Rolling bearing performance degradation evaluation method based on electrostatic monitoring

Abstract

The invention provides a rolling bearing performance degradation evaluation method based on electrostatic monitoring. The rolling bearing performance degradation evaluation method comprises the following steps: S1, fusing electrostatic charge signal features with conventional time domain, frequency domain and time-frequency domain features to form an electrostatic monitoring system feature parameter set representing a bearing state; the invention provides a rolling bearing performance degradation evaluation method based on electrostatic monitoring. According to the rolling bearing performance degradation evaluation method based on electrostatic monitoring, nonlinear features and different degradation stages of a bearing can be better identified through a deep noise reduction auto-encoder optimized by L1 sparse constraint and a Dropout mechanism; the provided adaptive density peak value clustering model can well measure the health state of the bearing, and compared with traditional DPC and SVDD, the method provided by the invention can recognize early degradation of the bearing earlier, and has better monotonicity, robustness and tendency; compared with vibration monitoring, static monitoring has superiority in bearing performance degradation evaluation, and degradation can be found earlier.

Description

Technical Field

[0001] The invention relates to the field of rolling bearings, and in particular to a rolling bearing performance degradation assessment method based on electrostatic monitoring. Background Art

[0002] The performance of rolling bearings largely determines whether the operation of mechanical equipment is safe and stable, because their degradation or failure will affect the performance of the entire machine. Under complex operating conditions, bearings are prone to performance degradation and failure as the operating time accumulates. If the beginning time of bearing degradation can be discovered as early as possible, there will be enough time to arrange maintenance reasonably and reduce the occurrence of accidents and economic losses. Therefore, it is very important to evaluate the performance degradation of bearings.

[0003] Conventional methods for online monitoring of rolling bearings include vibration monitoring, oil particle analysis, and temperature monitoring. Among them, online monitoring technology based on vibration signals is widely used in the study of bearing performance degradation. However, the internal structure of mechanical equipment is complex, and it is difficult to avoid the coupling of multiple excitation sources during operation. Its vibration signal often has a strong path noise influence. At present, electrostatic monitoring technology has become a very promising rolling bearing condition monitoring technology with its high sensitivity. This technology evaluates the operating status of the bearing by monitoring the change in the total amount of charge in the wear area of ​​the bearing. Its advantage is that it can directly measure the contact degradation of the bearing, rather than the secondary effect caused by degradation. Therefore, compared with vibration monitoring, electrostatic monitoring technology has the advantage of earlier identification of early degradation. Although electrostatic monitoring has high sensitivity, the signal characteristics of the bearing will be affected by many factors. Therefore, how to construct a health indicator that effectively reflects the operating status of the rolling bearing has become a research hotspot in the current field of equipment condition monitoring.

[0004] The existing rolling bearing performance degradation assessment method does not analyze the deep characteristics of the electrostatic signal data to construct a health indicator that can better reflect the degradation process. A single degradation indicator contains less fault information and has more external interference. It can only perform linear dimensionality reduction on the data. In order to map the nonlinear characteristics of the bearing electrostatic data, an autoencoder can be introduced to analyze and process the high-dimensional features. The selection of parameters is still highly subjective, which can easily affect the accuracy of the results and is not effective when the fluctuation is large. A linear regression model is established on a sliding window to repair the fluctuation. Although the sliding window can realize online monitoring and repair, the linear regression model lacks the ability to characterize the data fluctuation trend.

[0005] Therefore, it is necessary to provide a rolling bearing performance degradation assessment method based on electrostatic monitoring to solve the above technical problems. Summary of the invention

[0006] The present invention provides a rolling bearing performance degradation assessment method based on electrostatic monitoring, which solves the problem that the existing rolling bearing performance degradation assessment method does not analyze the deep characteristics of electrostatic signal data to construct a health indicator that can better reflect the degradation process. A single degradation indicator contains less fault information and has more external interference. It can only perform linear dimensionality reduction on the data. In order to map the nonlinear characteristics of the bearing electrostatic data, an autoencoder can be introduced to analyze and process the high-dimensional features. The selection of parameters is still relatively subjective, which can easily affect the accuracy of the results and has poor effect when the fluctuation is large. A method for establishing a linear regression model on a sliding window to repair the fluctuation. Although the sliding window can realize online monitoring and repair, the linear regression model lacks the ability to characterize the data fluctuation trend.

[0007] In order to solve the above technical problems, the present invention provides a rolling bearing performance degradation assessment method based on electrostatic monitoring, comprising the following steps S1: integrating electrostatic charge signal characteristics with conventional time domain, frequency domain and time-frequency domain characteristics to form an electrostatic monitoring system characteristic parameter set that characterizes the bearing state;

[0008] S2: The deep denoising autoencoder optimized with L1 regularization and Dropout mechanism is introduced to perform nonlinear mapping and dimensionality reduction on high-dimensional feature sets. IDDAE can not only improve the ability of feature selection, but also prevent overfitting of deep networks.

[0009] S3: Adaptive determination of DPC cluster centers is achieved through comprehensive optimization and entropy analysis, thereby avoiding the "chain effect" caused by incorrect allocation of cluster centers. Therefore, a model is established through the improved ADPC algorithm to measure the health status of the bearing, and the Mahalanobis distance MD between the test set and the cluster center is calculated to quantitatively characterize the performance degradation degree of the rolling bearing;

[0010] S4: Combined with the 3δ principle, a polynomial regression model is established on the sliding window to repair the false fluctuations of the MD index and obtain the improved index IMD that can accurately characterize the degradation process of the rolling bearing. The bearing performance degradation evaluation is realized by analyzing the relationship curve between the health threshold and IMD.

[0011] Preferably, the algorithm is improved in three ways. The first is to build a 5-layer deep denoising autoencoder, introduce L1 regularization, use its sparsity to improve the data feature extraction capability, and introduce the Dropout mechanism to prevent overfitting caused by network deepening; the second is to solve the problem that the traditional density peak clustering algorithm manually selects clustering centers and cutoff distances, resulting in a large difference between the clustering effect and the expected value, and propose an adaptive density peak clustering algorithm, and improve the original indicator calculation method through the Mahalanobis distance; the third is to establish a polynomial regression model on the sliding window to repair false fluctuations.

[0012] Preferably, the method for verifying the effectiveness of the rolling bearing performance degradation assessment is to conduct an accelerated fatigue life experiment of the bearing on an ABLT-1A testing machine. The monitoring platform includes a loading system, a transmission system, a lubrication system, an electrical control system, a computer monitoring system, an experimental head and an experimental head seat; the experimental head includes an experimental device, an electrostatic sensor installation position and a signal acquisition system, and the electrostatic sensor ES1 and the electrostatic sensor ES2 are used to monitor the electrostatic signal of the bearing.

[0013] Preferably, the L1 regularization is the sum of the absolute values ​​of each element in the weight vector, which guides the model to pay more attention to weights with larger absolute values ​​and continuously adjusts the feature weights so that the weights of some features become 0, thereby achieving a sparse effect.

[0014] Preferably, the principle of L1 generating a sparse model is as follows:

[0015]

[0016] In the formula, J0 is the loss function of the original model, and λ is the regularization coefficient;

[0017] In the field of machine learning, L1 regularization is often used in sparse models. Therefore, in the loss function of the deep denoising autoencoder, the L1 norm penalty term is added to achieve feature selection by minimizing the sum of the objective function and the L1 norm of the feature. This not only makes the model more robust and generalizable, but also enhances the feature extraction of electrostatic signals. At this time, the loss function of the deep denoising autoencoder after adding L1 regularization is:

[0018]

[0019] Adding L1 regularization to the loss function makes it easier to make data sparse, which will also reduce the difficulty of the learning task accordingly. It is also conducive to the automatic selection of features, making the model interpretable and providing good data for the degradation assessment model.

[0020] Preferably, the Dropout is a technique that can help reduce overfitting in training neural networks. In this study, the deep denoising autoencoder is composed of multiple layers of denoising autoencoders. As the complexity of the neural network increases, overfitting is prone to occur during model training. Therefore, the Dropout mechanism is introduced to solve the overfitting problem, so as to prevent complex co-adaptation on the training data and avoid repeated extraction of the same features.

[0021] In the training process of deep learning networks, Dropout temporarily and randomly discards hidden layer neurons according to the set probability. In the next training, different neurons will be recombined according to the set probability. After multiple Dropout processes, the original network will become a combination of multiple thinner neural networks. The calculation process of Dropout in the neural network can be expressed as:

[0022] y (l+1) =σ(W (l)T x (l) +b (l) )1≤l≤L-1

[0023] x (l) = dropout(y (l) ,p) 2≤l≤L

[0024] x (l+1) =(1-p)σ(W (l)T x (l) +b (l) )

[0025] Among them, l represents the number of neuron layers, L is the total number of neuron layers, and p represents the dropout rate;

[0026] By training the same set of data through different neural networks and averaging the outputs of each network, the purpose of reducing overfitting can be achieved. In practice, the activation outputs of some hidden neurons can be set to 0, so that these neurons are not included in the forward propagation training process, thereby optimizing the deep denoising autoencoder through dropout.

[0027] Preferably, the bearing abnormal data accounts for a very small part of the full life cycle data, and there is an extreme sample imbalance between normal samples and abnormal samples. The density peak clustering algorithm efficiently allocates sample points and removes outliers by finding density peak points. It is suitable for clustering analysis of large-scale bearing data, and uses ADPC to perform unsupervised learning on bearing data to construct the health status of the bearing.

[0028] Preferably, the fast search clustering algorithm DPC for density peak is a new density-based clustering algorithm, which is based on the following assumptions: the local density of the cluster center is higher than the local density of other points in the same cluster, and there is a large distance between any points with higher local density. When the normal state of electrostatic data is modeled, the cluster center has a large local density value and relative distance value;

[0029] In order to find the cluster center that satisfies both conditions, the algorithm introduces the definition of local density and relative distance. The traditional density peak clustering algorithm uses Euclidean distance to reflect the relationship between two points. However, Euclidean distance is easily affected by the feature scale. Therefore, the introduction of Mahalanobis distance to support DPC distance calculation can effectively integrate the characteristics of multiple feature vectors. The calculation formula of Mahalanobis distance between data points is:

[0030]

[0031] In the formula, S -1 is the covariance matrix. At this time, the local density calculation process is:

[0032]

[0033] When x<0, χ(x)=1, otherwise χ(x)=0; MD ij is the Mahalanobis distance between data points i and j, d c In order to truncate the distance and avoid the same local density of data points, the Gaussian function is used to replace the formula, as shown in the formula;

[0034]

[0035] The relative distance calculation process is:

[0036]

[0037] In the formula, S represents the set of data, and there are n samples in the set; represents the density ratio ρ i A collection of large points; represents the empty set;

[0038] Choose a value that can comprehensively consider the local density ρ i and the relative distance δ i The parameter μ i :

[0039] μ i =ρ i δ i (3)

[0040] In the DPC algorithm, the cluster center has a larger local density ρ than other points. i and the relative distance δ i Therefore, by using formula (18), we can increase the μ of the cluster center i μ with other points i The difference in values ​​is used to calculate μ i Then, μ i The values ​​are sorted in descending order, and then μ is calculated iThe downward trend of the value is calculated as follows:

[0041]

[0042] In the formula, μ i is the current μ value; μ i-1 and μ i+1 Represent the μ values ​​of the previous moment and the next moment respectively. After obtaining the downward trend, take the point with the largest downward trend and the previous points as the clustering center;

[0043] Traditional DPC algorithms usually rely on experience or experimental verification to obtain d c The value makes the clustering result have a lot of uncertainty. In order to improve the clustering effect, d can be determined based on the difference between the comprehensive measurement data. c This paper uses the entropy method to constrain the value of d c In information theory, the amount of information contained in an event is measured by information entropy. In comprehensive evaluation, if the numerical difference of the same indicator is small, that is, the probability distribution is more uniform, it means that the event "a certain value belongs to a certain evaluated object" has greater uncertainty and information entropy. Therefore, a lower weight is given to the indicator. On the contrary, if the numerical difference of the same indicator is large, the weight of the indicator is higher.

[0044] The calculation formula of information entropy is:

[0045]

[0046] Where n represents the number of samples in the data set S; p i is the ρ of point i i ×δ i The proportion of all points, p i The values ​​are as follows:

[0047]

[0048] The minimum value of d c The value is used as the final clustering d c value, instead of empirical method or experimental verification, so as to achieve adaptive selection of parameters.

[0049] Preferably, the performance of the bearing is in a state of continuous decline during operation. Due to some external factors, the decline characteristics may fluctuate abnormally, which will seriously affect the description of the degradation state of the bearing by the degradation index. Therefore, if there is a false fluctuation, it needs to be repaired before it can be used as an indicator for subsequent degradation evaluation.

[0050] In this paper, a sliding window is constructed based on health indicators. The window step size is set to 1, and the window length is 50. The fluctuation threshold θ is set according to the 3σ principle of the data within the sliding window and polynomial regression. Its expression is:

[0051]

[0052] Among them, the values of ω0, ω1, and ω2 are determined by the least squares method; x i is the time series of health indicators, and MD i is the distance value between the data to be measured and the clustering center;

[0053] According to the sliding window fluctuation threshold θ, the health indicator IMD after repairing false fluctuations is determined i , and its expression is:

[0054]

[0055] Among them, min{MD(i)} and max{MD(i)} represent the upper and lower limits of MD(i) respectively;

[0056]

[0057] In the formula, μ represents the mean of MD(i); σ represents the standard deviation of MD(i);

[0058] According to the above formula, it is judged whether there are false fluctuations in the bearing health indicator and repaired. When MD(i) ≤ min{MD(i)}, MD(i) moves to the average value to obtain a new predicted value IMD; when min{MD(i)} < MD(i) < max{MD(i)}, MD(i) is replaced by the value predicted by polynomial regression; when MD(i) ≥ max{MD(i)}, MD(i) moves to max{MD(i)}, and then a new predicted value is calculated. By establishing a regression model on the sliding window, the false fluctuations existing in the health indicator can be effectively repaired, so as to obtain a health indicator with better performance in characterizing the degradation trend.

[0059] Preferably, the experimental device includes an experimental device main body 100. A base 200 is fixedly installed at the bottom of the experimental device main body 100. Hydraulic cylinders 300 are fixedly installed around the top of the base 200. The bottom end of the output shaft of the hydraulic cylinder 300 is fixedly connected to a support block 400. Movable wheels 500 are arranged around the bottom of the base 200.

[0060] Compared with the related technology, a rolling bearing performance degradation evaluation method based on electrostatic monitoring provided by the present invention has the following beneficial effects:

[0061] The present invention provides a rolling bearing performance degradation assessment method based on electrostatic monitoring. The method realizes nonlinear dimensionality reduction of a high-dimensional feature set by using a deep denoising autoencoder after L1 sparse constraint, eliminates data redundancy, and introduces a Dropout mechanism to prevent model overfitting. An improved degradation model ADPC is constructed, and the Mahalanobis distance is introduced to calculate the length of a data point. The entropy method is used to find a cutoff distance that satisfies the distribution characteristics of electrostatic data, and an indicator that can simultaneously reflect the local density and the relative distance size is constructed. The point with the largest trend value is the cluster center. Compared with the traditional DPC, the improved model can adaptively select the cluster center and the cutoff distance to reduce the error of empirical selection. Moreover, compared with the original Euclidean distance, the Mahalanobis distance is not affected by the dimension and takes into account the correlation between variables. In order to meet the requirements of online monitoring to repair false fluctuations in the constructed degradation indicator, the set The sliding window is based on the principle of judgment. A polynomial regression model is established in the window. Nonlinear problems can be handled by adding new variables. The electrostatic charge signal feature is used to reflect the effective value of the charge in the signal. In electrostatic monitoring, it can be used as a good supplement to conventional time domain and frequency domain indicators. Compared with the traditional dimensionality reduction method PCA, the deep denoising autoencoder optimized by L1 sparse constraint and Dropout mechanism can better identify the nonlinear characteristics and different degradation stages of the bearing. The proposed adaptive density peak clustering model can measure the health status of the bearing well. Compared with traditional DPC and SVDD, the method proposed in this paper can identify the early degradation of the bearing earlier, and has better monotonicity, robustness and trend. Compared with vibration monitoring, electrostatic monitoring is superior in bearing performance degradation assessment and can detect the occurrence of degradation earlier. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic flow chart of a first embodiment of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention;

[0063] Figure 2 A schematic diagram of the DDAE network structure of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention;

[0064] Figure 3 A schematic diagram of a network improved by using Dropout for a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention;

[0065] Figure 4 A DPC principle diagram of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention;

[0066] Figure 5 A schematic diagram of the sliding window feature repair principle of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention;

[0067] Figure 6 A schematic diagram of the overall process of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention;

[0068] Figure 7 An electrostatic monitoring principle diagram of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention;

[0069] Figure 8 A schematic structural diagram of a second embodiment of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention.

[0070] Numbers in the figure: 100, main body of the experimental device, 200, base, 300, hydraulic cylinder, 400, supporting block, 500, moving wheel. DETAILED DESCRIPTION

[0071] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.

[0072] First embodiment

[0073] Please refer to Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 ,in, Figure 1 A schematic flow chart of a first embodiment of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention; Figure 2 A schematic diagram of the DDAE network structure of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention; Figure 3 A schematic diagram of a network improved by using Dropout for a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention; Figure 4 A DPC principle diagram of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention; Figure 5 A schematic diagram of the sliding window feature repair principle of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention; Figure 6 A schematic diagram of the overall process of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention; Figure 7 The electrostatic monitoring principle diagram of a rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention. A rolling bearing performance degradation assessment method based on electrostatic monitoring comprises the following steps S1: integrating electrostatic charge signal characteristics with conventional time domain, frequency domain and time-frequency domain characteristics to form an electrostatic monitoring system characteristic parameter set that characterizes the bearing state;

[0074] S2: The deep denoising autoencoder optimized with L1 regularization and Dropout mechanism is introduced to perform nonlinear mapping and dimensionality reduction on high-dimensional feature sets. IDDAE can not only improve the ability of feature selection, but also prevent overfitting of deep networks.

[0075] S3: Adaptive determination of DPC cluster centers is achieved through comprehensive optimization and entropy analysis, thereby avoiding the "chain effect" caused by incorrect allocation of cluster centers. Therefore, a model is established through the improved ADPC algorithm to measure the health status of the bearing, and the Mahalanobis distance MD between the test set and the cluster center is calculated to quantitatively characterize the performance degradation degree of the rolling bearing;

[0076] S4: Combined with the 3δ principle, a polynomial regression model is established on the sliding window to repair the false fluctuations of the MD index and obtain the improved index IMD that can accurately characterize the degradation process of the rolling bearing. The bearing performance degradation evaluation is realized by analyzing the relationship curve between the health threshold and IMD.

[0077] The algorithm has been improved in three ways. The first is to build a 5-layer deep denoising autoencoder, introduce L1 regularization, use its sparsity to improve the data feature extraction capability, and introduce the Dropout mechanism to prevent overfitting caused by network deepening; the second is to solve the problem that the traditional density peak clustering algorithm manually selects clustering centers and cutoff distances, resulting in a large difference between the clustering effect and the expected value, and propose an adaptive density peak clustering algorithm, and improve the original indicator calculation method through the Mahalanobis distance; the third is to establish a polynomial regression model on the sliding window to repair false fluctuations.

[0078] The method for verifying the effectiveness of the rolling bearing performance degradation evaluation is to conduct a bearing accelerated fatigue life test on an ABLT-1A testing machine. The monitoring platform includes a loading system, a transmission system, a lubrication system, an electrical control system, a computer monitoring system, an experimental head and an experimental head seat; the experimental head includes an experimental device, an electrostatic sensor installation position and a signal acquisition system, and the electrostatic sensor ES1 and the electrostatic sensor ES2 are used to monitor the electrostatic signal of the bearing.

[0079] The L1 regularization is the sum of the absolute values ​​of each element in the weight vector. It guides the model to pay more attention to weights with larger absolute values ​​and continuously adjusts the feature weights so that the weights of some features become 0, thereby achieving a sparse effect.

[0080] The principle of L1 generating a sparse model is as follows:

[0081]

[0082] In the formula, J0 is the loss function of the original model, and λ is the regularization coefficient;

[0083] In the field of machine learning, L1 regularization is often used in sparse models. Therefore, in the loss function of the deep denoising autoencoder, the L1 norm penalty term is added to achieve feature selection by minimizing the sum of the objective function and the L1 norm of the feature. This not only makes the model more robust and generalizable, but also enhances the feature extraction of electrostatic signals. At this time, the loss function of the deep denoising autoencoder after adding L1 regularization is:

[0084]

[0085] Adding L1 regularization to the loss function makes it easier to make data sparse, which will also reduce the difficulty of the learning task accordingly. It is also conducive to the automatic selection of features, making the model interpretable and providing good data for the degradation assessment model.

[0086] The Dropout is a technique that can help reduce overfitting of training neural networks. In this study, the deep denoising autoencoder is composed of multiple layers of denoising autoencoders. As the complexity of the neural network increases, overfitting is prone to occur during model training. Therefore, the Dropout mechanism is introduced to solve the overfitting problem to prevent complex co-adaptation on the training data and avoid repeated extraction of the same features.

[0087] In the training process of deep learning networks, Dropout temporarily and randomly discards hidden layer neurons according to the set probability. In the next training, different neurons will be recombined according to the set probability. After multiple Dropout processes, the original network will become a combination of multiple thinner neural networks. The calculation process of Dropout in the neural network can be expressed as:

[0088] y (l+1) =σ(W (l)T x (l) +b (l) ) 1≤l≤L-1

[0089] x (l) = dropout(y (l) ,p) 2≤l≤L

[0090] x (l+1) =(1-p)σ(W (l)T x (l) +b (l) )

[0091] Among them, l represents the number of neuron layers, L is the total number of neuron layers, and p represents the dropout rate;

[0092] By training the same set of data through different neural networks and averaging the outputs of each network, the purpose of reducing overfitting can be achieved. In practice, the activation outputs of some hidden neurons can be set to 0, so that these neurons are not included in the forward propagation training process, thereby optimizing the deep denoising autoencoder through dropout.

[0093] The bearing abnormal data accounts for a very small part of the full life cycle data. There is an extreme sample imbalance between normal samples and abnormal samples. The density peak clustering algorithm efficiently allocates sample points and removes outliers by finding density peak points. It is suitable for clustering analysis of large-scale bearing data. ADPC is used to perform unsupervised learning on bearing data to construct the health status of the bearing.

[0094] The fast search clustering algorithm DPC for density peaks is a new density-based clustering algorithm based on the following assumptions: the local density of the cluster center is higher than the local density of other points in the same cluster, and there is a large distance between any points with higher local density. When the normal state of electrostatic data is modeled, the cluster center has a large local density value and relative distance value;

[0095] In order to find the cluster center that satisfies both conditions, the algorithm introduces the definition of local density and relative distance. The traditional density peak clustering algorithm uses Euclidean distance to reflect the relationship between two points. However, Euclidean distance is easily affected by the feature scale. Therefore, the introduction of Mahalanobis distance to support DPC distance calculation can effectively integrate the characteristics of multiple feature vectors. The calculation formula of Mahalanobis distance between data points is:

[0096]

[0097] In the formula, S -1 is the covariance matrix. At this time, the local density calculation process is:

[0098]

[0099] When x<0, χ(x)=1, otherwise χ(x)=0; MD ij is the Mahalanobis distance between data points i and j, d c In order to truncate the distance and avoid the same local density of data points, the Gaussian function is used to replace the formula, as shown in the formula;

[0100]

[0101] The relative distance calculation process is:

[0102]

[0103] In the formula, S represents the set of data, and there are n samples in the set; represents the density ratio ρ i A collection of large points; represents the empty set;

[0104] Choose a value that can comprehensively consider the local density ρ i and the relative distance δ i The parameter μ i :

[0105] μ i =ρ i δ i (5)

[0106] In the DPC algorithm, the cluster center has a larger local density ρ than other points. i and the relative distance δ i Therefore, by using formula (18), we can increase the μ of the cluster center i μ with other points i The difference in values ​​is used to calculate μ i Then, μ i The values ​​are sorted in descending order, and then μ is calculated i The downward trend of the value is calculated as follows:

[0107]

[0108] In the formula, μ i is the current μ value; μ i-1 and μ i+1 Represent the μ values ​​of the previous moment and the next moment respectively. After obtaining the downward trend, take the point with the largest downward trend and the previous points as the clustering center;

[0109] Traditional DPC algorithms usually rely on experience or experimental verification to obtain d c The value makes the clustering result have a lot of uncertainty. In order to improve the clustering effect, d can be determined based on the difference between the comprehensive measurement data. c This paper uses the entropy method to constrain the value of d c In information theory, the amount of information contained in an event is measured by information entropy. In comprehensive evaluation, if the numerical difference of the same indicator is small, that is, the probability distribution is more uniform, it means that the event "a certain value belongs to a certain evaluated object" has greater uncertainty and information entropy. Therefore, a lower weight is given to the indicator. On the contrary, if the numerical difference of the same indicator is large, the weight of the indicator is higher.

[0110] The calculation formula of information entropy is:

[0111]

[0112] Where n represents the number of samples in the data set S; p i is the ρ of point i i ×δ i The proportion of all points, p i The values ​​are as follows:

[0113]

[0114] The minimum value of d c The value is used as the final clustering d c value, instead of empirical method or experimental verification, so as to achieve adaptive selection of parameters.

[0115] During the operation of the bearing, its performance is in a state of continuous decline. Due to some external factors, the decline characteristics may fluctuate abnormally, which will seriously affect the description of the degradation state of the bearing by the degradation index. Therefore, if there is a false fluctuation, it needs to be repaired before it can be used as an indicator for subsequent degradation assessment.

[0116] This paper constructs a sliding window based on health indicators, sets the window step size to 1, the window length to 50, and sets the fluctuation threshold θ based on the 3σ principle of the data in the sliding window and polynomial regression. The expression is:

[0117]

[0118] Among them, the values ​​of ω0, ω1 and ω2 are determined by the least squares method; x i For the time series of health indicators, MD i is the distance between the data to be tested and the cluster center;

[0119] Determine the health indicator IMD after repairing the false fluctuations according to the sliding window fluctuation threshold θ i , whose expression is:

[0120]

[0121] Among them, min{MD(i)} and max{MD(i)} represent the upper and lower limits of MD(i), respectively;

[0122]

[0123] Where, μ represents the mean of MD(i); σ represents the standard deviation of MD(i);

[0124] Judge whether there is false fluctuation in the bearing health index according to the above formula and repair it. When MD(i) ≤ min{MD(i)}, MD(i) moves to the average value to obtain a new predicted value IMD; when min{MD(i)} < MD(i) < max{MD(i)}, MD(i) is replaced by the value predicted by polynomial regression; when MD(i) ≥ max{MD(i)}, MD(i) moves to max{MD(i)}, and then a new predicted value is calculated. By establishing a regression model on the sliding window, the false fluctuation existing in the health index can be effectively repaired, so as to obtain a health index with better performance in characterizing the degradation trend.

[0125] Performance degradation assessment based on deep denoising autoencoder and adaptive density peak clustering:

[0126] A method of deep denoising autoencoder and adaptive peak density clustering is proposed for bearing performance degradation assessment. The bearing electrostatic signal is input into ADPC for feature fusion after dimensionality reduction by IDDAE, and the output result is used as the health index for performance degradation assessment. The health index is the distance value IMD between the sample to be measured and the clustering center after ADPC clustering. By judging the relationship between the IMD index and the threshold, it is determined whether the bearing is in a degraded state. If it exceeds the threshold, it means the bearing has degraded, otherwise the bearing is in a normal state.

[0127] The bearing performance degradation assessment is divided into an offline stage and an online stage. Combining the method proposed in the paper, the specific implementation steps of these two stages are as follows:

[0128] (1) Offline stage

[0129] A. Collect the electrostatic signals of the rolling bearing in the normal working state, and extract the original features from the bearing electrostatic signals, including 12 time-domain features, 3 frequency-domain features, 8 time-frequency domain features and 2 electrostatic charge signal features. Among them, the electrostatic charge signal features are unique parameters of electrostatic monitoring, which are divided into the active rate level (AL) and the event rate (ER).

[0130] The definition of the active rate level is:

[0131]

[0132] In the formula, Q f is the electrostatic charge signal, T is the intercepted signal length, usually taken as 1 s, N is the number of sampling points of Q f (t) within T after discrete sampling, and Q fn is Q f(t) The electrostatic charge signal after discrete sampling, the physical meaning of the activity level AL is the effective value of the induced charge on the sensor induced by the charged particles in the gas path passing through the detection space of the electrostatic sensor within a certain time interval.

[0133] The event rate ER is defined as:

[0134]

[0135] Where M is the number of points that exceed K times the AL(t) value in the electrostatic signal interval (1s signal) corresponding to each point A, K is the empirical multiple, and N is the total number of samples in this signal segment. The physical meaning of the event rate can be understood as the percentage of the number of amplitude abnormal values ​​to the total number of sampling points within a certain time interval.

[0136] B. The initially extracted original features are input into IDDAE for greedy training to learn the nonlinear features of the bearing layer by layer. In the IDDAE network used for feature dimensionality reduction, a network with a hidden layer structure of 5 layers is set, and the neurons of the hidden layers are configured to 128, 64, 32, 16, and 3 respectively. The learning rate is set to 0.001, the Gaussian noise coefficient is set to 0.2, the sparse coefficient is set to 0.05, and the Dropout discard rate is set to 0.2.

[0137] C. The feature vectors of the rolling bearing under normal conditions extracted from the IDDAE model are used as training samples for the ADPC model, and the data distribution shape and cluster center of the normal samples are found by calculating the local density and relative distance.

[0138] Online stage:

[0139] A. Same as the offline method, extract the original features of the rolling bearing electrostatic signal.

[0140] B. Input the original features into IDDAE to reconstruct the feature vector.

[0141] C. In the offline stage, the ADPC model of normal samples has been trained, and the distribution of data under normal conditions can be learned and used as a benchmark for measurement. Therefore, in the online stage, the Mahalanobis distance is used to display the data distribution difference between the data to be tested and the benchmark, thereby obtaining the indicator MD for measuring the degree of bearing degradation.

[0142] D. Because the measurement environment of online data is subject to a lot of external interference, it is inevitable that the data will have abnormal fluctuations. In order to eliminate the impact of data fluctuations on the accuracy of the results, a regression model is established on the sliding window, and the false fluctuations of the degradation trend curve of the rolling bearing are repaired to obtain a new health indicator IMD.

[0143] E. The performance degradation of rolling bearings can be evaluated by analyzing the relationship curve between the health threshold and the health indicator IMD.

[0144] Principle of rolling bearing electrostatic monitoring:

[0145] The wear area of ​​rolling bearings is accompanied by static electricity, and static electricity monitoring technology is a real-time monitoring of the total amount of static charge in the wear area. Figure 6 The diagram shows the principle diagram of electrostatic monitoring in the wear area. When charged particles pass through the effective area in front of the probe sensing surface of the electrostatic sensor, the electric field lines will terminate on the probe surface. Due to electrostatic induction, the electrostatic charge source on the probe surface will attract charges of different polarities. At this time, charges with the same polarity will gather at the other end of the probe surface. If the signal conditioning circuit is connected to the other end of the probe, the induced charge can be converted into a measurable voltage signal. Therefore, the operating status of the rolling bearing can be monitored in real time through the electrostatic signal.

[0146] When monitoring the electrostatic discharge of the rolling bearing wear area, the induced charge of the charged particles on the sensor probe is recorded as Q(t), and the output induced voltage is recorded as U(t). The default system initial state is 0. The charge Q(t) and the voltage U(t) are transformed by Laplace to Q(s) and U(s), respectively, and the two satisfy the following relationship:

[0147] U(s)=R·s·Q(s) / (R·C·s+1)

[0148] Where R is the equivalent resistance of the detection loop; s is the complex frequency corresponding to time t; and C is the equivalent capacitance of the loop.

[0149] When R·C·s≤1, the electrostatic signal monitoring model can be equivalent to:

[0150]

[0151] The original voltage collected by the rolling bearing acquisition system is proportional to the first-order derivative of the charge on the inductive surface of the electrostatic sensor, which shows that the electrostatic signal can directly reflect the operating status of the rolling bearing.

[0152] The working principle of the rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention is as follows:

[0153] When working, firstly, based on the time domain, frequency domain and time-frequency domain features, the electrostatic charge signal features are combined to form a high-dimensional feature set, and the feature set is reduced in dimension through IDDAE.

[0154] Second, the optimal cluster center is selected by using comprehensive index optimization and entropy method, a health status model is constructed to measure the operating status of the bearing, and the Mahalanobis distance is designed to represent the degree of difference between the health data and the test data.

[0155] Thirdly, the regression model and principle are integrated in the sliding window to repair the false fluctuations and obtain the IMD index that characterizes the degree of bearing degradation. Finally, the relationship between IMD and the threshold is analyzed to realize the bearing performance degradation assessment.

[0156] The experimental results show that: (1) The electrostatic charge signal feature is used to reflect the effective value of the charge in the signal. In electrostatic monitoring, it can serve as a good supplement to conventional time domain and frequency domain indicators; (2) Compared with the traditional dimensionality reduction method PCA, the deep denoising autoencoder optimized by L1 sparse constraint and Dropout mechanism can better identify the nonlinear characteristics and different degradation stages of the bearing; (3) The proposed adaptive density peak clustering model can well measure the health status of the bearing. Compared with traditional DPC and SVDD, the method proposed in this paper can identify the early degradation of the bearing earlier, and has better monotonicity, robustness and trend; (4) Compared with vibration monitoring, electrostatic monitoring is superior in bearing performance degradation assessment and can detect the occurrence of degradation earlier.

[0157] Compared with the related art, the rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention has the following beneficial effects:

[0158] The present invention provides a rolling bearing performance degradation assessment method based on electrostatic monitoring. The method realizes nonlinear dimensionality reduction of a high-dimensional feature set by using a deep denoising autoencoder after L1 sparse constraint, eliminates data redundancy, and introduces a Dropout mechanism to prevent model overfitting. An improved degradation model ADPC is constructed, and the Mahalanobis distance is introduced to calculate the length of a data point. The entropy method is used to find a cutoff distance that satisfies the distribution characteristics of electrostatic data, and an indicator that can simultaneously reflect the local density and the relative distance size is constructed. The point with the largest trend value is the cluster center. Compared with the traditional DPC, the improved model can adaptively select the cluster center and the cutoff distance to reduce the error of empirical selection. Moreover, compared with the original Euclidean distance, the Mahalanobis distance is not affected by the dimension and takes into account the correlation between variables. In order to meet the requirements of online monitoring to repair false fluctuations in the constructed degradation indicator, the set The sliding window is based on the principle of judgment. A polynomial regression model is established in the window. Nonlinear problems can be handled by adding new variables. The electrostatic charge signal feature is used to reflect the effective value of the charge in the signal. In electrostatic monitoring, it can be used as a good supplement to conventional time domain and frequency domain indicators. Compared with the traditional dimensionality reduction method PCA, the deep denoising autoencoder optimized by L1 sparse constraint and Dropout mechanism can better identify the nonlinear characteristics and different degradation stages of the bearing. The proposed adaptive density peak clustering model can measure the health status of the bearing well. Compared with traditional DPC and SVDD, the method proposed in this paper can identify the early degradation of the bearing earlier, and has better monotonicity, robustness and trend. Compared with vibration monitoring, electrostatic monitoring is superior in bearing performance degradation assessment and can detect the occurrence of degradation earlier.

[0159] Second embodiment

[0160] Please refer to Figure 8 Based on the rolling bearing performance degradation assessment method based on electrostatic monitoring provided in the first embodiment of the present application, the second embodiment of the present application proposes another rolling bearing performance degradation assessment method based on electrostatic monitoring. The second embodiment is only a preferred method of the first embodiment, and the implementation of the second embodiment will not affect the independent implementation of the first embodiment.

[0161] Specifically, the second embodiment of the present application provides a rolling bearing performance degradation assessment method based on electrostatic monitoring. The difference is that the rolling bearing performance degradation assessment method based on electrostatic monitoring, the experimental device includes an experimental device body 100, a base 200 is fixedly installed at the bottom of the experimental device body 100, a hydraulic cylinder 300 is fixedly installed around the top of the base 200, the bottom end of the output shaft of the hydraulic cylinder 300 is fixedly connected to a support block 400, and movable wheels 500 are arranged around the bottom of the base 200.

[0162] The working principle of the rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention is as follows:

[0163] During operation, firstly, the output shaft of the hydraulic cylinder 300 is controlled by an external power source to move up or down, thereby adjusting the height between the support block 400 and the moving wheel 500 .

[0164] Compared with the related art, the rolling bearing performance degradation assessment method based on electrostatic monitoring provided by the present invention has the following beneficial effects:

[0165] The present invention provides a rolling bearing performance degradation assessment method based on electrostatic monitoring. The output shaft of the hydraulic cylinder 300 is controlled to move up or down through an external power supply to adjust the height between the support block 400 and the moving wheel 500. The support block 400 supports the ground so that when the moving wheel 500 is suspended in the air, the storage stability of the experimental device main body 100 is maintained. When the moving wheel 500 supports the ground and the support block 400 is suspended in the air, it is convenient for the user to push the experimental device main body 100 for transportation and movement, thereby improving the convenience of transportation of the device.

[0166] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A rolling bearing performance degradation assessment method based on electrostatic monitoring, characterized in that: The following steps are involved: S1: The electrostatic charge signal characteristics are integrated with the conventional time domain, frequency domain and time-frequency domain characteristics to form a characteristic parameter set of the electrostatic monitoring system that characterizes the bearing status; S2: The deep denoising autoencoder optimized with L1 regularization and Dropout mechanism is introduced to perform nonlinear mapping and dimensionality reduction on high-dimensional feature sets. IDDAE can not only improve the ability of feature selection, but also prevent overfitting of deep networks. S3: Adaptive determination of DPC cluster centers is achieved through comprehensive optimization and entropy analysis, thereby avoiding the "chain effect" caused by incorrect allocation of cluster centers. Therefore, a model is established through the improved ADPC algorithm to measure the health status of the bearing, and the Mahalanobis distance MD between the test set and the cluster center is calculated to quantitatively characterize the performance degradation degree of the rolling bearing; S4: Combined with the 3δ principle, a polynomial regression model is established on the sliding window to repair the false fluctuations of the MD index and obtain the improved index IMD that can accurately characterize the degradation process of the rolling bearing. The bearing performance degradation evaluation is realized by analyzing the relationship curve between the health threshold and IMD.

2. A rolling bearing performance degradation assessment method based on electrostatic monitoring according to claim 1, characterized in that: The algorithm has been improved in three ways. The first one is to build a 5-layer deep denoising autoencoder, introduce L1 regularization, use its sparsity to improve the data feature extraction capability, and introduce the Dropout mechanism to prevent overfitting caused by network deepening. The second one is to solve the problem that the clustering effect is too different from the expected value due to the manual selection of cluster centers and cutoff distances in the traditional density peak clustering algorithm, and propose an adaptive density peak clustering algorithm, and improve the original indicator calculation method through the Mahalanobis distance. The third one is to build a polynomial regression model on the sliding window to fix the false fluctuations.

3. The rolling bearing performance degradation assessment method based on electrostatic monitoring according to claim 1 is characterized in that: The method for verifying the effectiveness of the rolling bearing performance degradation evaluation is to conduct a bearing accelerated fatigue life test on an ABLT-1A testing machine. The monitoring platform includes a loading system, a transmission system, a lubrication system, an electrical control system, a computer monitoring system, an experimental head and an experimental head seat; the experimental head includes an experimental device, an electrostatic sensor installation position and a signal acquisition system, and the electrostatic sensor ES1 and the electrostatic sensor ES2 are used to monitor the electrostatic signal of the bearing.

4. The rolling bearing performance degradation assessment method based on electrostatic monitoring according to claim 1 is characterized in that: The L1 regularization is the sum of the absolute values ​​of each element in the weight vector. It guides the model to pay more attention to weights with larger absolute values ​​and continuously adjusts the feature weights so that the weights of some features become 0, thereby achieving a sparse effect.

5. The rolling bearing performance degradation assessment method based on electrostatic monitoring according to claim 1 is characterized in that: The principle of L1 generating a sparse model is as follows: In the formula, J0 is the loss function of the original model, and λ is the regularization coefficient; In the field of machine learning, L1 regularization is often used in sparse models. Therefore, in the loss function of the deep denoising autoencoder, the L1 norm penalty term is added to achieve feature selection by minimizing the sum of the objective function and the L1 norm of the feature. This not only makes the model more robust and generalizable, but also enhances the feature extraction of electrostatic signals. At this time, the loss function of the deep denoising autoencoder after adding L1 regularization is: Adding L1 regularization to the loss function makes it easier to make data sparse, which will also reduce the difficulty of the learning task accordingly. It is also conducive to the automatic selection of features, making the model interpretable and providing good data for the degradation assessment model.

6. The rolling bearing performance degradation assessment method based on electrostatic monitoring according to claim 1 is characterized in that: The Dropout is a technique that can help reduce overfitting of training neural networks. In this study, the deep denoising autoencoder is composed of multiple layers of denoising autoencoders. As the complexity of the neural network increases, overfitting is prone to occur during model training. Therefore, the Dropout mechanism is introduced to solve the overfitting problem to prevent complex co-adaptation on the training data and avoid repeated extraction of the same features. In the training process of deep learning networks, Dropout temporarily and randomly discards hidden layer neurons according to the set probability. In the next training, different neurons will be recombined according to the set probability. After multiple Dropout processes, the original network will become a combination of multiple thinner neural networks. The calculation process of Dropout in the neural network can be expressed as: y (l+1) =σ(W (l)T x (l) +b (l) )1≤l≤L-1 x (l) =dropout(y (l) ,p)2≤l≤L x (l+1) =(1-p)σ(W (l)T x (l) +b (l) ) Among them, l represents the number of neuron layers, L is the total number of neuron layers, and p represents the dropout rate; By training the same set of data through different neural networks and averaging the outputs of each network, the purpose of reducing overfitting can be achieved. In practice, the activation outputs of some hidden neurons can be set to 0, so that these neurons are not included in the forward propagation training process, thereby optimizing the deep denoising autoencoder through dropout.

7. The rolling bearing performance degradation assessment method based on electrostatic monitoring according to claim 1 is characterized in that: The bearing abnormal data accounts for a very small part of the full life cycle data. There is an extreme sample imbalance between normal samples and abnormal samples. The density peak clustering algorithm efficiently allocates sample points and removes outliers by finding density peak points. It is suitable for clustering analysis of large-scale bearing data. ADPC is used to perform unsupervised learning on bearing data to construct the health status of the bearing.

8. The rolling bearing performance degradation assessment method based on electrostatic monitoring according to claim 2 is characterized in that: The fast search clustering algorithm DPC for density peaks is a new density-based clustering algorithm based on the following assumptions: the local density of the cluster center is higher than the local density of other points in the same cluster, and there is a large distance between any points with higher local density. When the normal state of electrostatic data is modeled, the cluster center has a large local density value and relative distance value; In order to find the cluster center that satisfies both conditions, the algorithm introduces the definition of local density and relative distance. The traditional density peak clustering algorithm uses Euclidean distance to reflect the relationship between two points. However, Euclidean distance is easily affected by the feature scale. Therefore, the introduction of Mahalanobis distance to support DPC distance calculation can effectively integrate the characteristics of multiple feature vectors. The calculation formula of Mahalanobis distance between data points is: In the formula, S -1 is the covariance matrix. At this time, the local density calculation process is: When x<0, χ(x)=1, otherwise χ(x)=0; MD ij is the Mahalanobis distance between data points i and j, d c In order to truncate the distance and avoid the same local density of data points, the Gaussian function is used to replace the formula, as shown in the formula; The relative distance calculation process is: In the formula, S represents the set of data, and there are n samples in the set; represents the density ratio ρ i A collection of large points; represents the empty set; Choose a value that can comprehensively consider the local density ρ i and the relative distance δ i The parameter μ i : m i =ρ i d i (1) In the DPC algorithm, the cluster center has a larger local density ρ than other points. i and the relative distance δ i Therefore, by using formula (18), we can increase the μ of the cluster center i μ with other points i The difference in values ​​is used to calculate μ i Then, μ i The values ​​are sorted in descending order, and then μ is calculated i The downward trend of the value is calculated as follows: In the formula, μ i is the current μ value; μ i-1 and μ i+1 Represent the μ values ​​of the previous moment and the next moment respectively. After obtaining the downward trend, take the point with the largest downward trend and the previous points as the clustering center; Traditional DPC algorithms usually rely on experience or experimental verification to obtain d c The value makes the clustering result have a lot of uncertainty. In order to improve the clustering effect, d can be determined based on the difference between the comprehensive measurement data. c This paper uses the entropy method to constrain the value of d c In information theory, the amount of information contained in an event is measured by information entropy. In comprehensive evaluation, if the numerical difference of the same indicator is small, that is, the probability distribution is more uniform, it means that the event "a certain value belongs to a certain evaluated object" has greater uncertainty and information entropy. Therefore, a lower weight is given to the indicator. On the contrary, if the numerical difference of the same indicator is large, the weight of the indicator is higher. The calculation formula of information entropy is: Where n represents the number of samples in the data set S; p i is the ρ of point i i ×δ i The proportion of all points, p i The values ​​are as follows: The minimum value of d c The value is used as the final clustering d c value, instead of empirical method or experimental verification, so as to achieve adaptive selection of parameters.

9. The rolling bearing performance degradation assessment method based on electrostatic monitoring according to claim 1, characterized in that: During the operation of the bearing, its performance is in a continuous state of decline. Due to some external factors, the decline characteristics will show abnormal fluctuations, which will seriously affect the description of the bearing degradation state by the degradation index. Therefore, if there are false fluctuations, they need to be repaired before being used as an index for subsequent degradation assessment; In this paper, a sliding window is constructed based on the health index. The window step size is set to 1, and the window length is 50. The fluctuation threshold θ is set according to the 3σ principle and polynomial regression of the data within the sliding window. Its expression is: Among them, the values ​​of ω0, ω1 and ω2 are determined by the least squares method; x i For the time series of health indicators, MD i is the distance between the data to be tested and the cluster center; Determine the health indicator IMD after repairing the false fluctuations according to the sliding window fluctuation threshold θ i , whose expression is: Among them, min{MD(i)} and max{MD(i)} represent the upper and lower limits of MD(i) respectively; In the formula, μ represents the mean value of MD(i); σ represents the standard deviation of MD(i); According to the above formula, it is judged whether there are false fluctuations in the bearing health index and repaired. When MD(i) ≤ min{MD(i)}, MD(i) moves to the average value to obtain a new predicted value IMD; when min{MD(i)} < MD(i) < max{MD(i)}, MD(i) is replaced with the value predicted by polynomial regression; when MD(i) ≥ max{MD(i)}, MD(i) moves to max{MD(i)}, and then a new predicted value is calculated. By establishing a regression model on the sliding window, the false fluctuations existing in the health index can be effectively repaired, so as to obtain a health index with better performance in characterizing the degradation trend.

10. The rolling bearing performance degradation assessment method based on electrostatic monitoring according to claim 4, characterized in that: The experimental device includes an experimental device main body 100. A base 200 is fixedly installed at the bottom of the experimental device main body 100. Hydraulic cylinders 300 are fixedly installed around the top of the base 200. The bottom end of the output shaft of the hydraulic cylinder 300 is fixedly connected to a support block 400. Movable wheels 500 are arranged around the bottom of the base 200.

Images