A bearing health status assessment method and system based on deep latent variable clustering model

Abstract

The present invention discloses a bearing health status assessment method and system based on a deep latent variable clustering model, which automatically assesses the health status of bearings during the degradation process. The vibration signal of the bearing throughout its life cycle is collected, and the extracted time-frequency characteristic parameters of the vibration signal are screened based on monotonicity, robustness and correlation standards. The screened parameters are subjected to principal component analysis PCA dimensionality reduction, and the principal component with the highest explained variance ratio is selected; based on the GRU-VAE network, the network is optimized in combination with particle swarm optimization and multi-layer perceptron, and the bearing degradation is objectively divided into healthy, decaying and faulty stages by means of potential information mining and unsupervised clustering training of the deep latent variable clustering model, thereby ensuring the objectivity and accuracy of the degradation stage division. It can solve the defects of the current bearing health assessment method, such as high computational complexity and poor real-time performance, as well as the manual feature extraction and processing of large amounts of data still relying on professional knowledge, which affects the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present application relates to the technical field of bearing fault detection, and in particular, to a bearing health status assessment method and system based on a deep latent variable clustering model. Background Art

[0002] Bearings are one of the most critical components in mechanical equipment, and their health directly affects the reliability and service life of the equipment. However, due to long-term operation and harsh working environment, bearings are prone to wear, fatigue and other failures. Therefore, real-time monitoring and accurate assessment of bearing health are essential to ensure the normal operation of equipment and prevent unexpected downtime.

[0003] Traditional bearing health assessment methods mainly rely on vibration analysis, temperature monitoring, and lubricant analysis. Although these methods are effective to a certain extent, they usually require professional knowledge and experience, and it is difficult to achieve real-time monitoring and prediction. In addition, traditional methods have limited response capabilities to the complex operating conditions and changing working conditions of equipment, and cannot fully capture the degradation process of bearings. With the rapid development of sensor technology, data acquisition and processing technology, data-driven bearing health assessment methods have gradually become a research hotspot. By using modern artificial intelligence algorithms for health status identification and prediction, the limitations of traditional physical degradation models that require high professional knowledge of researchers can be overcome, thereby achieving more accurate and efficient health assessments.

[0004] Although some AI-based health assessment solutions are now available, there are still some significant shortcomings. These solutions usually use complex deep learning models, which leads to high computational complexity, poor real-time performance, and manual feature extraction and processing of large amounts of data still rely on professional knowledge. In addition, the internal mechanisms of many deep learning models, especially complex recurrent neural networks, are difficult to explain, resulting in poor model interpretability. This poses challenges to fault diagnosis and decision support, and affects users' trust and acceptance of model results. Summary of the invention

[0005] The purpose of this application is to provide a bearing health status assessment method and system based on a deep latent variable clustering model to solve the technical defects of the current intelligent bearing health assessment method, which has high computational complexity, poor real-time performance, manual feature extraction and processing of large amounts of data still rely on professional knowledge, poor model interpretability, and affect the accuracy of fault diagnosis. The health status of the bearing is automatically assessed during the degradation process.

[0006] This application adopts the following technical solutions to solve the above technical problems:

[0007] A bearing health status assessment method based on a deep latent variable clustering model is characterized by comprising the following steps:

[0008] (1) Collect vibration signals of the bearing throughout its life cycle;

[0009] (2) Extract and calculate the time domain and frequency domain features of the vibration signal at a certain time interval; each feature obtains a string of feature sequence features and time series;

[0010] (3) Based on the comprehensive score weight formula, calculate the comprehensive score of each feature and screen the features whose comprehensive scores meet the set standards; the comprehensive score weight formula is:

[0011] ;

[0012] In the formula, , , They are monotonicity indices Mon , robustness index Rob , correlation index Rec The weight index of Score is the comprehensive score;

[0013] (4) Perform PCA dimensionality reduction on the selected parameters. When the cumulative variance reaches a set percentage, stop increasing the dimension and select the principal component with the highest explained variance ratio;

[0014] (6) Use exponentially weighted moving average to smooth the reduced-dimensional curve;

[0015] (5), use In principle, the curves after dimension reduction and smoothing are divided into healthy stage, decay stage and failure stage, and then normalized as training sets;

[0016] (7) Build a network: Build a neural network including GRU and VAE to reconstruct the data, and build an MLP network to use MLP to divide the health status of the reconstructed signal;

[0017] (8) Use PSO to optimize the parameters of the hidden layer size of GRU and the potential feature size of VAE;

[0018] (9) Import the training set into the network after parameter optimization for training, and output the trained model;

[0019] (10) Monitor the bearing vibration signal in real time, calculate the time domain and frequency domain parameters after optimization at the current moment, and import them into the trained model after PCA dimension reduction to evaluate the current bearing health status.

[0020] In the above technical solution, the characteristics described in step (2) include at least one of the following: effective value of the through-frequency speed, effective value of the low-frequency speed, effective value of the acceleration, acceleration peak value, vibration impact value, acceleration kurtosis, acceleration skewness, velocity peak value, peak factor, and acceleration signal power spectrum.

[0021] In the above technical solution, the monotonicity index of step (3) is:

[0022] ;

[0023] In the formula, is the difference between the rankings of the two variables; n is the number of data points; ,The larger the monotonicity index is, the stronger the characteristic monotonicity trend is as the bearing performance degradation increases.

[0024] In the above technical solution, the correlation index is:

[0025] ;

[0026] In the formula, It is the characteristic i elements; is the time series i elements; yes x The mean of . yes t The mean of .

[0027] Among the above technical solutions, the comprehensive scores are:

[0028]

[0029] In the formula, , , They are the weight indicators of monotonicity, robustness and correlation respectively.

[0030] In the above technical solution, , , They are weight indicators of monotonicity, robustness, and correlation, with weights of 0.4, 0.3, and 0.3 respectively.

[0031] In the above technical solution, the comprehensive score of each feature in step (3) is greater than or equal to 70 points as the screening target.

[0032] In the above technical solution, in step (4), when the cumulative variance reaches more than 95%, the dimension is stopped from being increased.

[0033] In the above technical solution, the exponentially weighted moving average in step (5) is defined as:

[0034] ;

[0035] In the formula, It's at the time t The exponentially weighted moving average of ; It's at the time t The original data value of is the smoothing coefficient, usually in the range of , which represents the weight of the current data.

[0036] In the above technical solution, step (6) intercepts several health data points before the full life data to calculate the variance and mean According to the bearing experiments of multiple full life curves, the data point value is less than Marked as the health stage, and Between is the recession stage, greater than as the fault stage, and then normalized as the training set.

[0037] In the above technical solution, step (6) intercepts the first 30 health data points of the whole life data to calculate the variance and mean .

[0038] In the above technical solution, step (7) constructs a dual-gate GRU network and connects it to a VAE network to form a GRU-VAE network to reconstruct the data. In addition, a three-layer MLP network is built to use MLP to divide the health status of the reconstructed signal.

[0039] In the above technical solution, step (8) takes the hidden layer dimension hidden_dim of the GRU and the potential feature dimension latent_dim of the VAE network as the objects of PSO optimization, and the fitness function is selected to minimize the network loss function.

[0040] In the above technical solution, in step (9), the training set is imported into the GRU-VAE network with optimized parameters to reconstruct the signal and mine the potential features. Then, the reconstructed high-dimensional signal is imported into the MLP network for supervised classification. Finally, the trained model is saved.

[0041] In the above technical solution, step (10) collects the bearing vibration signal in real time, calculates the time domain and frequency domain parameters at the current moment, specifically calculates the feature quantity for each second, performs PCA dimensionality reduction on the calculated feature quantity, and imports it into the trained model to output the current health status.

[0042] Based on the above method, the present invention can also provide a bearing health status assessment system of a deep latent variable generation model or a storage medium storing a computer program; when the system or computer program is running, it is used to implement the steps of the above method.

[0043] Therefore, the technical solution of the present invention is based on the GRU-VAE network, combined with particle swarm optimization (PSO), principal component analysis (PCA) and multi-layer perceptron (MLP) for bearing health assessment. The potential information mining and unsupervised clustering training of the deep latent variable clustering model are used to objectively divide the different stages of bearing degradation (healthy, decayed or warning, fault), ensuring the objectivity and accuracy of the degradation stage division.

[0044] Compared with the prior art, the beneficial effects of this application are:

[0045] PCA dimensionality reduction is performed on the optimized features, and the principal components are smoothed using exponential weighted moving averages. The use of PCA dimensionality reduction technology and exponential weighted moving averages in data preprocessing and feature selection ensures the accuracy and real-time performance of the model.

[0046] GRU is used to capture time series features, VAE maps data to latent space, and MLP is used to classify health status. The GRU-VAE and MLP combined method and model architecture are creatively used to solve the objective factors of big data analysis of bearing failures.

[0047] GRU-VAE and MLP are combined and applied to the assessment of bearing health status, and PSO optimization and PCA dimensionality reduction technology are used to improve the performance and real-time performance of the model. Combining optimization algorithms and automated data processing technology, real-time monitoring and rapid response capabilities for dynamic equipment health status are achieved. Automatic detection and assessment of the health status of bearings during the degradation process.

[0048] High computational efficiency: GRU is used to replace complex networks such as LSTM that process time series, simplifying the model structure and reducing computational costs and resource consumption.

[0049] Strong real-time performance: PCA dimensionality reduction improves the model's real-time monitoring capability, making it suitable for application scenarios that require immediate health status feedback.

[0050] Automation and intelligence: The MLP algorithm is integrated to realize automatic health status classification, which reduces manual intervention and improves the intelligence level of the system.

[0051] Strong generalization ability: Optimization algorithms and dimensionality reduction techniques enhance the adaptability and generalization ability of the model under different working conditions and equipment types, and improve the stability and accuracy of predictions.

[0052] Improved interpretability: Through automated feature selection and PCA dimensionality reduction, the interpretability of model results is improved, making it easier for engineers and decision makers to understand and apply. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 This is a flow chart of the bearing health status assessment method based on the deep latent variable clustering model of the present invention.

[0055] Figure 2 This is a network structure diagram of the present invention.

[0056] Figure 3 This is the time domain diagram of the bearing full life vibration signal of the verification test 1 of the present invention.

[0057] Figure 4 This is a curve diagram of the network state division of the bearing full life vibration signal of the verification test 1 of the present invention.

[0058] Figure 5 This is the time domain diagram of the bearing full life vibration signal of the verification test 2 of the present invention.

[0059] Figure 6 This is a graph of the network state division of the bearing full life vibration signal of the verification test 2 of the present invention.

[0060] Figure 7 This is the time domain diagram of the bearing full life vibration signal of the verification test 3 of the present invention.

[0061] Figure 8 This is a curve diagram of the network state division of the bearing full life vibration signal of the verification test 3 of the present invention. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0063] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0064] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0065] In the present application, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0066] The features and performance of the present application are further described in detail below in conjunction with the embodiments.

[0067] Example 1

[0068] According to the present invention, an automatic detection method and system for evaluating the health status of a bearing during the degradation process of use, see Example 1, a bearing health status evaluation method based on a deep latent variable clustering model is as follows: Figure 1 As shown,

[0069] The specific steps include:

[0070] S1: Signal acquisition. Install the acceleration sensor to collect the original acceleration vibration signal data of the bearing during its entire life cycle;

[0071] S2: Feature extraction. Calculate the time domain and frequency domain features of the bearing for 1 second every 1 minute, and obtain a series of feature sequence features for each feature. and time series The features include: through-frequency velocity effective value, low-frequency velocity effective value, acceleration effective value, acceleration peak value, vibration impact value, acceleration kurtosis, acceleration skewness, velocity peak value, peak factor, acceleration signal power spectrum;

[0072] S3: Feature engineering. Using the monotonicity, robustness and correlation criteria, the comprehensive score of each indicator is evaluated, and the comprehensive scores above the baseline of 70 points are screened out, which are consistent with the characteristics of the bearing degradation process.

[0073] The monotonicity evaluation index is defined as:

[0074] ;

[0075] In the formula, is the difference between the rankings of the two variables; n is the number of data points, which is a natural number; ,The larger the monotonicity index is, the stronger the characteristic monotonicity trend is as the bearing performance degradation increases.

[0076] The robustness evaluation index is defined as:

[0077] ;

[0078] In the formula, Representative i The eigenvalues ​​corresponding to the sample points are Representative i The average trend characteristic value obtained after smoothing the sample points.

[0079] The correlation index is defined as:

[0080]

[0081] In the formula, It is the characteristic i elements. is the time series i elements. yes x The mean of . yes t The mean of .

[0082] The comprehensive score is:

[0083] ;

[0084] In the formula, , , They are weight indicators of monotonicity, robustness, and correlation, with weights of 0.4, 0.3, and 0.3 respectively.

[0085] S4: Feature dimensionality reduction. Perform PCA dimensionality reduction on the selected parameters. When the cumulative variance reaches more than 95%, stop increasing the dimension and then select the principal component with the highest explained variance ratio.

[0086] Principal component analysis (PCA) is a statistical method used to transform a set of possibly correlated variables into a set of linearly uncorrelated variables, called principal components, through an orthogonal transformation. PCA is used to reduce the dimensionality of sampling point features in order to more clearly display the signal distribution network status.

[0087] S5: Smoothing: Use exponentially weighted moving average to smooth the reduced-dimensional curve.

[0088] The exponentially weighted moving average is defined as:

[0089] ;

[0090] In the formula, It's at the time t The exponentially weighted moving average of ; It's at the time t The original data value of is the smoothing coefficient, usually in the range of , which represents the weight of the current data.

[0091] S6: State division. Calculate the variance by intercepting the first 30 health data points of the whole life data and mean According to the bearing experiments of multiple full life curves, the data point value is less than Marked as the health stage, and Between is the recession stage, greater than as the fault stage, and then normalized as the training set;

[0092] S7: Build a network. Build a two-gate GRU network connected to a VAE network, and build a three-layer MLP network. Figure 2 The network structure of the present invention involving state transition and data processing is shown, which includes encoding and decoding processes, and uses activation functions to enhance the nonlinear processing capability of the network.

[0093] Among them, the state label may represent the hidden state in the network or the label of the time step (t, t+1). In many types of networks, such as recurrent neural networks (RNN) or long short-term memory networks (LSTM), state labels are used to represent the state of the network at different time steps. The decoder is usually responsible for converting the internal representation of the network into output. In sequence generation tasks, the decoder receives the encoded information from the encoder and converts it into the desired output format, such as images, text, or other types of data. The encoder's task is to convert the input data into a compact internal representation, usually in the form of a vector. This representation captures the important features of the input data and provides a basis for further processing by the decoder.

[0094] The hyperbolic tangent function tanh is used as the respective activation function of the GRU network of the gate. It maps the input values ​​to values ​​between -1 and 1, which helps introduce nonlinearity and enhances the expressiveness of the network.

[0095] S8: Parameter optimization. The hidden layer dimension hidden_dim of GRU and the potential feature dimension latent_dim of VAE network are taken as the objects of PSO optimization, and the fitness function is selected to minimize the network loss function.

[0096] S9: Training model. Import the prepared training set into the parameter-optimized GRU-VAE network for signal reconstruction and mining of potential features. Then import the reconstructed high-dimensional signal into the MLP network for supervised classification. Finally, save the trained model.

[0097] S10: Real-time monitoring. Collect sensor vibration signals in real time, calculate the optimized feature quantity every second, perform PCA dimension reduction on the calculated feature quantity, import it into the trained model, and output the current health status.

[0098] Compared to the existing time domain diagram of the bearing's full life vibration signal, PCA can help identify which vibration modes are important and which may be caused by noise. Through the state division after PCA dimension reduction, the health status and potential failure modes of the bearing can be more accurately identified, thereby improving the accuracy of predictive maintenance and fault diagnosis. It helps to extract meaningful information from complex data and identify key fault characteristics.

[0099] Correspondingly, the bearing health status assessment method based on the deep latent variable clustering model implemented according to the present invention can also be integrated into electronic equipment to form a bearing health status assessment system based on the deep latent variable clustering model to monitor and diagnose bearing faults.

[0100] Alternatively, preferably, the carrier is a storage medium, which is stored in the storage medium in the form of a computer program. When the computer program is executed, the steps of the bearing health status assessment method based on the deep latent variable clustering model are implemented to monitor and diagnose bearing faults.

[0101] Example 2

[0102] Based on Example 1, the comprehensive scores of various features are obtained according to step S3 as shown in Table 1.

[0103] Table 1 Comprehensive scores of each feature

[0104]

[0105] The variance is calculated according to step S6 =0.1536 and mean =-1.524, and the decline stage and failure stage are divided according to the thresholds of -1.0627 and 0.7809.

[0106] According to the PSO particle swarm optimization setting parameters in step S8, as shown in Table 2:

[0107] Table 2 PSO parameter settings

[0108]

[0109] Other parameters can be left as default.

[0110] Example 3

[0111] In the examples that can be selected, the applicant used the Xi'an Jiaotong University's public bearing data set to conduct multiple experimental verifications, and the results proved the feasibility of the invention.

[0112] like Figure 3 and 4 In the verification test 1, the vibration of the bearing throughout its life was verified. Figure 3 The horizontal axis of the full life vibration time domain diagram represents time in minutes. The time range is from 0 to 160 minutes, which represents the operation or test cycle of the embodiment equipment. The vertical axis vibration signal represents the amplitude of vibration, and the vibration amplitude ranges from -40 to 30Hz.

[0113] Figure 4 This is the curve after network state division (PCA dimension reduction + state division). In this figure, PCA is used to reduce the dimension of sampling point features in order to more clearly display the network state. The vertical axis represents the feature value after PCA dimension reduction. Figure 3Different divisions are made based on the network's behavior or performance at different time points, and each state corresponds to a different health state of the network. 0-80 minutes is a healthy state, 80-100 minutes is a warning state, and after 100 minutes is a fault state.

[0114] visible Figure 3 The sampling signal vibrates violently at 100 minutes, but it cannot be found at the normal detection level before 100 minutes. After adopting the present invention, Figure 4 At 80 minutes, the system prompts that it is in an unhealthy state and starts to alarm, which greatly improves the accuracy of health monitoring and avoids the equipment from working with problems when it exceeds the healthy state, causing greater failures, providing timely assistance for equipment monitoring and maintenance at the production site.

[0115] Example 4

[0116] like Figure 5 and 6 In the verification test 2, compared with the embodiment 3, this embodiment selected a shorter monitoring time (70 minutes) and a larger vibration amplitude variation range (-50 to 50 Hz) for experimental verification.

[0117] Likewise, in Figure 6 The health status was detected to be problematic at 36-40 minutes in the middle timeline, and a warning was issued, and it was officially guaranteed at 40 minutes. Figure 5 There is no obvious signal jump in the time period from 40 minutes to 45 minutes. At 45 minutes, the signal amplitude begins to increase slowly, and at 50 minutes, there is an obvious fault. Figure 6 The monitoring ratio Figure 5 It was nearly 10 minutes earlier than expected, avoiding the equipment from operating with problems when it was beyond a healthy state and causing greater failures, providing timely assistance for equipment monitoring and maintenance at the production site.

[0118] Example 5

[0119] like Figure 7 The time domain diagram of the bearing life vibration signal in the verification test 3 and Figure 8 The curve diagram after network status division in verification test 3, compared with embodiment 3, this embodiment selects a longer monitoring time (250 minutes) and a larger vibration amplitude variation range (-40 to 40 Hz) for experimental verification.

[0120] For long-term monitoring, the method and system of the present invention are also very sensitive and can fully ensure accurate detection and assessment of health status.

[0121] exist Figure 8 The problem was discovered at around 43 minutes and a warning was issued, and the fault was diagnosed at around 60 minutes. Figure 7The signal amplitude starts to change slowly around 53 minutes, and an obvious sudden jump fault appears around 75 minutes.

[0122] The embodiments described above are part of the embodiments of the present application, rather than all of the embodiments. The detailed description of the embodiments of the present application is not intended to limit the scope of the present application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

Claims

1. A bearing health status assessment method based on a deep latent variable clustering model, characterized in that The steps include: (1) Collect vibration signals of the bearing throughout its life cycle; (2) Extract and calculate the time domain and frequency domain features of the vibration signal according to the set time interval, and obtain a string of feature sequence features and time series for each feature; (3) Based on the comprehensive score weight formula, calculate the comprehensive score of each feature and select the feature parameters whose comprehensive scores are above the set standards; the comprehensive score weight formula is: ； In the formula, , , They are monotonicity indices Mon , robustness index Rob , correlation index Rec The weight index of Score is the comprehensive score; (4) Perform PCA dimensionality reduction on the selected feature parameters. When the cumulative variance reaches a set percentage, stop increasing the dimension and select the principal component with the highest explained variance ratio. (5) Use exponentially weighted moving average to smooth the reduced-dimensional curve; (6), use In principle, the curves after dimension reduction and smoothing are divided into healthy stage, decay stage and failure stage, and then normalized as training sets; (7) Build a network model: Build a neural network including GRU and VAE for data reconstruction, and build an MLP network. Use the MLP network to classify the health status of the reconstructed signal; (8) Use PSO to optimize the parameters of the hidden layer size of GRU and the potential feature size of VAE; (9) Import the training set into the network model after parameter optimization for training, and output the trained model; (10) Monitor the bearing vibration signal in real time, calculate the time domain and frequency domain parameters at the current moment, and import them into the trained model after PCA dimension reduction to evaluate the current bearing health status.

2. The bearing health status assessment method based on the deep latent variable clustering model according to claim 1 is characterized in that The characteristics of step (2) include at least one of the following: effective value of full-frequency speed, effective value of low-frequency speed, effective value of acceleration, acceleration peak value, vibration impact value, acceleration kurtosis, acceleration skewness, velocity peak value, peak factor, and acceleration signal power spectrum.

3. The bearing health status assessment method based on the deep latent variable clustering model according to claim 1 is characterized in that The monotonicity index of step (3) is: ； In the formula, is the difference between the rankings of the two variables; n is the number of data points; ,The larger the monotonicity index is, the stronger the characteristic monotonicity trend is as the bearing performance degradation increases.

4. The bearing health status assessment method based on the deep latent variable clustering model according to claim 1 is characterized in that The correlation indicators are: ； In the formula, It is the characteristic i elements; is the time series i elements; yes x The mean of yes t The mean of .

5. The bearing health status assessment method based on the deep latent variable clustering model according to claim 1 is characterized in that The screening target is a comprehensive score of 70 points or more for each feature.

6. The bearing health status assessment method based on the deep latent variable clustering model according to claim 1 is characterized in that Step (4) When the cumulative variance reaches more than 95%, stop increasing the dimension.

7. The bearing health status assessment method based on the deep latent variable clustering model according to claim 1 is characterized in that Step (6) Calculate the variance of several health data points before intercepting the whole life data and mean According to the bearing experiments of multiple full life curves, the data point value is less than Marked as the health stage, and Between is the recession stage, greater than as the fault stage, and then normalized as the training set.

8. The bearing health status assessment method based on the deep latent variable clustering model according to claim 1 is characterized in that A GRU-VAE network is constructed, which is composed of a double-gate GRU network connected to a VAE network. The GRU-VAE network is used to reconstruct the data. In addition, a three-layer MLP network is built, and the MLP network is used to divide the health status of the reconstructed signal.

9. The bearing health status assessment method based on the deep latent variable clustering model according to claim 1 is characterized in that Step (8) takes the hidden layer dimension hidden_dim of GRU and the potential feature dimension latent_dim of VAE network as the objects of PSO optimization, and the fitness function is selected to minimize the network loss function.

10. A bearing health status assessment system based on a deep latent variable generation model, characterized in that When the system is running, it is used to implement the steps of the method described in any one of claims 1 to 9.

Images