Rolling bearing fault monitoring method and system based on vibration acoustic method

Abstract

The invention provides a rolling bearing fault monitoring method and system based on a vibration acoustic method, and relates to the technical field of fault diagnosis, and the method comprises the steps of data collection, model establishment, model training, fault recognition and the like. The system comprises a data acquisition module, a model building module, a model training module and a fault identification module, the fault modes of the rolling bearing can be distinguished in an environment where multiple fault modes coexist through building a fault monitoring model based on environmental condition changes, and the accuracy of fault monitoring and diagnosis of the rolling bearing is improved.

Description

Technical Field

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a rolling bearing fault monitoring method and system based on a vibroacoustic method. Background Art

[0002] With the advancement of industrialization, mechanical equipment is increasingly used in production. Rolling bearings are a vital component of mechanical transmission systems and are widely used in various types of industrial equipment. However, rolling bearings are prone to failure during long-term operation due to factors such as excessive load, insufficient lubrication, and improper installation. Rolling bearing failure not only affects the performance and operating efficiency of the equipment, but may also cause equipment downtime, production losses, and safety hazards. During use, the load, speed, and environmental conditions (such as temperature and humidity) that rolling bearings bear have an important impact on their working conditions. When rolling bearings are overloaded, poorly lubricated, or work at high speeds for a long time, their wear and fatigue will be significantly aggravated. In addition, corrosion, vibration, and excessive temperature will further increase the risk of rolling bearing failure, resulting in reduced equipment reliability. Therefore, effective monitoring and diagnosis of rolling bearing failures can timely detect potential problems, avoid greater losses caused by equipment failures, and thus improve equipment maintenance efficiency and production safety.

[0003] The Chinese invention patent with publication number CN117828350A provides a bearing fault diagnosis method and system for unbalanced data across working conditions. The patent is based on transfer learning and improved deep convolutional generative adversarial networks. By improving the deep convolutional generative adversarial network to pseudo-expand the small sample parameters of the target domain, the generated simulated signal retains the complete high-frequency and low-frequency characteristics of the real signal. At the same time, the improved SwinTransformer network is used to complete the identification of the health status of rolling bearings with unknown labels across working conditions.

[0004] With regard to the above technical solution, vibration signals are only collected when the environmental conditions are constant to analyze the failure of rolling bearings under different working conditions. The impact of changes in environmental conditions on rolling bearings is not considered. Therefore, when faced with an environment where multiple fault modes coexist, if only relying on vibration signal feature extraction for judgment, the failure mode of the rolling bearing cannot be effectively distinguished, resulting in reduced accuracy in rolling bearing fault monitoring and diagnosis. Summary of the invention

[0005] In order to improve the accuracy of rolling bearing fault monitoring, the present application provides a rolling bearing fault monitoring method and system based on a vibroacoustic method.

[0006] In a first aspect, the present application provides a rolling bearing fault monitoring method based on a vibroacoustic method, which adopts the following technical solution: A rolling bearing fault monitoring method based on a vibroacoustic method comprises the following steps: Data collection: including real-time data collection and historical data collection; Collect real-time data: collect real-time data of rolling bearings; Collect historical data: collect historical data of rolling bearings and the fault type labels corresponding to the historical data; Building a model: Building a fault monitoring model based on a generative adversarial network model and a deep learning algorithm. The fault monitoring model includes a VAE-GAN sub-model and a fault classification sub-model. The VAE-GAN sub-model includes an encoder, a generator, and a discriminator. Model training: including latent space modeling, pseudo data generation, pseudo data discrimination, and fault classification; Latent space modeling: historical data and the fault type labels corresponding to the historical data are input into the encoder to obtain latent variables; Generate pseudo data: Take the latent variables as the input of the generator to generate pseudo data; Discriminating pseudo data: using pseudo data as the input of the discriminator to obtain the pseudo data discrimination result; If the judgment result is not historical data, the parameters of the generator are optimized, and the steps of generating pseudo data to judging pseudo data are executed; If the judgment result is historical data, the fault classification step is executed; Fault classification: The pseudo data, historical data, and the fault type labels corresponding to the historical data are used as inputs to the fault classification sub-model to obtain the fault type labels of the pseudo data; Fault identification: Input real-time data into the trained fault monitoring model and output the fault type label of the real-time data.

[0007] By adopting the above technical solution, the fault monitoring model constructed by training the equipment information and environmental information of the rolling bearing is enabled to learn the potential representations of different fault modes, which helps the model to extract effective features to distinguish different fault modes from the latent space when facing the coexistence of multiple faults, instead of relying on a single feature for discrimination. Moreover, the model can simulate sample data under different fault modes for model training when data is scarce or lacking a specific fault mode, which enhances the diversity and integrity of the training sample data set, especially when the actual collected data is difficult to cover all fault modes, which helps to avoid the problems of insufficient samples and overfitting during model training, and enhances the generalization ability of the model. In summary, the fault monitoring model constructed by this technical solution can better adapt to the rolling bearing fault monitoring in an environment where multiple fault modes coexist, and accurately distinguish the rolling bearing status under different fault modes.

[0008] Optionally, the VAE-GAN sub-model further includes a decoder, and after performing the step of latent space modeling and before performing the step of generating pseudo data, further includes: Generate reconstructed data: Use the latent variables as input to the decoder to obtain reconstructed data; Calculate reconstruction loss: Calculate the error between the reconstructed data and the historical data to get the reconstruction loss; Calculate KL divergence: Calculate the difference between the potential distribution and the standard normal distribution, and use the difference as the KL divergence; Optimize the encoder: define the loss function of the encoder, where the loss function includes reconstruction loss and KL divergence, and iteratively optimize the encoder by minimizing the loss function of the encoder to obtain an optimized encoder.

[0009] By adopting the above technical solution, the distribution of the latent space is optimized using KL divergence, so that the distribution of the latent variables approaches the standard normal distribution, reducing the risk of model overfitting. In addition, the encoder and decoder in the VAE-GAN sub-model use historical data and the fault type labels corresponding to the historical data to train the model, so that the fault monitoring model can extract effective features to distinguish different fault modes from the latent space, overcoming the difficulties faced by single data extraction and improving the distinction of fault modes and monitoring accuracy.

[0010] Optionally, after performing the step of latent space modeling and before performing the step of generating pseudo data, the method further includes: Historical data splicing: splice the historical data with the fault type labels corresponding to the historical data to obtain condition information; Latent variable concatenation: concatenating conditional information with latent variables to obtain concatenated data, and using the concatenated data as new latent variables.

[0011] By adopting the above technical solution, historical data and its fault type labels are spliced ​​with latent variables as conditional information, and the spliced ​​data is used as the input of the generator. This can provide the generator with more contextual information, so that the generator can generate data of specific fault types in a targeted manner when generating pseudo data, thereby enhancing the control ability of the generator and improving the quality of pseudo data. The generated pseudo data not only conforms to the original data distribution, but also carries clear classification information, which is helpful for the optimization of downstream fault classification tasks.

[0012] Optionally, after executing the step of generating pseudo data and before executing the step of identifying pseudo data, the method further includes: Extract historical data features: Extract the key features of historical data and record them as the first features; Extract pseudo data features: Extract the key features of pseudo data and record them as the second features; Calculate feature differences: calculate the feature differences between the first feature and the second feature respectively; Evaluate pseudo data: Evaluate the calculated feature differences to determine whether the pseudo data is qualified.

[0013] By adopting the above technical solution, before the discriminator discriminates the pseudo data, the generated pseudo data is screened according to the key features of the historical data to ensure that the data features of the pseudo data are consistent with the data features of the historical data, achieving the technical effect of constraining the quality of the data generated by the generator, which helps to improve the fault classification sub-model's ability to identify fault modes. Moreover, the feature difference can be used as an additional loss term to guide the generator to generate higher quality pseudo data, accelerate the generator optimization, and at the same time, by filtering low-quality pseudo data, enhance the robustness of the fault monitoring model.

[0014] Optionally, after executing the step of evaluating the pseudo data and before executing the step of distinguishing the pseudo data, the method further includes: Evaluation result feedback: When the pseudo data is judged to be qualified, the steps of distinguishing the pseudo data are executed; when the pseudo data is judged to be unqualified, the feature difference is input into the generator; Error back propagation: The generator uses the feature differences of the input to optimize the weights of the generator; Optimize the generator: define the generator loss function, iteratively optimize the generator by minimizing the generator loss function, and obtain the optimized generator.

[0015] By adopting the above technical solution, dynamic evaluation and feedback are used to ensure that the pseudo data entering the discriminator has high-quality characteristics and is consistent with historical data. And according to the quality of the generated pseudo data, dynamic feedback is used to optimize the generator, that is, when the pseudo data is unqualified, the feature difference is fed back to the generator, and the weight of the generator is optimized through back propagation, so that the generator gradually learns the potential distribution of data characteristics that is closer to historical data, reduces the number of generator training iterations, and accelerates the training of the generator, so that the generator can generate pseudo data that is more in line with the requirements of fault classification tasks according to specific data distribution and task requirements.

[0016] Optionally, after executing the step of identifying the pseudo data and before executing the step of fault classification, the method further includes: Fault classification sub-model training: including data input, shared feature extraction, task-specific branching, task processing, computing multi-task losses, and optimizing fault classification sub-models; Data input: input historical data and the fault type labels corresponding to the historical data into the fault classification sub-model; Shared feature extraction: The shared layer of the fault classification sub-model is used to extract the features of the historical data and the fault type labels corresponding to the historical data, which are recorded as shared features. Task-specific branches: Create task-specific branches based on the fault type labels corresponding to historical data. Task-specific branches include fault type classification tasks and fault severity classification tasks. Task processing: input shared features into task-specific branches to process tasks; Calculate multi-task loss: define the multi-task loss function and calculate the multi-task loss value in the task processing process; Optimize the fault classification sub-model: Iteratively optimize the fault classification sub-model by minimizing the task loss function.

[0017] By adopting the above technical solution and using a multi-task learning framework, the tasks of fault type classification and fault severity classification are completed at the same time, which improves the efficiency of the fault monitoring model and realizes efficient fault diagnosis and evaluation. Moreover, the shared layer can utilize the synergy between multiple tasks to make the extracted features more universal, while reducing the risk of overfitting of a single task and improving the generalization ability of the model.

[0018] Optionally, after the step of performing task processing and before the step of calculating the multi-task loss, the method further includes: Measuring task relevance: Calculate the mutual information between the fault type classification task and the fault severity classification task as task relevance; Adjust the number of shared layers: Determine whether the task correlation is less than the preset correlation threshold. If so, reduce the number of shared layers of the fault type classification task and the fault severity classification task; if not, increase the number of shared layers of the fault type classification task and the fault severity classification task.

[0019] By adopting the above technical solution, mutual information is used to calculate task relevance, the degree of dependence between the fault type classification task and the fault severity classification task is quantified, the number of shared layers is dynamically adjusted according to the calculation results, and the network structure is optimized to make it more in line with actual task requirements. This ensures that each task can be processed in the optimal feature representation method, improves the overall accuracy, stability and robustness of the fault classification sub-model, and enables it to adapt to diverse task requirements under different task relevance.

[0020] Optionally, after executing the step of adjusting the number of shared layers and before executing the step of calculating the multi-task loss, the method further includes: Define the shared layer structure: Define the search space for shared layers, which includes the number of shared layers, the type of shared layers, and the type of task-specific branches; Shared layer architecture optimization: Use evolutionary algorithms to search for the shared layer network architecture and select the optimal shared layer network architecture.

[0021] By adopting the above technical solutions, using evolutionary algorithms for automated search, exploring the optimal network architecture of the shared layer, comprehensively considering multi-task performance and computing efficiency, selecting the shared layer architecture with the best performance, taking into account task-specific requirements and overall architecture complexity, making the shared layer more efficient and stable, avoiding performance bottlenecks caused by fixed architectures, and reducing the number of repeated experiments and debugging. At the same time, the optimized shared layer architecture can better handle conflicts and collaborative relationships between tasks, improve the classification accuracy of the two tasks, and improve the accuracy of the fault classification sub-model.

[0022] Optionally, the generator loss function is: ;

[0023] in, represents the adversarial loss function, represents the feature alignment loss function, Represents a hyperparameter.

[0024] By adopting the above technical solution, introducing adversarial loss function and feature alignment loss function, and avoiding model overfitting or divergence problems caused by a single loss function, it is helpful to generate high-quality data that meets task requirements, so that the pseudo data generated has both authenticity and feature consistency, which improves the training stability of the generator, and at the same time provides higher quality data support for downstream tasks, thereby improving the performance and practicality of the fault monitoring model.

[0025] In a second aspect, the present invention provides a rolling bearing fault monitoring system based on a vibroacoustic method, which adopts the following technical solution: A rolling bearing fault monitoring system based on a vibroacoustic method, comprising: Data collection module, including real-time data collection module and historical data collection module; Real-time data collection module, used to collect real-time data of rolling bearings; A historical data collection module is used to collect historical data of rolling bearings and fault type labels corresponding to the historical data; Establish a model module for constructing a fault monitoring model based on a generative adversarial network model and a deep learning algorithm, wherein the fault monitoring model includes a VAE-GAN sub-model and a fault classification sub-model; the VAE-GAN sub-model includes an encoder, a generator, and a discriminator; The model training module is connected to the data acquisition module and the model building module, and includes a latent space modeling unit, a pseudo data generation unit, a pseudo data discrimination unit, and a fault classification unit; A latent space modeling unit, used for inputting historical data and fault type labels corresponding to the historical data into an encoder to obtain latent variables; A pseudo data generation unit is used to generate pseudo data by taking the latent variables as input of the generator; A pseudo data discriminating unit, used for taking the pseudo data as an input of the discriminator to obtain a pseudo data discriminating result; If the judgment result is not historical data, the parameters of the generator are optimized, and the generation of pseudo data units and the judgment of pseudo data units are triggered; If the judgment result is historical data, the fault classification unit is triggered; A fault classification unit, used to use the pseudo data, the historical data, and the fault type labels corresponding to the historical data as inputs of the fault classification sub-model to obtain the fault type labels of the pseudo data; The fault identification module is connected to the data acquisition module and the model training module for inputting real-time data into the trained fault monitoring model and outputting the fault type label of the real-time data.

[0026] By adopting the above technical solution, the fault monitoring model constructed by the model building module is trained using the historical data collected by the data acquisition module, so that the model can simulate sample data under different fault modes when data is scarce or lacking specific fault mode data, thereby conducting more comprehensive model training. In addition, the trained fault monitoring model can accurately distinguish the state of the rolling bearing under different fault modes based on the input data, and better adapt to the environment where multiple fault modes coexist.

[0027] In summary, the present application includes at least one of the following beneficial technical effects: 1. The fault monitoring model constructed by training using the equipment information and environmental information of rolling bearings enables the model to learn the potential representations of different fault modes, which helps the model to extract effective features to distinguish different fault modes from the latent space when facing the coexistence of multiple faults, instead of relying on a single feature for discrimination. Moreover, the model can simulate sample data under different fault modes for model training when data is scarce or lacks specific fault modes, which enhances the diversity and completeness of the training sample data set, especially when the actual collected data is difficult to cover all fault modes, which helps to avoid the problems of insufficient samples and overfitting during model training and enhances the generalization ability of the model.

[0028] 2. Historical data and its fault type labels are concatenated with latent variables as conditional information, and the concatenated data is used as the input of the generator. This can provide the generator with more contextual information, so that the generator can generate data of specific fault types in a targeted manner when generating pseudo data, thereby enhancing the control ability of the generator and improving the quality of pseudo data. The generated pseudo data not only conforms to the original data distribution, but also carries clear classification information, which is helpful for optimizing downstream fault classification tasks.

[0029] 3. Before the discriminator discriminates the pseudo data, the generated pseudo data is screened according to the key features of the historical data to ensure that the data features of the pseudo data are consistent with the data features of the historical data, achieving the technical effect of constraining the quality of the data generated by the generator, which helps to improve the fault classification sub-model's ability to identify fault modes. Moreover, the feature difference can be used as an additional loss term to guide the generator to generate higher-quality pseudo data, accelerate the generator optimization, and at the same time, by filtering low-quality pseudo data, enhance the robustness of the fault monitoring model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of Example 1 of the present application; Figure 2 This is a flowchart of the S3 model training of Example 1 of the present application; Figure 3 is a flowchart of optimization of the S32 latent space of Example 1 of the present application; Figure 4 is a flow chart of preliminary screening of S35 pseudo data of Example 1 of the present application; Figure 5 This is a flowchart of the S37 fault classification sub-model training of Example 1 of the present application. DETAILED DESCRIPTION

[0031] The following combination Figures 1 to 5 This application is described in further detail.

[0032] Embodiment 1: This embodiment discloses a rolling bearing fault monitoring method based on a vibroacoustic method, such as Figure 1 As shown, the monitoring method includes: collecting real-time data of the rolling bearing and historical data of the rolling bearing and fault type labels corresponding to the historical data, and then constructing a fault monitoring model, wherein the fault monitoring model includes a VAE-GAN sub-model and a fault classification sub-model; the VAE-GAN sub-model includes an encoder, a generator and a discriminator, and the fault monitoring model is trained using the historical data of the rolling bearing and the fault type labels corresponding to the historical data to obtain a trained fault monitoring model, and finally the collected real-time data is input into the trained fault monitoring model, and the fault type labels of the real-time data are output. This embodiment includes the following steps: S1 data collection: includes S11 real-time data collection and S12 historical data collection.

[0033] S11 collects real-time data: collects real-time data of rolling bearings, which includes environmental data and equipment data. Environmental data includes temperature and humidity; equipment data includes vibration signals of rolling bearings, speed of rolling bearings, loads of rolling bearings, and lubricating oil metal powder of rolling bearings.

[0034] In this embodiment, the vibration signal of the rolling bearing can be collected by an accelerometer or vibration sensor installed outside the rolling bearing and close to the support position of the rolling bearing. The rotation speed of the rolling bearing is collected by a magnetic encoder or a photoelectric sensor installed on the shaft of the rolling bearing or on the bearing assembly. The load of the rolling bearing is collected by a force sensor, a strain gauge or a piezoelectric sensor installed on the bearing seat of the rolling bearing or the bearing support structure. The metal powder of the lubricating oil is collected by an oil filter installed in the lubricating oil of the rolling bearing, or the metal powder of the lubricating oil is collected by a magnetic particle sensor or an oil contamination sensor installed in the oil flow path. The temperature of the rolling bearing is collected by a temperature sensor, a thermocouple or a platinum resistance temperature detector installed outside the bearing or in the oil circuit system of the rolling bearing. The humidity of the rolling bearing is collected by a humidity sensor installed in the working environment of the rolling bearing.

[0035] In this embodiment, when collecting real-time data of the rolling bearing, the temperature, humidity, vibration signal, rotation speed, load and lubricating oil metal powder of the rolling bearing collected at the same time point are used as a set of real-time data of the rolling bearing.

[0036] S12 collects historical data: collects historical data of rolling bearings and fault type labels corresponding to the historical data. The historical data of rolling bearings include historical vibration signals, historical speeds, historical temperatures, historical loads, historical humidity, and historical lubricating oil metal powder. Among them, the fault type labels include normal state, wear fault, crack fault, peeling fault, overload fault, insufficient lubrication fault, corrosion fault, imbalance fault, overheating fault, and low temperature fault. Wear faults include mild wear fault, moderate wear fault, and severe wear fault; crack faults include mild crack fault, moderate crack fault, and severe crack fault; peeling faults include mild peeling fault, moderate peeling fault, and severe peeling fault; overload faults include mild overload fault, moderate overload fault, and severe overload fault; insufficient lubrication faults include mild insufficient lubrication fault, moderate insufficient lubrication fault, and severe insufficient lubrication fault; corrosion faults include mild corrosion fault, moderate corrosion fault, and severe corrosion fault; imbalance faults include mild imbalance fault, moderate imbalance fault, and severe imbalance fault.

[0037] In this embodiment, when collecting historical data of the rolling bearing, the historical temperature, historical humidity, historical vibration signal, historical speed, historical load and historical lubricating oil metal powder of the rolling bearing collected at the same historical time point are taken as a set of historical data of the rolling bearing, and a set of historical data corresponds to at least one fault type label.

[0038] In this embodiment, after collecting the real-time data and historical data of the rolling bearing, the data needs to be preprocessed. The steps of data preprocessing include data cleaning, data normalization and feature extraction.

[0039] Data cleaning: De-noising, processing of missing values, and removal of outliers are performed on the collected real-time data and historical data.

[0040] Denoising includes using smoothing filters (such as sliding average) to eliminate small fluctuations in temperature, humidity, speed, and load data; using wavelet transform to decompose the vibration signal data, reconstructing after removing the noise component; and using smoothing filters to remove measurement fluctuations in lubricating oil metal powder data. Processing missing values ​​includes directly deleting data with missing values. Removing outliers includes detecting outliers in the data based on the 3-sigma rule or box plot, and deleting the detected outlier data.

[0041] Data normalization: Normalize the collected real-time data and historical data.

[0042] Extract features: Extract the time domain features (including mean, variance, skewness, kurtosis, peak-to-peak value, envelope features, etc.), frequency domain features (main frequency, spectrum energy, harmonic features) and time-frequency domain features of vibration signal data (extract time-frequency features through wavelet transform and short-time Fourier transform). Calculate the rate of change or trend features of the speed (such as speed fluctuation) and the rate of change or trend features of the load (such as load fluctuation). Extract the statistical features of temperature and humidity (such as average value, change amplitude) respectively. Extract the statistical features of the particle size distribution of lubricating oil metal powder.

[0043] S2 establishes a model: based on a generative adversarial network model and a deep learning algorithm, a fault monitoring model is constructed, wherein the fault monitoring model includes a VAE-GAN sub-model and a fault classification sub-model; the VAE-GAN sub-model includes an encoder, a decoder, a generator and a discriminator.

[0044] In this embodiment, the VAE-GAN sub-model represents a combination of a variational autoencoder (VAE) and a generative adversarial network (GAN). In this embodiment, the VAE part in the VAE-GAN sub-model selects a conditional variational autoencoder as the VAE part model in the VAE-GAN sub-model, and the conditional variational autoencoder includes an encoder and a decoder. The GAN part in the VAE-GAN sub-model is a generative adversarial network model, and the generative adversarial network model includes a generator and a discriminator. The fault classification sub-model can adopt a convolutional neural network, a recurrent neural network or a Transformer model, etc. In this embodiment, the Transformer model is selected as the basic model of the fault classification sub-model.

[0045] S3 model training: including S31 latent space modeling, S32 latent space optimization, S33 latent variable splicing, S34 pseudo data generation, S35 pseudo data preliminary screening, S36 pseudo data discrimination, S37 fault classification sub-model training and S38 fault classification.

[0046] In this embodiment, if Figure 2 As shown, S3 model training includes the following: S31 Latent space modeling: historical data and the fault type labels corresponding to the historical data are input into the conditional variational autoencoder. The encoder in the conditional variational autoencoder maps the high-dimensional historical data and the corresponding fault type labels into a low-dimensional latent space, and outputs the latent distribution of the low-dimensional latent space. The latent distribution includes the mean and variance, and uses the reparameterization technique to sample from the latent distribution to generate latent variables.

[0047] In this embodiment, the introduction of fault type labels makes the distribution of latent variables under different types of faults different, thereby helping the fault monitoring model learn to distinguish different fault types.

[0048] S32 Optimization of latent space: including S321 generating reconstruction data, S322 calculating reconstruction loss, S323 calculating KL divergence and S324 optimizing encoder, such as Figure 3 shown.

[0049] S321 generates reconstructed data: the latent variables are input into the decoder, and the decoder maps the low-dimensional latent variables back to the high-dimensional space of the historical data to obtain the reconstructed data, so as to restore the input historical data as much as possible to ensure that the latent variables can effectively describe the characteristics of the input historical data.

[0050] S322 calculates the reconstruction loss: calculates the error between the reconstructed data and the historical data through the cross entropy loss function to obtain the reconstruction loss.

[0051] S323 calculates KL divergence: calculates the difference between the potential distribution and the standard normal distribution, and uses the difference as the KL divergence to constrain the distribution of the potential space to be close to the standard normal distribution, so that the distribution of the potential variables has regularity, which is convenient for the subsequent generation of new sample data.

[0052] S324 Optimize encoder: Define the loss function of the encoder, which includes reconstruction loss and KL divergence, and use gradient descent method (such as Adam optimizer) to optimize the parameters of the encoder. By minimizing the loss function of the encoder and iteratively optimizing the encoder, a conditional variational autoencoder that can effectively generate latent space representation and can generate real data distribution is finally obtained.

[0053] S33 Latent variable concatenation: Perform one-hot encoding on the fault type labels corresponding to the historical data, concatenate the fault type labels after the one-hot encoding with the corresponding historical data, and record the concatenated data as condition information. Concatenate the condition information with the latent variables to obtain concatenated data, and use the concatenated data as a new latent variable.

[0054] In the above steps, the historical temperature, historical humidity, historical vibration signal, historical speed, historical load and historical lubricating oil metal powder in each set of historical data correspond to the same fault type label.

[0055] S34 generates pseudo data: the latent variables are input into the generator. After calculation of each convolution layer in the generator, the low-dimensional latent variables are gradually mapped to the high-dimensional space. The high-dimensional data output by the last convolution layer is the pseudo data generated by the generator.

[0056] S35 Preliminary screening of pseudo data: Preliminary screening of generated pseudo data based on key features of historical data, such as Figure 4 As shown, S35 preliminary screening of pseudo data includes S351 extracting historical data features, S352 extracting pseudo data features, S353 calculating feature differences, S354 evaluating pseudo data, S355 feedback of evaluation results, S356 error back propagation and S357 optimizing the generator.

[0057] S351 extracts historical data features: extracts key features of historical data, including the average temperature, temperature gradient, maximum temperature, minimum temperature and temperature fluctuation range of historical temperature; the average humidity, humidity fluctuation range and humidity deviation value of historical humidity; the root mean square value, peak value, spectrum characteristics, harmonic components, envelope characteristics, time domain characteristics and frequency domain characteristics of historical vibration signals; the average speed and speed frequency of historical speeds; the average load and load change rate of historical loads; the number of particles of historical lubricating oil metal powder, and records the key features of the extracted historical data as the first feature.

[0058] S352 extracts pseudo data features: extracts key features of pseudo data, including the average temperature, temperature gradient, maximum temperature, minimum temperature and temperature fluctuation range in the pseudo data; the average humidity, humidity fluctuation range and humidity deviation value in the pseudo data; the root mean square value, peak value, spectrum characteristics, harmonic components, envelope characteristics, time domain characteristics and frequency domain characteristics of the vibration signal in the pseudo data; the average speed and speed frequency in the pseudo data; the average load and load change rate in the pseudo data; the number of particles of lubricating oil metal powder in the pseudo data, and records the key features of the data in the extracted pseudo data as the second feature.

[0059] S353 Calculate feature differences: Calculate feature differences between each feature in the first feature and the corresponding feature in the second feature according to methods such as Euclidean distance or cosine similarity.

[0060] S354 evaluates pseudo data: compares each calculated feature difference with the corresponding pre-set feature difference threshold. If the feature differences are all less than or equal to the corresponding feature difference threshold, the pseudo data is judged to be qualified; if there is a feature difference greater than the corresponding feature difference threshold, the pseudo data is judged to be unqualified.

[0061] S355 Evaluation result feedback: When the pseudo data is judged to be qualified, S36 is executed to distinguish the pseudo data; when the pseudo data is judged to be unqualified, the feature difference is input into the generator and S356 error back propagation is executed.

[0062] S356 Error back propagation: The generator uses the feature differences of the input to optimize the weights of the generator.

[0063] S357 optimizes the generator: define the generator loss function, the generator loss function is , the generator is iteratively optimized by minimizing the generator loss function to obtain the optimized generator.

[0064] ; .

[0065] in, represents the generator loss function, represents the adversarial loss function, represents the feature alignment loss function, represents the hyperparameter, Express expectations, Represents a random noise distribution The random noise sampled from , Indicates that the generator is based on random noise Generated pseudo data, Represents the fake data generated by the generator to the discriminator The predicted probability, N represents the number of historical data, Indicates The feature vector of historical data, The generator generates The feature vector of the pseudo data.

[0066] In this embodiment, a grid search or a random search can be used to select a series of candidate values, and evaluate the generator in different The performance under the value, select the best performance value.

[0067] S36: Distinguishing pseudo data: taking the pseudo data as the input of the discriminator to obtain the pseudo data discrimination result.

[0068] If the discrimination result is not historical data, that is, the discriminator discriminates that the pseudo data is not historical data, the parameters of the generator are optimized, and S34 is executed to generate pseudo data to S36 to discriminate pseudo data.

[0069] If the discrimination result is historical data, that is, the discriminator discriminates the pseudo data as historical data, then S37 fault classification sub-model training is executed.

[0070] S37 Fault classification sub-model training: including S371 data input, S372 shared feature extraction, S373 task-specific branching, S374 task processing, S375 shared layer structure adjustment, S376 calculation of multi-task loss and S377 optimization of fault classification sub-model, such as Figure 5 shown.

[0071] In this embodiment, the Transformer model is selected as the base model of the fault classification sub-model, so a part of the Transformer encoder layers are used as shared layers of the fault classification sub-model, and the remaining Transformer encoder layers and the fully connected layers that perform specific tasks in the Transformer model are used as task-specific layers of the fault classification sub-model.

[0072] S371 Data input: input historical data and the fault type labels corresponding to the historical data into the fault classification sub-model.

[0073] S372 Shared feature extraction: Use the shared layer to extract the features of historical data, extract the feature matrix represented by low-dimensional features, and record it as shared features.

[0074] S373 Task-specific branch: Create task-specific branches based on the fault type labels corresponding to the historical data. The task-specific branches include fault type classification tasks and fault severity classification tasks. The fault type classification task is a task that outputs the probability distribution of fault categories through a set of task-specific layers. The fault severity classification task is a task that outputs the probability distribution of fault severity through another set of task-specific layers.

[0075] S374 Task processing: The shared features are input into the fault type classification task and the fault severity classification task respectively. The fault type classification task and the fault severity classification task respectively use the corresponding task specific layers to perform task processing on the input shared features.

[0076] S375 shared layer structure adjustment: including S3751 measuring task relevance, S3752 adjusting the number of shared layers, S3753 defining the shared layer structure and S3754 shared layer architecture optimization.

[0077] S3751 measures task correlation: the outputs of the fault type classification task and the fault severity classification task are taken as their random variables respectively, the mutual information between the two is calculated using the mutual information formula, and the calculated mutual information is taken as the task correlation between the fault type classification task and the fault severity classification task.

[0078] The mutual information formula is: .

[0079] Among them, T represents the output of the fault type classification task, S represents the output of the fault severity classification task, represents the joint distribution probability of the fault type classification task and the fault severity classification task, represents the marginal distribution probability of the output of the fault type classification task, Represents the marginal distribution probability of the output of the fault severity classification task.

[0080] S3752 adjusts the number of shared layers: determines whether the task relevance is less than the preset relevance threshold. If so, reduces the number of shared layers, for example, removes the last few layers in the shared layer from the shared layer and redefines them as task-specific layers, that is, transforms the last few layers in the Transformer encoder layer that belong to the shared layer into task-specific layers for fault type classification tasks and fault severity classification tasks respectively; if not, increases the number of shared layers, for example, adds a new Transformer encoder layer in front of the existing shared layer as a newly added shared layer.

[0081] S3753 defines the shared layer structure: defines the search space for shared layers, which includes the number of shared layers, the types of shared layers (such as convolutional layers, fully connected layers, attention mechanisms, etc.), and the types of task-specific branches (such as Softmax classification heads, ReLU activation heads, etc.).

[0082] In this embodiment, the number of shared layers in the search space of shared layers is defined according to the number of shared layers adjusted according to S3752.

[0083] S3754 Shared layer architecture optimization: Using evolutionary algorithms, a batch of shared layer architectures are randomly generated from the defined shared layer search space. The architecture performance is evaluated by the loss values ​​of the two tasks, the computational efficiency of the model, and the classification accuracy of the two tasks. The current shared layer architecture is used for crossover and mutation to generate a new shared layer architecture. According to the score of the fitness function, the optimal shared layer network architecture is selected according to the criteria of task relevance, minimum weighted sum of losses of the two tasks, and high classification accuracy.

[0084] S376 Calculate multi-task loss: Define the multi-task loss function, which is: . Calculate the task loss of each task according to the prediction result and label of each task, substitute the task loss of each task into the above formula to get the multi-task loss value.

[0085] in, represents the multi-task loss function, is the cross entropy loss for the fault type classification task, is the cross entropy loss for the fault severity classification task, and are weight coefficients, .

[0086] In this embodiment, cross validation or leave-out method can be used to evaluate the fault classification sub-model in and The performance of different weight settings and choose the best performing one and Weighted combination.

[0087] S377 optimizes the fault classification sub-model: Use a back-propagation algorithm (such as gradient descent or Adam optimizer) to update the weights of the shared layer and task branch, and iterate the optimization until the loss function converges or the training round is reached to obtain the optimized fault classification sub-model.

[0088] S38 Fault classification: The pseudo data is used as the input of the fault classification sub-model to obtain the fault type label of the pseudo data.

[0089] S4 fault identification: The collected real-time data of rolling bearings is input into the trained fault monitoring model, and the fault type label of the real-time data is output.

[0090] In this embodiment, the fault monitoring model constructed by training the equipment information and environmental information of the rolling bearing enables the model to learn the potential representations of different fault modes, which helps the model to extract effective features to distinguish different fault modes from the latent space when facing the situation of multiple faults coexisting, without relying on a single feature for discrimination. Moreover, the present application can simulate sample data under different fault modes for model training when data is scarce, especially when the actual collected data is difficult to cover all fault modes, thereby enhancing the diversity and integrity of the training sample data set. The fault monitoring model constructed by the present application can better adapt to the rolling bearing fault monitoring in an environment where multiple fault modes coexist, and accurately distinguish the rolling bearing status under different fault modes.

[0091] For example, the researcher collected data on rolling bearings at a certain moment, including temperature (100°C), humidity (55%), vibration signal (6.0m / s²), speed (2650rpm), load (1400N) and lubricating oil metal powder (32ppm), and performed fault type monitoring and diagnosis. The specific steps are as follows: 1. Collect historical data of rolling bearings and their fault type labels, see Table 1: Table 1 Rolling bearing historical data and fault type label table Serial number Vibration signal RMS (m / s²) Speed ​​(rpm) Temperature(℃) Load(N) humidity(%) Lubricating oil metal dust (ppm) Fault Type Label 1 1.2 3000 75 1000 35 5 Normal state 2 2.5 2900 80 1100 38 10 Minor wear and tear 3 5.0 2750 90 1250 42 20 Severe wear and tear 4 6.0 2650 95 1350 50 28 Moderate crack failure 5 7.2 2600 100 1400 55 35 Severe crack failure 6 3.0 2900 78 1150 38 12 Mild peeling failure 7 4.5 2800 82 1200 42 18 Moderate peeling 8 3.6 2800 85 1150 40 14 Moderate overload fault 9 4.8 2700 90 1250 45 20 Severe overload fault 10 4.0 2800 85 1200 42 20 Moderate corrosion failure 11 5.0 2750 90 1300 45 25 Severe corrosion failure 12 3.0 2900 78 1150 48 12 Minor lubrication deficiency 13 4.2 2850 85 1200 50 16 Moderate lubrication deficiency fault 14 4.5 2700 95 1350 50 28 Minor imbalance fault 15 7.5 2600 105 1450 60 35 Severe imbalance fault 16 5.8 2700 88 1300 45 22 Mild peeling failure, severe wear failure 17 4.8 2700 90 1250 45 20 Mild overload fault, moderate corrosion fault 18 4.8 2700 92 1300 45 25 Slight crack failure

[0092] 2. Data processing: The collected historical data of the rolling bearings are cleaned, normalized and features are extracted to obtain processed historical data.

[0093] 3. Model construction: Based on the generative adversarial network model and deep learning algorithm, a fault monitoring model is constructed. The fault monitoring model includes a VAE-GAN sub-model and a fault classification sub-model; the VAE-GAN sub-model includes an encoder, a decoder, a generator and a discriminator.

[0094] 4. Use the processed historical data and its fault type labels to train the model: The processed historical data and its fault type labels are input into the fault monitoring model. The encoder uses the input historical data and its fault type labels to generate latent variables. The historical data and its fault type labels are spliced ​​together, and the spliced ​​data is spliced ​​with the latent variables to obtain new latent variables. The new latent variables are input into the generator to generate pseudo data. After the discriminator determines that the pseudo data is historical data, the generated pseudo data is imported into the fault classification sub-model, and the fault classification model of the pseudo data is output to obtain a trained fault monitoring model.

[0095] 5. Input the collected data into the trained fault monitoring model to obtain the fault classification results: Temperature (100°C), humidity (55%), vibration signal (6.0m / s²), speed (2650rpm), load (1400N) and lubricating oil metal powder (32ppm) are used as a set of data input into the trained fault monitoring model, and the output fault type label is moderate imbalance fault.

[0096] Embodiment 2: This embodiment discloses a rolling bearing fault monitoring system based on a vibroacoustic method, the monitoring system comprising: The data acquisition module includes a real-time data acquisition module and a historical data acquisition module.

[0097] The real-time data collection module is used to collect real-time data of rolling bearings.

[0098] The historical data collection module is used to collect the historical data of the rolling bearing and the fault type labels corresponding to the historical data.

[0099] A model module is established to construct a fault monitoring model based on a generative adversarial network model and a deep learning algorithm. The fault monitoring model includes a VAE-GAN sub-model and a fault classification sub-model; the VAE-GAN sub-model includes an encoder, a decoder, a generator and a discriminator.

[0100] The model training module communicates with the data acquisition module and the model module, and includes a latent space modeling unit, a latent space optimization unit, a latent variable splicing unit, a pseudo data generation unit, a pseudo data preliminary screening unit, a pseudo data identification unit and a fault classification unit.

[0101] The latent space modeling unit is used to input the historical data and the fault type labels corresponding to the historical data into the encoder to obtain the latent variables.

[0102] Latent space optimization unit: communicates with the latent space modeling unit and the latent variable splicing unit, including a sub-unit for generating reconstructed data, a sub-unit for calculating reconstruction loss, a sub-unit for calculating KL divergence, and a sub-unit for optimizing the encoder.

[0103] Generate reconstructed data subunit: used to input latent variables into the decoder. The decoder maps the low-dimensional latent variables back to the high-dimensional space of historical data to obtain reconstructed data, so as to restore the input historical data as much as possible to ensure that the latent variables can effectively describe the characteristics of the input historical data.

[0104] Calculate reconstruction loss subunit: used to calculate the error between the reconstructed data and the historical data through the cross entropy loss to obtain the reconstruction loss.

[0105] KL divergence calculation subunit: used to calculate the difference between the potential distribution and the standard normal distribution, and use the difference as the KL divergence to constrain the distribution of the potential space to be close to the standard normal distribution, so that the distribution of the potential variables has regularity, which is convenient for the subsequent generation of new samples.

[0106] Optimize encoder subunit: used to define the loss function of the encoder, which includes reconstruction loss and KL divergence, and uses gradient descent method (such as Adam optimizer) to optimize the parameters of the encoder. By minimizing the loss function of the encoder and iteratively optimizing the encoder, a conditional variational autoencoder that can effectively generate latent space representation and can generate real data distribution is finally obtained.

[0107] Latent variable concatenation unit: used to perform one-hot encoding on the fault type labels corresponding to the historical data, concatenate the one-hot encoded fault type labels with the corresponding historical data, and record the concatenated data as conditional information. The conditional information is concatenated with the latent variables to obtain new latent variables.

[0108] The pseudo data generation unit is used to generate pseudo data by taking the latent variables as the input of the generator.

[0109] A preliminary screening unit for pseudo data: communicates with a pseudo data generating unit and a pseudo data judging unit, and includes a historical data feature extracting sub-unit, a pseudo data feature extracting sub-unit, a feature difference calculating sub-unit, a pseudo data evaluating sub-unit, an evaluation result feedback sub-unit, an error back propagation sub-unit and a generator optimization sub-unit.

[0110] Extract historical data feature subunit: used to extract key features of historical data, recorded as the first feature.

[0111] Extract pseudo data feature subunit: used to extract key features of pseudo data, recorded as the second feature.

[0112] The feature difference calculation subunit is used to calculate the feature difference between the first feature and the second feature respectively.

[0113] Evaluate pseudo data subunit: used to evaluate the calculated feature differences and determine whether the pseudo data is qualified.

[0114] Evaluation result feedback subunit: used to execute the steps of distinguishing the pseudo data unit when the pseudo data is judged to be qualified; when the pseudo data is judged to be unqualified, input the feature difference into the generator.

[0115] Error back propagation subunit: used by the generator to optimize the weights of the generator using the feature differences of the input.

[0116] Optimize generator subunit: used to define the generator loss function, iteratively optimize the generator by minimizing the generator loss function, and obtain the optimized generator.

[0117] The pseudo data discriminating unit is used to take the pseudo data as the input of the discriminator to obtain the pseudo data discriminating result.

[0118] If the judgment result is not historical data, the parameters of the generator are optimized, and the generation of pseudo data units and the judgment of pseudo data units are triggered.

[0119] If the judgment result is historical data, the fault classification unit is triggered.

[0120] The fault classification unit is used to take the pseudo data, the historical data and the fault type labels corresponding to the historical data as inputs of the fault classification sub-model to obtain the fault type labels of the pseudo data.

[0121] The fault identification module is connected to the data acquisition module and the model training module for inputting real-time data into the trained fault monitoring model and outputting the fault type label of the real-time data.

[0122] In this embodiment, the fault monitoring model constructed by the model construction module is trained using the historical data collected by the data acquisition module, so that the model can simulate sample data under different fault modes when data is scarce or lacking specific fault mode data, thereby performing more comprehensive model training. In addition, the trained fault monitoring model can accurately distinguish the state of the rolling bearing under different fault modes based on the input data, and better adapt to the environment where multiple fault modes coexist.

[0123] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A rolling bearing fault monitoring method based on a vibroacoustic method, characterized in that: include: Data collection: including real-time data collection and historical data collection; Collect real-time data: collect real-time data of rolling bearings; Collect historical data: collect historical data of rolling bearings and the fault type labels corresponding to the historical data; Building a model: Building a fault monitoring model based on a generative adversarial network model and a deep learning algorithm. The fault monitoring model includes a VAE-GAN sub-model and a fault classification sub-model. The VAE-GAN sub-model includes an encoder, a generator, and a discriminator. Model training: including latent space modeling, pseudo data generation, pseudo data discrimination, and fault classification; Latent space modeling: historical data and the fault type labels corresponding to the historical data are input into the encoder to obtain latent variables; Generate pseudo data: Take the latent variables as the input of the generator to generate pseudo data; Discriminating pseudo data: using pseudo data as the input of the discriminator to obtain the pseudo data discrimination result; If the judgment result is not historical data, the parameters of the generator are optimized, and the steps of generating pseudo data to judging pseudo data are executed; If the judgment result is historical data, the fault classification step is executed; Fault classification: The pseudo data, historical data, and the fault type labels corresponding to the historical data are used as inputs to the fault classification sub-model to obtain the fault type labels of the pseudo data; Fault identification: Input real-time data into the trained fault monitoring model and output the fault type label of the real-time data.

2. The rolling bearing fault monitoring method based on the vibroacoustic method according to claim 1 is characterized in that: The VAE-GAN sub-model also includes a decoder, and after performing the step of latent space modeling and before performing the step of generating pseudo data, it also includes: Generate reconstructed data: Use the latent variables as input to the decoder to obtain reconstructed data; Calculate reconstruction loss: Calculate the error between the reconstructed data and the historical data to get the reconstruction loss; Calculate KL divergence: Calculate the difference between the potential distribution and the standard normal distribution, and use the difference as the KL divergence; Optimize the encoder: define the loss function of the encoder, where the loss function includes reconstruction loss and KL divergence, and iteratively optimize the encoder by minimizing the loss function of the encoder to obtain an optimized encoder.

3. The rolling bearing fault monitoring method based on the vibroacoustic method according to claim 1, characterized in that: After performing the step of latent space modeling and before performing the step of generating pseudo data, it also includes: Historical data splicing: splice the historical data with the fault type labels corresponding to the historical data to obtain condition information; Latent variable concatenation: concatenating conditional information with latent variables to obtain concatenated data, and using the concatenated data as new latent variables.

4. The rolling bearing fault monitoring method based on the vibroacoustic method according to claim 1, characterized in that: After executing the step of generating pseudo data and before executing the step of identifying pseudo data, the method further includes: Extract historical data features: Extract the key features of historical data and record them as the first features; Extract pseudo data features: Extract the key features of pseudo data and record them as the second features; Calculate feature differences: calculate the feature differences between the first feature and the second feature respectively; Evaluate pseudo data: Evaluate the calculated feature differences to determine whether the pseudo data is qualified.

5. The rolling bearing fault monitoring method based on the vibroacoustic method according to claim 4 is characterized in that: After executing the step of evaluating the pseudo data and before executing the step of discriminating the pseudo data, the method further includes: Evaluation result feedback: When the pseudo data is judged to be qualified, the steps of distinguishing the pseudo data are executed; when the pseudo data is judged to be unqualified, the feature difference is input into the generator; Error back propagation: The generator uses the feature differences of the input to optimize the weights of the generator; Optimize the generator: define the generator loss function, iteratively optimize the generator by minimizing the generator loss function, and obtain the optimized generator.

6. The rolling bearing fault monitoring method based on the vibroacoustic method according to claim 1, characterized in that: After executing the step of identifying the pseudo data and before executing the step of fault classification, the method further includes: Fault classification sub-model training: including data input, shared feature extraction, task-specific branching, task processing, computing multi-task losses, and optimizing fault classification sub-models; Data input: input historical data and the fault type labels corresponding to the historical data into the fault classification sub-model; Shared feature extraction: The shared layer of the fault classification sub-model is used to extract the features of the historical data and the fault type labels corresponding to the historical data, which are recorded as shared features. Task-specific branches: Create task-specific branches based on the fault type labels corresponding to historical data. Task-specific branches include fault type classification tasks and fault severity classification tasks. Task processing: input shared features into task-specific branches to process tasks; Calculate multi-task loss: define the multi-task loss function and calculate the multi-task loss value in the task processing process; Optimize the fault classification sub-model: Iteratively optimize the fault classification sub-model by minimizing the task loss function.

7. The rolling bearing fault monitoring method based on the vibroacoustic method according to claim 6 is characterized in that: After the step of performing task processing and before the step of calculating the multi-task loss, the following steps are also included: Measuring task relevance: Calculate the mutual information between the fault type classification task and the fault severity classification task as task relevance; Adjust the number of shared layers: Determine whether the task correlation is less than the preset correlation threshold. If so, reduce the number of shared layers of the fault type classification task and the fault severity classification task; if not, increase the number of shared layers of the fault type classification task and the fault severity classification task.

8. The rolling bearing fault monitoring method based on the vibroacoustic method according to claim 7 is characterized in that: After the step of adjusting the number of shared layers and before the step of calculating the multi-task loss, the following steps are also included: Define the shared layer structure: Define the search space for shared layers, which includes the number of shared layers, the type of shared layers, and the type of task-specific branches; Shared layer architecture optimization: Use evolutionary algorithms to search for the shared layer network architecture and select the optimal shared layer network architecture.

9. The rolling bearing fault monitoring method based on the vibroacoustic method according to claim 5, characterized in that: The generator loss function is: ； in, represents the adversarial loss function, represents the feature alignment loss function, Represents a hyperparameter.

10. A rolling bearing fault monitoring system based on a vibroacoustic method, the system being applicable to the method according to any one of claims 1 to 9, characterized in that: include: Data collection module, including real-time data collection module and historical data collection module; Real-time data collection module, used to collect real-time data of rolling bearings; A historical data collection module is used to collect historical data of rolling bearings and fault type labels corresponding to the historical data; Establish a model module for constructing a fault monitoring model based on a generative adversarial network model and a deep learning algorithm, wherein the fault monitoring model includes a VAE-GAN sub-model and a fault classification sub-model; the VAE-GAN sub-model includes an encoder, a generator, and a discriminator; The model training module is connected to the data acquisition module and the model building module, and includes a latent space modeling unit, a pseudo data generation unit, a pseudo data discrimination unit, and a fault classification unit; A latent space modeling unit, used for inputting historical data and fault type labels corresponding to the historical data into an encoder to obtain latent variables; A pseudo data generation unit is used to generate pseudo data by taking the latent variables as input of the generator; A pseudo data discriminating unit, used for taking the pseudo data as an input of the discriminator to obtain a pseudo data discriminating result; If the judgment result is not historical data, the parameters of the generator are optimized, and the generation of pseudo data units and the judgment of pseudo data units are triggered; If the judgment result is historical data, the fault classification unit is triggered; A fault classification unit, used to use the pseudo data, the historical data, and the fault type labels corresponding to the historical data as inputs of the fault classification sub-model to obtain the fault type labels of the pseudo data; The fault identification module is connected to the data acquisition module and the model training module for inputting real-time data into the trained fault monitoring model and outputting the fault type label of the real-time data.

Images