A Fault Diagnosis Method for Rolling Bearings Based on Time-Frequency Enhancement and Generative Learning
This method for diagnosing rolling bearing faults using time-frequency enhancement and generative learning leverages GAF and CVAE models to augment data and incorporates a lightweight deep learning network. This approach addresses the issues of high manpower and time requirements inherent in traditional methods, achieving efficient and accurate fault diagnosis suitable for real-time monitoring and early warning.
Patent Information
- Application Number
- CN202410744936.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-11
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-06-11
AI Technical Summary
Traditional rolling bearing fault diagnosis methods require a lot of manpower and time, and their diagnostic accuracy is insufficient under conditions of small sample data, making it difficult to achieve real-time monitoring and early warning.
A small-sample rolling bearing fault diagnosis method based on time-frequency enhancement and generative learning is adopted. The one-dimensional vibration signal is converted into a GADF time-frequency feature map by the GAF algorithm, and the data is augmented by the CVAE model. Finally, the MobileNet-v3 lightweight deep learning model is used for fault feature learning and classification.
It improves the accuracy and practicality of fault diagnosis in small sample situations, reduces the computational resource requirements, is suitable for real-time monitoring and fault early warning systems, and has broad application prospects.
Smart Images

Figure CN118626929B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent manufacturing and industrial big data technology, and relates to a small-sample rolling bearing fault diagnosis method based on time-frequency enhancement and generative learning. Background Technology
[0002] With the development of modern industry, rolling bearings, as key components widely used in various mechanical equipment, are crucial to the normal operation of these devices due to their stability and performance. Rolling bearings are a common type that withstands rotational and sliding loads. However, rolling bearings can fail, such as due to wear, fatigue, or excessive clearance, affecting normal equipment operation and even leading to downtime and equipment damage. Therefore, effective diagnosis of rolling bearing faults is of great significance. Timely and accurate identification of rolling bearing faults can prevent serious accidents, reduce equipment downtime, and improve production efficiency and economic benefits.
[0003] Diagnosing and predicting rolling bearing failures also helps in developing reasonable maintenance plans and replacement strategies, reducing operating and maintenance costs. Applications include wind power generation equipment, automobiles, aircraft, high-speed trains, and industrial machinery, where high performance and stability requirements for rolling bearings are necessary, and failures can lead to serious economic losses and safety hazards. Traditional rolling bearing failure diagnosis requires significant manpower and time, while modern methods utilize sensors and data acquisition technologies, along with signal processing and modal analysis, to achieve real-time monitoring and analysis, assessing the health status of rolling bearings and predicting their lifespan. This avoids downtime for maintenance and equipment damage, improving production efficiency and economic benefits. Summary of the Invention
[0004] In view of this, the purpose of this invention is to provide a small-sample rolling bearing fault diagnosis method based on time-frequency enhancement and generative learning.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] A few-sample rolling bearing fault diagnosis method based on time-frequency enhancement and generative learning includes the following steps:
[0007] S1: Collect rolling bearing fault data and preprocess it to establish a rolling bearing fault dataset;
[0008] S2: The one-dimensional vibration fault signal is converted into a GADF time-frequency feature map using the GAF algorithm, and training sample data, validation sample data and test sample data are divided.
[0009] S3: Use the CVAE model to perform generative data sample augmentation on small training data;
[0010] S4: Construct the GAF-CVAE-MobileNet-v3 rolling bearing fault diagnosis model; train and optimize the model using training set sample data and validation set sample data;
[0011] S5: Use the test set sample data to verify the rolling bearing fault diagnosis model. If it does not meet the requirements, return to step S4 until the result meets the requirements.
[0012] Furthermore, step S1 specifically includes the following steps:
[0013] S11: Data Collection: Vibration signal data of the rolling bearing device in operation is captured in real time by sensors and stored in the data acquisition system;
[0014] S12: Data Preprocessing: For the acquired one-dimensional vibration data, let the time series be X = {x1, x2, ..., x...} n}, x i For the i-th sampled signal, the one-dimensional time-series signal is scaled to the range [-1, 1] using a normalization algorithm, following the formula:
[0015]
[0016] In the formula For x i Normalized values, i = 1, 2, ..., n;
[0017] The scaled time series was then converted to polar coordinates and represented as follows:
[0018]
[0019] In the formula The normalized values of one-dimensional time series data are encoded into angles in a polar coordinate image using a cosine function. t i The timestamp is encoded as the radius r in the polar coordinate image, and N is a constant factor for regularization and the coordinate system generation space.
[0020] Furthermore, step S2 specifically includes the following steps:
[0021] S21: Time-Frequency Conversion: Using the Gramian Angular Field (GAF) algorithm, the preprocessed one-dimensional vibration signal is mapped to a two-dimensional time-frequency space to generate a GADF time-frequency feature map, following the formula below:
[0022]
[0023] In the formula This is the angle encoded into a polar coordinate image using a cosine function;
[0024] S22: Sample partitioning: The generated GADF time-frequency feature map is divided into training samples, validation samples and test samples according to experimental requirements.
[0025] Furthermore, step S3 specifically includes the following steps:
[0026] S31: Encoder Operation: The encoder transforms the input data into a representation in a latent space, which contains the key features of the data; in CVAE, the encoder receives the input data and condition variables as inputs and outputs the mean and variance of the latent space; a probability distribution is defined to represent the latent representation of the input data under given conditions;
[0027] S32: Sampling: Based on the parameters output by the encoder, samples are taken from the latent space, introducing randomness, and following the formula:
[0028] μ = f φ (x)
[0029] σ 2 =g φ (x)
[0030] Where f φ and g φ These represent the output layers of the encoder network, respectively. x is used to output the mean and variance of the latent space, and is the input data.
[0031] z=μ+σ·∈
[0032] In the reparameterization process, z represents the latent variable, and ∈ is random noise drawn from the standard normal distribution;
[0033] S33: Decoder Operation: The function of the decoder is to transform the representation of the latent space back into the original data space and generate new data; in CVAE, the decoder receives the sampling points and condition variables of the latent space as input and outputs the generated data.
[0034] S34: Loss Function Evaluation: In the CVAE model, reconstruction loss is used to measure the similarity between the input and generated data, and KL divergence loss is used to measure the difference between the latent spatial distribution and the prior distribution, following the formula:
[0035]
[0036] in, For the reconstruction loss, KL is the Kullback-Leibler divergence. Combining these two loss functions forms the total loss function of the model, and the model is trained by minimizing the total loss.
[0037] S35: Data Augmentation: After training the CVAE model, a large amount of new data is generated by changing the condition variables and sampling of the latent space for data augmentation.
[0038] Furthermore, step S4 specifically includes the following steps:
[0039] S41: Model Architecture Design: Determine the architecture of the GAF-CVAE-MobileNet-v3 model, including GAF time-frequency feature extraction, CVAE data augmentation, and the integration of the MobileNet-v3 lightweight deep learning model; the GAF algorithm is used to convert the original one-dimensional vibration signal into a two-dimensional time-frequency feature map, CVAE is responsible for generative data augmentation to expand the small sample dataset, and MobileNet-v3 is used to realize fault feature learning and classification;
[0040] S42: Feature Learning and Optimization: Using GAF time-frequency feature maps as input, the MobileNet-v3 model is trained with CVAE-enhanced data. During this process, the deep learning model automatically learns and extracts fault features, and optimization algorithms such as backpropagation and gradient descent are used to adjust the model parameters to minimize classification error.
[0041] S43: Hyperparameter Tuning: Adjust the learning rate, batch size, number of convolutional layers, and configure hyperparameters. Use the Adam optimizer to adjust the learning rate of each parameter, following the formula below:
[0042] m t =β1·m t-1 +(1-β1)·g t
[0043]
[0044] Where g t The gradient at real time step t, m t and v t These are the first and second moment estimates of the gradient, respectively, where β1 and β2 are the decay rates; for m t and v t Perform deviation correction:
[0045]
[0046] Last updated parameters:
[0047]
[0048] Where, θ t The parameter is at time step t, α is the learning rate, and ∈ is a decimal added to maintain numerical stability;
[0049] S44: Model Evaluation and Validation: Evaluate the model using the validation sample dataset, and use accuracy, recall, and F1 score metrics to assess the model's performance; analyze the model's ability to diagnose various types of faults to ensure that the model has good generalization ability and robustness;
[0050] S45: Iterative optimization and termination judgment: Based on the model evaluation results, iteratively optimize the model architecture, parameters and hyperparameters; set termination conditions, and if the termination conditions are met, the model training is considered to have converged, the optimization process is terminated and the optimal model is selected.
[0051] The beneficial effects of this invention are as follows: By combining time-frequency enhancement technology, generative zero-shot learning, and lightweight deep learning networks, this invention effectively improves the accuracy and practicality of rolling bearing fault diagnosis under small sample conditions. In particular, the method of this invention can significantly reduce the computational resource requirements of the model while maintaining high diagnostic accuracy, making it suitable for real-time monitoring and fault early warning systems and possessing broad application prospects.
[0052] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:
[0054] Figure 1 This is a structural diagram of the small-sample rolling bearing fault diagnosis method based on time-frequency enhancement and generative small-sample learning described in this invention.
[0055] Figure 2 This is a structural diagram of the CVAE data augmentation algorithm described in this invention;
[0056] Figure 3 This is a schematic diagram of the MobileNet-V3 block structure described in this invention;
[0057] Figure 4 This is a flowchart illustrating the training process of the few-sample rolling bearing fault diagnosis algorithm based on time-frequency enhancement and generative few-sample learning described in this invention. Detailed Implementation
[0058] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0059] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0060] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0061] like Figure 1-4 As shown, this invention provides a few-sample rolling bearing fault diagnosis method based on time-frequency enhancement and generative few-sample learning, applicable to few-sample fault diagnosis of rolling bearings. The specific steps are as follows:
[0062] 101: Data Preprocessing: Collect and preprocess rolling bearing fault data to establish a rolling bearing fault dataset. Specifically, perform noise filtering and normalization on the collected vibration signal data to improve data quality and ensure the accuracy of subsequent analysis. For the collected one-dimensional vibration data, let the time series be X = {x1, x2, ..., x...} n}, x i For the i-th (i = 1, 2, ..., n) sampled signal, the one-dimensional time-series signal is scaled to the range [-1, 1] using a normalization algorithm, following the formula:
[0063]
[0064] In the formula For x i Normalized values. The scaled time series is then converted to polar coordinates and represented as follows:
[0065]
[0066] In the formula The normalized values of one-dimensional time series data are encoded into angles in a polar coordinate image using a cosine function. t i is the timestamp, encoded as the radius r in the polar coordinate image. N is a constant factor for regularization and the coordinate system generation space.
[0067] 102: Data Time-Frequency Feature Enhancement: The one-dimensional vibration fault signal is converted into a GADF time-frequency feature map using the GAF algorithm, specifically including:
[0068] Time-frequency transformation: Using the Gramian Angular Field (GAF) algorithm, the preprocessed one-dimensional vibration signal is mapped to a two-dimensional time-frequency space, generating a GADF time-frequency feature map. This transformation not only preserves the time-frequency characteristics of the original signal but also enhances the visual representation of these characteristics. The process follows the formula:
[0069]
[0070] In the formula That is, the angle in equation (2) is encoded into the polar coordinate image by the cosine function;
[0071] Sample partitioning: The generated GADF time-frequency feature map is divided into training samples, validation samples and test samples according to experimental requirements.
[0072] 103: Sample Augmentation: Generative data augmentation was performed using the CVAE model to augment the small training data. The generated GADF time-frequency feature maps were divided into training, validation, and test samples according to experimental requirements, preparing the data foundation for model training, validation, and testing. (Reference) Figure 2 As shown, based on the small sample data obtained in step 102, the CVAE generation model is trained to obtain expanded samples, specifically including the following steps:
[0073] 1031: Encoder Operation: The encoder transforms input data into a representation in a latent space. This representation is a compressed representation of the input data, containing its key features. In CVAE, the encoder receives input data and condition variables as input and outputs the mean and variance of the latent space. Define a probability distribution representing the latent representation of the input data under given conditions.
[0074] 1032: Sampling: Sample from the latent space based on the parameters (mean and variance) output by the encoder. This step introduces randomness, which is key to generating diverse new data. Follow the formula below:
[0075] μ = f φ (x)
[0076] σ 2 =g φ (x)
[0077] Where f φ and g φ , representing the output layer of the encoder network, respectively; x is used to output the mean and variance of the latent space, and is the input data.
[0078] z=μ+σ·∈
[0079] In the reparameterization process, z represents the latent variable, and ∈ is random noise drawn from the standard normal distribution.
[0080] 1033: Decoder Operation: The decoder's role is to transform the representation of the latent space back into the original data space, generating new data. In CVAE, the decoder receives the sampled points of the latent space and condition variables as input, and outputs the generated data. Due to the randomness of the sampling process, even with the same input data and conditions, different new data can be generated under different sampling conditions.
[0081] 1034: Loss Function Evaluation: In the CVAE model, reconstruction loss (MSE loss) is used to measure the similarity between the input and generated data, and KL divergence loss is used to measure the difference between the latent spatial distribution and the prior distribution (normal distribution). The following formula applies:
[0082]
[0083] in, For the reconstruction loss, KL represents the Kullback-Leibler divergence. Combining these two loss functions forms the model's total loss function. The model is trained by minimizing this total loss, thereby generating new data that is as close as possible to the real data.
[0084] 1035: Data Augmentation: After training the CVAE model, a large amount of new data can be generated by changing the condition variables and the sampling of the latent space for data augmentation. This is especially important for improving the model's performance on finite or imbalanced datasets.
[0085] 104: Constructing a GAF-CVAE-MobileNet-v3 rolling bearing fault diagnosis model. The model was trained and optimized using training and validation set sample data; see references... Figure 3 , Figure 4 As shown, based on the expanded and partitioned training, validation, and test sets obtained in step 103, a lightweight diagnostic model is trained to enable it to diagnose rolling bearing faults. This specifically includes the following steps:
[0086] 1041: Model Design: A rolling bearing fault diagnosis model based on GAF-CVAE-MobileNet-v3 is constructed. This model combines the advantages of time-frequency feature extraction, generative learning, and lightweight deep learning networks, aiming to achieve efficient and accurate fault diagnosis.
[0087] 1042: Model Training and Optimization: Train the model using the training set and evaluate and optimize its performance using the validation set. Adjust the model parameters based on the evaluation results to achieve optimal fault diagnosis performance. Use the Adam optimizer to adjust the learning rate of each parameter, following the formula:
[0088] m t =β1·m t-1 +(1-β1)·g t
[0089]
[0090] Where g t The gradient at real time step t, m t and v t These are the first and second moment estimates of the gradient, respectively, where β1 and β2 are the decay rates, typically close to 1. Then, for m... t and v t Perform deviation correction:
[0091]
[0092]
[0093] The bias correction here helps adjust for biases introduced during the initial training phase due to initialization to 0. Last updated parameters:
[0094]
[0095] Where, θ t The parameter is at time step t, α is the learning rate, and ∈ is a small number added to maintain numerical stability, ranging from 10. -7 Up to 10 -8 Within the range.
[0096] Step 105: Validate the rolling bearing fault diagnosis model using test set sample data. If the requirements are not met, return to step 104 until the results meet the requirements. In step 105, the rolling bearing fault diagnosis model is tested to determine whether it meets the highest fault diagnosis accuracy or whether the predetermined number of training iterations has been completed. If the model test results reach the set optimal performance standard, i.e., the model exhibits excellent accuracy, recall, and F1 score on the independent test set, then the model will be confirmed as the final fault diagnosis model. At this time, the key parameter configurations of the model will be recorded and saved, including the optimal layer settings and hyperparameters in the GAF-CVAE-MobileNet-v3 structure, to ensure that the optimal performance of the model can be reproduced and applied. Subsequently, the optimal model is used to process the independent test sample dataset to output the fault diagnosis results of the rolling bearing, providing accurate data support for maintenance decisions. This process not only demonstrates the model's application capability in a real industrial environment but also ensures the efficiency and reliability of the fault diagnosis method.
[0097] This invention is applicable to rolling bearing fault diagnosis. It utilizes a small-sample rolling bearing fault diagnosis method based on time-frequency enhancement and generative few-sample learning, as disclosed in this invention. By combining efficient time-frequency feature extraction techniques and advanced generative learning enhancement methods, it effectively improves fault diagnosis accuracy in small-sample data environments. The GAF algorithm and CVAE model not only enhance the time-frequency features of the original signal but also significantly expand the number and diversity of training samples. Furthermore, the lightweight MobileNet-v3 deep learning framework enables efficient and accurate fault feature learning and classification. This integration of technologies not only improves the accuracy of fault diagnosis but also significantly enhances the model's generalization ability and real-time performance, making it suitable for real-time monitoring and early warning of rolling bearing faults in various industrial application scenarios.
[0098] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When the program is executed, it can implement the steps of the method. The storage medium may be, for example, ROM / RAM, magnetic disk, optical disk, etc.
[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A small sample rolling bearing fault diagnosis method based on time-frequency enhancement and generative learning, characterized by: Includes the following steps: S1: Collect rolling bearing fault data and preprocess it to establish a rolling bearing fault dataset; S2: The one-dimensional vibration fault signal is converted into a GADF time-frequency feature map using the GAF algorithm, and training sample data, validation sample data and test sample data are divided. S3: Use the CVAE model to perform generative data sample augmentation on small training data; S4: Construct the GAF-CVAE-MobileNet-v3 rolling bearing fault diagnosis model; train and optimize the model using training set sample data and validation set sample data; S5: Use the test set sample data to verify the rolling bearing fault diagnosis model. If it does not meet the requirements, return to step S4 until the result meets the requirements. Step S1 specifically includes the following steps: S11: Data Collection: Vibration signal data of the rolling bearing device in operation is captured in real time by sensors and stored in the data acquisition system; S12: Data Preprocessing: For the collected one-dimensional vibration data, let the time series be... X ={ x 1, x 2, ..., x n }, x i For the first i A sampled signal is used to scale a one-dimensional time-series signal to the range [-1, 1] using a normalization algorithm, following the formula: In the formula for Normalized value, i =1, 2, ..., n; The scaled time series was then converted to polar coordinates and represented as follows: In the formula The normalized values of one-dimensional time series data are encoded into angles in a polar coordinate image using a cosine function. , The timestamp is encoded as a radius in the polar coordinate image. r , N A constant factor for regularization and coordinate system generation space; Step S2 specifically includes the following steps: S21: Time-Frequency Conversion: Using the Gram angle field algorithm, the preprocessed one-dimensional vibration signal is mapped to a two-dimensional time-frequency space to generate a GADF time-frequency feature map, following the formula below: In the formula This is the angle encoded into a polar coordinate image using a cosine function; S22: Sample partitioning: Divide the generated GADF time-frequency feature map into training samples, validation samples and test samples according to experimental requirements; Step S3 specifically includes the following steps: S31: Encoder Operation: The encoder transforms the input data into a representation in a latent space, which contains the key features of the data; in CVAE, the encoder receives the input data and condition variables as inputs and outputs the mean and variance of the latent space; a probability distribution is defined to represent the latent representation of the input data under given conditions; S32: Sampling: Based on the parameters output by the encoder, samples are taken from the latent space, introducing randomness, and following the formula: in and These represent the output layers of the encoder network, respectively. x is used to output the mean and variance of the latent space, and is the input data. During reparameterization, z represents a latent variable. It is random noise drawn from a standard normal distribution; S33: Decoder Operation: The function of the decoder is to transform the representation of the latent space back into the original data space and generate new data; in CVAE, the decoder receives the sampling points and condition variables of the latent space as input and outputs the generated data. S34: Loss Function Evaluation: In the CVAE model, reconstruction loss is used to measure the similarity between the input and generated data, and KL divergence loss is used to measure the difference between the latent spatial distribution and the prior distribution, following the formula: in, To reconstruct the loss, The Kullback-Leibler divergence is used to combine these two loss functions to form the model's total loss function. The model is trained by minimizing the total loss. S35: Data Augmentation: After training the CVAE model, a large amount of new data is generated by changing the condition variables and sampling of the latent space for data augmentation.
2. The method for small-sample rolling bearing fault diagnosis based on time-frequency enhancement and generative learning according to claim 1, characterized in that: Step S4 specifically includes the following steps: S41: Model Architecture Design: Determine the architecture of the GAF-CVAE-MobileNet-v3 model, including GAF time-frequency feature extraction, CVAE data augmentation, and the integration of the MobileNet-v3 lightweight deep learning model; the GAF algorithm is used to convert the original one-dimensional vibration signal into a two-dimensional time-frequency feature map, CVAE is responsible for generative data augmentation to expand the small sample dataset, and MobileNet-v3 is used to realize fault feature learning and classification; S42: Feature Learning and Optimization: Using GAF time-frequency feature maps as input, the MobileNet-v3 model is trained with CVAE-enhanced data. During this process, the deep learning model automatically learns and extracts fault features, and optimization algorithms such as backpropagation and gradient descent are used to adjust the model parameters to minimize classification error. S43: Hyperparameter Tuning: Adjust the learning rate, batch size, number of convolutional layers, and configure hyperparameters. Use the Adam optimizer to adjust the learning rate of each parameter, following the formula below: in The gradient at real time step t, and These are the first and second moment estimates of the gradient, respectively. and It is the attenuation rate; for and Perform deviation correction: Last updated parameters: in, It is a parameter at time step t. It's the learning rate. The decimal part was added to maintain numerical stability; S44: Model Evaluation and Validation: Evaluate the model using the validation sample dataset, and use accuracy, recall, and F1 score metrics to assess the model's performance; analyze the model's ability to diagnose various types of faults to ensure that the model has good generalization ability and robustness; S45: Iterative optimization and termination judgment: Based on the model evaluation results, iteratively optimize the model architecture, parameters and hyperparameters; set termination conditions, and if the termination conditions are met, the model training is considered to have converged, the optimization process is terminated and the optimal model is selected.
Citation Information
Patent Citations
Rolling bearing fault diagnosis method based on GAF-DRSN
CN114595730A
Rolling bearing fault diagnosis method under small sample data
CN114707534A
Fault diagnosis method of three-phase inverter based on improved deep residual network
CN117454156A
