. The model is trained using mini-batch gradient descent combined with an improved Adam optimization algorithm. At the same time, dynamically adjusted Dropout technology is used to prevent model overfitting. Algorithm verification and optimization module: This module is used to collect different and widely representative vehicle reducer bearing fault signals from the training samples as test samples. The test samples are then input into the trained hybrid model to obtain diagnostic results. The results of this algorithm are compared with the diagnostic results of various other deep network algorithms. The performance advantages of this algorithm are comprehensively evaluated based on multiple evaluation indicators such as accuracy, recall, and F1 score. Based on the evaluation results, the network structure parameters of the hybrid model and the optimization algorithm parameters are dynamically adjusted and optimized. Operating status identification and decision-making module: Includes an optimized and trained Softmax classifier, which can identify the operating status of reducer bearings in test samples. The operating status includes, but is not limited to, normal status, different types and degrees of fault status, and formulates detailed and targeted maintenance decisions based on the diagnostic results. Maintenance decisions include, but are not limited to, immediate maintenance, planned maintenance, and monitoring operation.

[0009] Furthermore, a new energy vehicle reducer vibration signal noise reduction and bearing fault diagnosis software, based on the method described in claim 1, utilizes a computer programming language for functional design, including: Data input module: It has an intuitive and convenient file selection control, which allows users to select the collected vibration signal file of the new energy vehicle reducer housing from local storage devices or external data sources. The file format includes, but is not limited to, common audio file formats or specific data file formats. Model training module: It has a training button to trigger the training process of the hybrid model built on Stacked Denoising Autoencoder Network (SDAE) and Bidirectional Gated Recurrent Neural Network (BiGRU). During the training process, the mini-batch gradient descent method is combined with the improved Adam optimization algorithm and the dynamically adjusted Dropout technology is used. At the same time, a visual training process monitoring interface is provided to display the changes of key indicators such as loss function value and accuracy in real time during the training process. Fault diagnosis module: It is equipped with a diagnosis button to start the fault diagnosis algorithm. The input vibration signal is input into the trained hybrid model for diagnosis. During the diagnosis process, the diagnosis progress and status information can be displayed in real time. Results display module: Displays fault diagnosis results in a variety of intuitive and easy-to-understand forms, including but not limited to displaying the probability distribution of different working states in bar charts, displaying the time-domain and frequency-domain characteristics of signals in line charts, and providing detailed descriptions of the fault type, location, and severity in text form, as well as specific maintenance suggestions, including but not limited to information on maintenance time, cost, and required parts; Data storage and management module: Used to store the collected vibration signals, parameters during model training, and fault diagnosis results. The data storage format adopts a standardized database format (such as SQLite, MySQL, etc.) to facilitate subsequent querying, analysis, and sharing. It also provides data backup and recovery functions to ensure data security and integrity.

[0010] Furthermore, in the fault signal acquisition step, the experimental platform is also equipped with other types of sensors such as temperature sensors and speed sensors to collect auxiliary information such as temperature and speed during the operation of the reducer. This auxiliary information, together with the vibration signal, serves as the basis for subsequent analysis and diagnosis, so as to improve the accuracy and reliability of fault diagnosis.

[0011] Furthermore, in the hybrid model construction and training steps, the improved Adam optimization algorithm dynamically adjusts the momentum factor and learning rate in the original Adam algorithm, and adaptively adjusts the optimization direction and step size according to the changes in the loss function during training, so as to improve the convergence speed and training effect of the model.

[0012] Furthermore, in the algorithm accuracy verification and optimization step, the various other deep network algorithms are trained and tested using the same training and test sets to ensure the fairness and accuracy of the comparison results. At the same time, when conducting the comparison evaluation, in addition to indicators such as accuracy, recall, and F1 score, performance indicators such as model training time and inference time are also considered.

[0013] Furthermore, the vibration sensor in the signal acquisition module transmits the acquired vibration signal to the preprocessing module wirelessly. The wireless transmission method includes, but is not limited to, Wi-Fi, Bluetooth, ZigBee, etc., to improve the flexibility and convenience of data transmission and reduce the interference and inconvenience caused by cable connections.

[0014] Furthermore, the preprocessing module also includes a signal enhancement unit, which is used to further enhance the signal after preliminary preprocessing to improve the signal-to-noise ratio and feature clarity. The signal enhancement algorithm includes, but is not limited to, wavelet transform and empirical mode decomposition (EMD).

[0015] Furthermore, the results display module also has a data export function, allowing users to export diagnostic results in common file formats, which facilitates subsequent report writing, data analysis and sharing.

[0016] Furthermore, the software also has remote monitoring and diagnostic functions, allowing users to monitor the operating status of the reducer of new energy vehicles in real time and diagnose faults through network connection, thereby improving the timeliness and efficiency of fault handling.

[0017] The beneficial effects of this invention are: 1. In the signal acquisition stage, an experimental platform capable of simulating various complex vehicle operating conditions is built, equipped with high-precision vibration sensors with strong anti-interference capabilities, and reasonably installed in key parts of the reducer housing. This allows for the comprehensive and accurate acquisition of vibration signals that reflect the actual operating state of the reducer bearings. During the preprocessing stage, various methods such as digital filtering, detrending processing, and normalization are used to effectively remove noise and interference components from the signal, thereby enhancing the signal quality and feature clarity.

[0018] 2. In terms of signal acquisition, the vibration sensor uses wireless transmission to transmit the acquired vibration signals to the preprocessing module, reducing interference and inconvenience caused by cable connections and improving the flexibility and convenience of data transmission. At the same time, the results display module has a data export function, and users can choose the format of the exported diagnostic results according to their needs, such as PDF format for generating a complete report, and Excel format for convenient subsequent data processing and analysis, meeting the usage needs of different users in different scenarios.

[0019] 3. The remote monitoring platform adopts a strict identity authentication and authorization mechanism. Users need to undergo multi-factor authentication when logging in, and different users are set with different permission levels. Only authorized users can access specific data and functions, which effectively prevents unauthorized access and data leakage and ensures the security of the system. In terms of data storage, a distributed storage system is used to store data on multiple nodes, which improves the reliability and availability of data. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention. Detailed Implementation

[0021] To more clearly and completely illustrate the technical solution of the present invention, the present invention will be further described below in conjunction with the accompanying drawings. Example

[0022] like Figure 1 As shown, an embodiment of the present invention proposes a method for noise reduction of vibration signals in new energy vehicle reducers and for bearing fault diagnosis. (I) Signal Acquisition An experimental platform capable of accurately simulating various complex in-vehicle operating conditions was constructed. This platform includes an adjustable road surface simulation device capable of simulating different road conditions such as bumpy roads, smooth roads, and high-speed roads. For example, bumpy roads can be simulated by setting up protrusions and depressions of different shapes and arrangements, smooth roads can be simulated by using flat metal plates, and high-speed roads can be simulated by using tracks with specific textures and slopes.

[0023] Equipped with an adjustable speed drive, the reducer can operate at different speeds, such as low speed (e.g., 10-30km / h), medium speed (e.g., 30-60km / h), and high speed (e.g., 60-120km / h).

[0024] The load adjustment device can simulate different load conditions such as no load, half load (e.g., loading 50% of the rated load of the reducer), and full load (loading the rated load of the reducer).

[0025] Installing a timing device allows for precise recording of the reducer's operating time, including short periods of operation (e.g., 1-10 minutes) and long periods of continuous operation (e.g., several hours).

[0026] High-precision, high-sensitivity vibration sensors with strong anti-interference capabilities are selected, such as sensors with shielded housings and special filtering circuits. The sampling frequency is adjustable, and the adjustment range covers the operating frequency of new energy vehicle reducers (e.g., 0-5000Hz) and fault characteristic frequencies (determined according to the specific structure and fault type of the reducer, generally within a specific frequency band).

[0027] Vibration sensors are installed in critical parts of the reducer housing, such as near the bearing, to ensure that vibration signals reflecting the bearing's operating status are accurately collected.

[0028] (ii) Pretreatment Advanced digital filtering algorithms, such as Butterworth filters, are used to remove high-frequency noise and low-frequency interference from signals. For example, by setting an appropriate cutoff frequency, high-frequency noise above a certain frequency (e.g., 3000Hz) and low-frequency interference below a certain frequency (e.g., 10Hz) can be filtered out.

[0029] Detrending algorithms, such as polynomial fitting detrending, are used to eliminate long-term trend terms in a signal. For example, the trend of the signal is approximated by fitting a quadratic polynomial, and then the trend term is subtracted from the original signal to obtain the detrended signal.

[0030] Normalization is performed to adjust the amplitude range of the signal to a pre-defined appropriate interval, such as [-1, 1]. For example, the maximum-minimum normalization method is used to map the minimum value of the signal to -1, the maximum value to 1, and other values ​​to a proportional mapping.

[0031] (III) Model Building and Training A hybrid model based on Stacked Denoising Autoencoder Network (SDAE) and Bidirectional Gated Recurrent Neural Network (BiGRU) was constructed. The number of nodes in the input, hidden, and output layers of the SDAE was optimized based on the characteristics of the acquired vibration signals and the requirements for fault diagnosis. For example, the number of nodes in the input layer was set to be related to the number of vibration signal sampling points (e.g., 1024), the number of nodes in the hidden layer was determined by experimentally trying different values ​​(e.g., 256, 512, etc.), and the number of nodes in the output layer was set according to the number of fault types (e.g., 4 nodes for 4 fault types). All layers were fully connected.

[0032] During training, the encoder is first trained on samples with added noise of specific types and intensities using the SDAE encoder. For example, Gaussian white noise is added, with the noise intensity set to a certain proportion of the signal standard deviation (e.g., 0.1 times). Through training, the encoder learns robust and representative features.

[0033] The output of SDAE is used as the input to BiGRU, leveraging BiGRU's ability to extract temporal features by simultaneously utilizing past and future information. Model training employs a mini-batch gradient descent method combined with an improved Adam optimization algorithm. For example, the mini-batch size is set to 32. The improved Adam optimization algorithm dynamically adjusts the momentum factor and learning rate, adaptively adjusting the optimization direction and step size based on changes in the loss function during training. Simultaneously, a dynamically adjusted Dropout technique is used to prevent overfitting; for instance, the initial Dropout probability is set to 0.5, and dynamically adjusted during training based on validation set performance.

[0034] (iv) Verification and Optimization Different vehicle reducer bearing fault signals, which are distinct from the training samples and broadly representative, were collected as test samples. For example, fault signals were collected under different experimental conditions (such as different road conditions, speeds, and load combinations).

[0035] The test samples are input into the trained hybrid model to obtain diagnostic results. The results of this algorithm are then compared with those of various other deep network algorithms (such as Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and their improved versions). The performance advantages of this algorithm are comprehensively evaluated based on multiple evaluation metrics, including accuracy, recall, and F1 score. For example, accuracy = number of correctly diagnosed samples / total number of samples; recall = number of correctly diagnosed samples of a certain type of fault / actual number of samples of that type of fault; F1 score = 2 * accuracy * recall / (accuracy + recall).

[0036] Based on the evaluation results, the network structure parameters (such as the number of layers, number of nodes, and connection methods) and optimization algorithm parameters (such as learning rate and momentum factor) of the hybrid model are dynamically adjusted and optimized. For example, if the model is found to have a low recall rate in diagnosing a certain type of fault, the number of hidden layer nodes related to that type of fault can be appropriately increased or the learning rate can be adjusted.

[0037] (v) State recognition and decision making An optimized and trained Softmax classifier is used to identify the operating state of the reducer bearings in the test samples. Operating states include normal conditions and different types and degrees of fault conditions (such as inner race fault, outer race fault, rolling element fault, etc.). For example, the Softmax classifier outputs the probability of each operating state, and the state with the highest probability is selected as the diagnostic result.

[0038] Detailed and targeted maintenance decisions can be made based on the diagnostic results. For example, if a minor inner race fault is diagnosed, a decision can be made to monitor operation and regularly check the development of the fault; if a serious outer race fault is diagnosed, a decision can be made to repair immediately and arrange for professional personnel to perform repair or replacement.

[0039] Furthermore, during the signal acquisition process described above, in addition to vibration sensors, the experimental platform also installed other types of sensors, such as temperature sensors and speed sensors. Temperature sensors are installed on the surface or key internal parts of the reducer housing to collect temperature information during reducer operation; speed sensors are installed on the input or output shaft of the reducer to collect speed information. This auxiliary information, along with the vibration signals, is stored in the data acquisition system and serves as an important reference in subsequent analysis and diagnosis to improve the accuracy and reliability of fault diagnosis. For example, when the vibration signal shows abnormalities, combining temperature and speed information can more accurately determine the type and cause of the fault. For instance, high temperature may indicate poor lubrication or excessive bearing wear, while abnormal vibration at high speed may be related to gear meshing problems.

[0040] Furthermore, during the construction and training of the hybrid model, the improved Adam optimization algorithm dynamically adjusts the momentum factor and learning rate in the original Adam algorithm. The specific implementation is as follows: Initialization parameters: Set the initial momentum factor =0.9, =0.999, initial learning rate =0.001.

[0041] Calculate the gradient: In each iteration, calculate the gradient of the loss function with respect to the model parameters.

[0042] Update momentum estimates: Update the first-order momentum estimate m and the second-order momentum estimate v based on the gradient, as shown in the following formula: ; ; in, It is the gradient of the current iteration. and It is the momentum estimate from the previous iteration.

[0043] Dynamically adjust the momentum factor and learning rate: Adjust the momentum factor and learning rate based on changes in the loss function. For example, if the loss function decreases slowly over several iterations, the learning rate can be appropriately reduced. Simultaneously adjust the momentum factor and This can be used to change the direction of optimization. Specific adjustment strategies can be determined through experiments, such as setting a threshold. When the decrease in the loss function is less than the threshold, the learning rate is multiplied by a decay factor (such as 0.9), and the momentum factor is fine-tuned at the same time.

[0044] Parameter Update: Update the model parameters based on the adjusted momentum factor and learning rate, using the following formula:

[0045] in, These are the model parameters for the current iteration. These are the model parameters from the previous iteration. It is a very small constant (such as 1e-8) to prevent division by zero errors.

[0046] Furthermore, during signal acquisition, to enhance the flexibility and convenience of data transmission and reduce interference and inconvenience caused by cable connections, the vibration sensor transmits the acquired vibration signals to the preprocessing module wirelessly. For example, a Wi-Fi transmission method can be used, integrating a Wi-Fi module at the vibration sensor end to encode the acquired signals and transmit them wirelessly. A Wi-Fi receiver is set up at the preprocessing module end to receive and decode the signals. Simultaneously, to ensure the stability and security of data transmission, encryption algorithms, such as AES encryption, can be used to encrypt the transmitted data.

[0047] Furthermore, in the preprocessing module, in addition to the preprocessing operations described above, a signal enhancement unit is also included. For example, wavelet transform is used for signal enhancement. The specific steps are as follows: Wavelet decomposition: Selecting appropriate wavelet basis functions (such as Daubechies wavelets) to perform multi-level wavelet decomposition on the pre-processed signal, decomposing the signal into sub-signals of different frequency bands.

[0048] Thresholding: The decomposed sub-signals are subjected to thresholding to remove noise components. For example, a soft thresholding method can be used, setting an appropriate threshold to set coefficients below the threshold to zero and shrinking coefficients above the threshold.

[0049] Wavelet reconstruction: The sub-signal after thresholding is reconstructed using wavelets to obtain the enhanced signal. Wavelet transform can effectively improve the signal-to-noise ratio and feature clarity of the signal, making subsequent fault diagnosis more accurate.

[0050] Furthermore, during the algorithm accuracy verification and optimization process, to ensure the fairness and accuracy of the comparison results, various other deep network algorithms (such as CNN, LSTM, and their improved algorithms) are trained and tested using the same training and test sets. For example, the collected vibration signals are divided into training and test sets according to a certain ratio (e.g., 70% for training and 30% for testing), and all algorithms are trained using the same training set and evaluated using the same test set. Simultaneously, in the comparative evaluation, in addition to metrics such as accuracy, recall, and F1 score, performance metrics such as model training time and inference time are also considered. For example, the training time of each algorithm on the training set and the inference time required to perform a complete diagnosis on the test set are recorded to comprehensively evaluate the algorithm's merits.

[0051] Furthermore, the results display module includes a data export function to facilitate subsequent report writing, data analysis, and sharing. For example, a data export interface could be developed, allowing users to choose the format for exporting diagnostic results, such as PDF or Excel. When the user chooses to export in PDF format, the text and chart information from the diagnostic results would be integrated to generate a complete PDF report; when the user chooses to export in Excel format, the data from the diagnostic results would be exported in tabular form, facilitating further data processing and analysis.

[0052] Furthermore, to achieve remote monitoring and diagnostic functions for the software, a network connection allows users to monitor the operating status and diagnose faults of the reducer in new energy vehicles in real time from different locations. For example, a web-based remote monitoring platform can be developed. Users can log in to the platform via a browser from anywhere with an internet connection, and access the reducer's real-time operating data and diagnostic results after entering their username and password. The platform uses real-time data transmission technology to transmit data such as vibration signals, temperature, and speed from the reducer to the server in real time. The server processes and analyzes the data and displays the results to the user. Simultaneously, the platform also has an alarm function; when an abnormality is detected, it promptly sends alarm information to the user, such as SMS or email notifications.

[0053] Furthermore, the remote monitoring platform employs an identity authentication and authorization mechanism. Users are required to undergo strict identity verification upon login, including multi-factor authentication such as username, password, and verification code. Different access levels are set for different users, ensuring that only authorized users can access specific data and functions. For data storage, a distributed storage system is used, storing data across multiple nodes to improve reliability and availability. Regular data backups are also performed, with backups stored off-site to prevent data loss.

[0054] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for noise reduction of vibration signals and diagnosis of bearing faults in new energy vehicle reducers, characterized in that, Includes the following steps: Fault signal acquisition: An experimental platform capable of simulating various complex vehicle operating conditions was built. Vibration sensors with anti-interference capabilities, adjustable sampling frequency, and coverage of the operating and fault characteristic frequencies of the reducer were used to acquire the vibration signal of the reducer housing during operation. Signal preprocessing: The acquired signal is fully processed, and advanced digital filtering algorithms are used to remove high-frequency noise and low-frequency interference. Detrending algorithms are used to eliminate long-term trend terms, and normalization is then performed to adjust the amplitude to an appropriate range. Hybrid Model Construction and Training: A hybrid model based on Stacked Denoising Autoencoder Network (SDAE) and Bidirectional Gated Recurrent Neural Network (BiGRU) is constructed. The number of nodes in each layer of SDAE is optimized according to signal characteristics and diagnostic requirements. After full connection, its encoder is used to train noisy samples to obtain robust features, and the output is used as input to BiGRU to extract temporal features. Mini-batch gradient descent combined with an improved Adam optimization algorithm is used for training, and dynamic Dropout is used to prevent overfitting. Algorithm accuracy verification and optimization: The diagnostic results of this algorithm are compared with those of various other deep network algorithms using different test samples. The model and algorithm parameters are dynamically adjusted and optimized based on the evaluation of accuracy, recall, F1 score, etc. Operating status identification and decision-making: The optimized Softmax classifier is used to identify the bearing operating status, and maintenance decisions are made based on the diagnostic results.

2. The vibration signal noise reduction and bearing fault diagnosis system for a new energy vehicle reducer according to claim 1, characterized in that, include: Signal acquisition module: includes hardware equipment for building an experimental platform, capable of accurately simulating various different vehicle operating conditions, as well as a high-precision, high-sensitivity vibration sensor. The vibration sensor has strong anti-interference ability, adjustable sampling frequency, and the sampling frequency range covers the operating frequency and fault characteristic frequency of the new energy vehicle reducer. It is used to collect the vibration signal of the housing of the new energy vehicle reducer during operation. Preprocessing module: It is equipped with a dedicated signal processing unit that can perform comprehensive and detailed preprocessing operations on the vibration signals acquired by the signal acquisition module. The preprocessing operations include, but are not limited to, using advanced digital filtering algorithms to remove high-frequency noise and low-frequency interference from the signal, using detrending algorithms to eliminate long-term trend terms from the signal, and performing normalization processing to adjust the amplitude range of the signal to a pre-set appropriate range. Hybrid Model Construction and Training Module: This module is capable of constructing a hybrid model based on Stacked Denoising Autoencoder Network (SDAE) and Bidirectional Gated Recurrent Neural Network (BiGRU). The number of nodes in the input, hidden, and output layers of the SDAE network structure can be optimized according to the characteristics of the collected vibration signals and the requirements of fault diagnosis. Each layer is fully connected. This module can use the encoder of SDAE to train samples with added noise of specific types and intensities to obtain robust features. The output of SDAE is used as the input of BiGRU to extract temporal features. The model is trained using mini-batch gradient descent combined with an improved Adam optimization algorithm. At the same time, dynamically adjusted Dropout technology is used to prevent model overfitting. Algorithm verification and optimization module: This module is used to collect different and widely representative vehicle reducer bearing fault signals from the training samples as test samples. The test samples are then input into the trained hybrid model to obtain diagnostic results. The results of this algorithm are compared with the diagnostic results of various other deep network algorithms. The performance advantages of this algorithm are comprehensively evaluated based on multiple evaluation indicators such as accuracy, recall, and F1 score. Based on the evaluation results, the network structure parameters of the hybrid model and the optimization algorithm parameters are dynamically adjusted and optimized. Operating status identification and decision-making module: Includes an optimized and trained Softmax classifier, which can identify the operating status of reducer bearings in test samples. The operating status includes, but is not limited to, normal status, different types and degrees of fault status, and formulates detailed and targeted maintenance decisions based on the diagnostic results. Maintenance decisions include, but are not limited to, immediate maintenance, planned maintenance, and monitoring operation.

3. A software for noise reduction of vibration signals and diagnosis of bearing faults in a new energy vehicle reducer, characterized in that, Based on the method described in claim 1, functional design is performed using a computer programming language, including: Data input module: It has an intuitive and convenient file selection control, which allows users to select the collected vibration signal file of the new energy vehicle reducer housing from local storage devices or external data sources. The file format includes, but is not limited to, common audio file formats or specific data file formats. Model training module: It has a training button to trigger the training process of the hybrid model built on Stacked Denoising Autoencoder Network (SDAE) and Bidirectional Gated Recurrent Neural Network (BiGRU). During the training process, the mini-batch gradient descent method is combined with the improved Adam optimization algorithm and the dynamically adjusted Dropout technology is used. At the same time, a visual training process monitoring interface is provided to display the changes of key indicators such as loss function value and accuracy in real time during the training process. Fault diagnosis module: It is equipped with a diagnosis button to start the fault diagnosis algorithm. The input vibration signal is input into the trained hybrid model for diagnosis. During the diagnosis process, the diagnosis progress and status information can be displayed in real time. Results display module: Displays fault diagnosis results in a variety of intuitive and easy-to-understand forms, including but not limited to displaying the probability distribution of different working states in bar charts, displaying the time-domain and frequency-domain characteristics of signals in line charts, and providing detailed descriptions of the fault type, location, and severity in text form, as well as specific maintenance suggestions, including but not limited to information on maintenance time, cost, and required parts; Data storage and management module: Used to store the collected vibration signals, parameters during model training, and fault diagnosis results. The data storage format adopts a standardized database format (such as SQLite, MySQL, etc.) to facilitate subsequent querying, analysis, and sharing. It also provides data backup and recovery functions to ensure data security and integrity.

4. The method for noise reduction of vibration signals and diagnosis of bearing faults in new energy vehicle reducers according to claim 1, characterized in that, In the fault signal acquisition step, the experimental platform is also equipped with other types of sensors such as temperature sensors and speed sensors to collect auxiliary information such as temperature and speed during the operation of the reducer. This auxiliary information, together with the vibration signal, serves as the basis for subsequent analysis and diagnosis to improve the accuracy and reliability of fault diagnosis.

5. The method for noise reduction of vibration signals and bearing fault diagnosis of new energy vehicle reducers according to claim 1, characterized in that, In the hybrid model construction and training steps, the improved Adam optimization algorithm dynamically adjusts the momentum factor and learning rate in the original Adam algorithm, and adaptively adjusts the optimization direction and step size according to the changes in the loss function during training, so as to improve the convergence speed and training effect of the model.

6. The method for noise reduction of vibration signals and bearing fault diagnosis of new energy vehicle reducers according to claim 1, characterized in that, In the algorithm accuracy verification and optimization steps, the various other deep network algorithms are trained and tested using the same training and test sets to ensure the fairness and accuracy of the comparison results. In addition to accuracy, recall, F1 score and other metrics, the model's training time, inference time and other performance metrics are also considered when conducting the comparison evaluation.

7. The new energy vehicle reducer vibration signal noise reduction and bearing fault diagnosis system according to claim 2, characterized in that, The vibration sensor in the signal acquisition module transmits the acquired vibration signal to the preprocessing module wirelessly. The wireless transmission method includes, but is not limited to, Wi-Fi, Bluetooth, ZigBee, etc., to improve the flexibility and convenience of data transmission and reduce the interference and inconvenience caused by cable connections.

8. The vibration signal noise reduction and bearing fault diagnosis system for new energy vehicle reducers according to claim 2, characterized in that, The preprocessing module also includes a signal enhancement unit, which is used to further enhance the preprocessed signal to improve the signal-to-noise ratio and feature clarity. The signal enhancement algorithm includes, but is not limited to, wavelet transform and empirical mode decomposition (EMD).

9. The software for noise reduction of vibration signals and diagnosis of bearing faults in new energy vehicle reducers according to claim 3, characterized in that, The results display module also has a data export function, allowing users to export diagnostic results in common file formats, which facilitates subsequent report writing, data analysis and sharing.

10. The software for noise reduction of vibration signals and diagnosis of bearing faults in new energy vehicle reducers according to claim 3, characterized in that, The software also has remote monitoring and diagnostic functions, allowing users to monitor the operating status of the reducer of new energy vehicles in real time and diagnose faults through network connection, thereby improving the timeliness and efficiency of fault handling.

Images