Rolling bearing degradation point detection method based on sparse time-frequency self-coding confrontation

Abstract

A rolling bearing degradation point detection method based on sparse time-frequency self-coding adversarial belongs to the field of rolling bearings, and comprises the following steps: an HI construction stage: carrying out feature learning and data enhancement on vibration signals collected by a rolling bearing in a healthy state by adopting a sparse time-frequency self-coding adversarial network model; an HI curve capable of sensitively reflecting performance degradation of the rolling bearing is constructed through depth feature extraction; an IDP determination stage: performing local feature analysis on an HI sequence output by the sparse time-frequency self-encoding adversarial network model by adopting a dynamic sliding window technology, and constructing a dynamic threshold in combination with a linear residual analysis method so as to determine a potential initial degradation interval; and then introducing a curvature optimization method into the initial degradation interval, and finally accurately positioning an initial degradation point with significant statistical significance by quantifying the change rate of geometric characteristics of an HI curve. According to the method, the accuracy and reliability of initial degradation point detection are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of rolling bearings, and particularly relates to a rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation. BACKGROUND

[0002] As a key basic component in mechanical systems, rolling bearings play an irreplaceable and important role in many fields such as mechanical manufacturing, automobile industry, aerospace, etc. Due to the core position of rolling bearings in mechanical equipment, its failure often leads to non-planned shutdown of equipment, performance degradation, and even serious damage. Accurate prediction of the remaining useful life (RUL) of rolling bearings is of great significance for implementing predictive maintenance, reducing accident risk, and ensuring safe operation of equipment. However, the degradation process of rolling bearings has obvious stage characteristics, which can usually be divided into a healthy stage and a degradation stage. Traditional RUL prediction methods usually use full life cycle data for model training, but the health stage data contains limited fault information, which will affect the accuracy of the prediction model. Accurate identification of the initial degradation point (IDP) to divide the healthy stage and the degradation stage is a key prerequisite for improving the accuracy of RUL prediction.

[0003] In rolling bearing RUL prediction research, the construction of unsupervised health index (HI) is a key step to represent the degradation process of equipment. Existing research mainly adopts deep learning-based methods to construct HI. For example, Wang et al. (Wang Q, Xu K, Kong X, Huai T. A linear mapping method for predicting accurately the RUL of rolling bearing. Measurement 2021; 176: 109127. http: / / dx.doi.org / 10.1016 / j.measurement.2021.109127.) proposed a method based on reliability metrics, which extracts deep features through multiple deep autoencoder models and constructs a linear health index. Guo et al. (Guo L, Li N, Jia F, Lei Y, Lin J. A recurrent neural network based health indicator for remaining useful life prediction of bearings. Neurocomputing 2017; 240: 98-109. http: / / dx.doi.org / 10.1016 / j.neucom.2017.02.045.) adopted a recurrent neural network (RNN) to process the features extracted from the original signal to construct HI. However, these methods all adopt supervised learning to construct HI, which has obvious limitations in actual industrial applications: on the one hand, there is much more equipment health status data than degradation data; on the other hand, the monitoring and labeling cost of the whole life cycle data is high. To overcome this limitation, some scholars turn to unsupervised learning methods. The research of Guo et al. (Guo L, Yu Y, Duan A, Gao H, Zhang J. An unsupervised feature learning based health indicator construction method for performance assessment of machines. Mech Syst Signal Process 2022; 167: 108573.) shows that in some application scenarios, the output of the autoencoder can be directly used as an effective HI.Li et al. (Li X, Zhang W, Ma H, Luo Z, Li X. Data alignments in machinery remaining useful life prediction using deep adversarial neural networks. Knowl-Based Syst 2020;197:105843. http: / / dx.doi.org / 10.1016 / j.knosys.2020.105843.) developed an end-to-end prediction framework based on generative adversarial networks (GAN). But these methods failed to fully consider the noise interference in actual working conditions (such as sensor loosening), changes in operating conditions, and other component vibrations, which can significantly affect the monotonicity and robustness of the HI. Therefore, to address the noise problem, existing research has mostly used linear fitting and other methods for smoothing. Zhang et al. (Zhang N, Wu L, Wang Z, Guan Y. Bearing remaining useful life prediction based on naive Bayes and Weibull distributions. Entropy 2018;20:944. http: / / dx.doi.org / 10.3390 / e20120944.) used an improved Weibull distribution to fit the characteristics of rolling bearings at different degradation stages to eliminate HI noise. However, this type of method tends to eliminate the important feature of the “running-in stage” while reducing noise, and this feature is crucial for determining the degradation stage and degree (El-Thalji I, Jantunen E. A descriptive model of wear evolution in rolling bearings. Eng Fail Anal 2014;45:204-24. http: / / dx.doi.org / 10.1016 / j.engfailanal.2014.06.004).

[0004] In existing research, the initial IPT is usually determined based on a degradation threshold obtained from relevant standards or prior knowledge. A typical example is Khazaee et al. (Khazaee M, Banakar A, Ghobadian B, Mirsalim MA, Minaei S. Remaining useful life (RUL) prediction of internal combustion enginetiming belt based on vibration signals and artificial neural network. NeuralComput Appl 2021;33:7785–801. http: / / dx.doi.org / 10.1007 / s00521-020-05520-3.) who used the root mean square (RMS) of the original vibration signal as the HI and set the degradation threshold based on industry standards and expert knowledge. However, this method has obvious limitations: on the one hand, some industries lack comprehensive standards and specifications; on the other hand, over-reliance on highly subjective expert knowledge significantly increases the difficulty of threshold determination. To overcome these shortcomings, researchers generally adopt statistical principles... The criterion method (Chen Z, Xia T, Li Y, Pan E. A hybrid prognostic method based on gated recurrent unit network and an adaptive Wiener process model considering measurement errors. Mech Syst Signal Process 2021;158:107785. http: / / dx.doi.org / 10.1016 / j.ymssp.2021.107785.) constructs a dynamic threshold interval by calculating the mean and standard deviation of the observed data. Typical applications include 2 Guidelines and 3 Criteria. For example, Li et al. (LI N, LEIY, LIN J, et al. An improved exponential model for predicting remaining useful life of rolling element bearings[J]. IEEE Transactions on Industrial Electronics, 2015, 62(12): 7762-7773.) used vibration signal kurtosis as a detection index and established a 3 The threshold range is defined, and when multiple consecutive kurtosis values ​​exceed this range, it is determined as the initial degradation point. Li et al. (Li W, Shang Z, Gao M, Qian S, Feng Z. Remaining useful life prediction based on transfer multi-stage shrinkage attention temporal convolutional network under variable working conditions. Reliab Eng Syst Saf 2022;226:108722. http: / / dx.doi.org / 10.1016 / j.ress.2022.108722.) developed a multi-scale convolutional neural network (CNN) model and adopted 2 The criteria optimize IPT identification accuracy to improve remaining useful life (RUL) prediction performance. Cheng et al. (Cheng H, Kong X, Wang Q, Ma H, Yang S. The two-stage RUL prediction across operation conditions using deep transfer learning and insufficient degradation data. Reliab Eng Syst Saf 2022;225:108581. http: / / dx.doi.org / 10.1016 / j.ress. 2022.108581.) proposed a two-stage RUL prediction method, and Chen et al. (Chen J, Jing H, Chang Y, Liu Q. Gated recurrent unit based recurrent neural network for remaining useful life prediction of nonlinear deterioration process. ReliabEng Syst Saf 2019;185:372–82. http: / / dx.doi.org / 10.1016 / j.ress.2019.01. 006.) proposed a hybrid prediction model based on gated recurrent units (GRU) and Wiener processes. Both methods employ 3 The criterion for IPT detection is to trigger a degradation alarm when the observed data exceeds three standard deviations of the mean, and determine the degradation initiation time after the continuous triggering condition is met. For variable speed operation, Li et al. (Li N, Xu P, Lei Y, Cai X, Kong D. A self-data-driven method for remaining useful life prediction of wind turbines considering continuously varying speeds. Mech Syst Signal Process 2022;165:108315. http: / / dx.doi.org / 10.1016 / j.ymssp.2021. 108315) innovatively incorporated 3... A dynamic early warning threshold system was constructed by combining the criteria with a continuous triggering algorithm. However, these existing methods for detecting the initial degradation point (IDP) of rolling bearings suffer from false alarms and missed detections due to vibration signal noise interference. Summary of the Invention

[0005] To address the false detection problem in existing rolling bearing initial degradation point (IDP) detection methods, this invention provides a rolling bearing degradation point detection method based on sparse time-frequency autoencoder countermeasures.

[0006] The technical solution adopted by this invention to solve the technical problem is as follows:

[0007] This invention provides a method for detecting degradation points in rolling bearings based on sparse time-frequency autoencoders, comprising the following steps:

[0008] HI construction phase: A sparse time-frequency autoencoder adversarial network model is used to perform feature learning and data augmentation on the vibration signals collected under the healthy state of the rolling bearing. Through deep feature extraction, an HI curve that can sensitively reflect the performance degradation of the rolling bearing is constructed.

[0009] In the IDP determination stage: First, the HI sequence output by the sparse time-frequency autoencoder adversarial network model is analyzed for local features using dynamic sliding window technology. Then, a dynamic threshold is constructed using linear residual analysis to determine the potential initial degradation interval. Subsequently, within this initial degradation interval, the curvature optimization method is introduced to quantify the rate of change of the geometric features of the HI curve, and finally, the initial degradation point with statistical significance is accurately located.

[0010] Furthermore, the sparse time-frequency autoencoder adversarial network model includes a generator and a discriminator. The generator consists of a parallel time-frequency domain autoencoder, a channel attention mechanism, a sparse regularization network with sparse long short-term memory layers, and a decoder. The parallel time-frequency domain autoencoder includes a time-domain encoder and a frequency-domain encoder, which have the same structure, consisting of two convolutional layers and two pooling layers. A batch normalization function and a ReLU activation function are added to the convolutional layers of the encoder. The decoder is responsible for outputting the reconstructed vibration signal, and its structure includes two deconvolutional layers and two upsampling layers, corresponding to the encoder's structure. The original vibration signal sequence is input into the generator to train it to generate spurious vibration signals and reconstruct the vibration signal, while simultaneously optimizing the sparse coding in the space using sparse regularization.

[0011] The discriminator consists of multiple convolutional layers and a fully connected layer. The multiple convolutional layers use the same activation function, and the fully connected layer is used to help the generator learn the distribution of the input data. The original vibration signal and the reconstructed vibration signal are simultaneously passed to the discriminator, enabling it to distinguish between the generated signal and the real signal, thereby driving the generator to generate a high-quality reconstructed vibration signal that is closer to the original input.

[0012] Furthermore, the original vibration signal is simultaneously input into a parallel structure of a time-domain encoder and a frequency-domain encoder. Before the frequency-domain encoder, the original vibration signal needs to be subjected to a fast Fourier transform. The features extracted by the parallel time-frequency domain autoencoder are input into the channel attention mechanism to enhance the feature representation. The time-domain and frequency-domain features are concatenated and then passed to a sparse regularization network with a sparse long short-term memory layer for sparse regularization processing. This layer is applied to the latent state in the generator to learn a complete dictionary and generate sparse code.

[0013] Furthermore, during adversarial training, the training loss L of the sparse time-frequency autoencoder adversarial network model is composed of the time-frequency reconstruction loss. Combating losses and sparsity loss composition:

[0014] (1)

[0015] The time-frequency reconstruction loss The mathematical expression is:

[0016] (4)

[0017] in, All are weighting coefficients. For temporal reconstruction loss, For frequency domain reconstruction loss;

[0018] The resistance loss The mathematical expression is:

[0019] (6)

[0020] in, For generator loss, For discriminator loss;

[0021] The sparse loss The mathematical expression is:

[0022] (9)

[0023] in, A sparsely regularized dictionary. Sparse code generated by sparse regularization The sparse code that serves as input to the generator. It is the Frobenius norm. for Norm, with the superscript T indicating transpose.

[0024] Furthermore, the temporal reconstruction loss and frequency domain reconstruction loss The mathematical expression is:

[0025] (2)

[0026] (3)

[0027] in, For batch size, The input is the original vibration signal. To reconstruct the vibration signal, This is the frequency domain signal after Fast Fourier Transform.

[0028] Furthermore, the time-frequency domain weights are automatically adjusted based on the loss magnitude to ensure dynamic updates during each training round and maintain a balance in the loss magnitude. The mathematical expression for this is:

[0029] (5).

[0030] Furthermore, the discriminator loss and generator loss The mathematical expression is:

[0031] (7)

[0032] (8)

[0033] in, For the discriminator input, To determine the reconstructed input signal, These are the weighting coefficients. A signal created by randomly combining the original vibration signal and the reconstructed vibration signal. and As a Wasserstein distance label The input is the original vibration signal. To reconstruct the vibration signal, This refers to the batch size.

[0034] Furthermore, during adversarial training, the distribution of the sparse time-frequency autoencoder adversarial network model's learning data under the rolling bearing's health condition is used. When the real-time monitored vibration signal is input again, the discriminator score and the residual loss generated during the iteration process are used as indicators of the degree of degradation. The mathematical expression is as follows:

[0035] (10)

[0036] in, These are the weighting coefficients; The residual loss is expressed mathematically as follows:

[0037] (11)

[0038] in, The input is the original vibration signal. For the discriminator input, To reconstruct the vibration signal.

[0039] Furthermore, in the IDP determination stage, the initial degradation interval is determined by using residual analysis combined with a sliding window to construct a dynamic threshold, including the following steps:

[0040] S3.1.1: Calculate the residual sequence within the window;

[0041] Window size is determined by the grid search method; residual sequence The calculation formula is:

[0042] (12)

[0043] in, For unsupervised health indices, The fitted value is the value within the window, and t is the value at time t;

[0044] S3.1.2: Construct a dynamic threshold;

[0045] As the window slides across the time series, the window for different time series is calculated. mean of residuals within and standard deviation Construct dynamic threshold Its mathematical expression is:

[0046] (13)

[0047] (14)

[0048] (15)

[0049] in, The window start time, The end time of the window; This is the sensitivity coefficient;

[0050] S3.1.3: Determine if the number of consecutive points exceeding the threshold N within the window is greater than or equal to 3; if not, return to step S3.1.1; if yes, accumulate the degradation interval; within the window In the middle, a threshold number is set. At that time, the first time point exceeding the threshold will be... The starting point of the degradation range is taken as the first time point that falls below the threshold. This allows for the identification of the first degradation region. The first five degradation regions were identified sequentially using this method. As a candidate set for the initial degenerate interval;

[0051] S3.1.4: HI Consistency Test;

[0052] The HI monotonicity test was performed sequentially on the five identified degradation intervals. The initial degradation interval was selected by comparing the HI fitting slope of the previous window with that of the identified degradation interval. The mathematical expression for the HI consistency test is:

[0053] (16)

[0054] in, The current window slope, These represent the slope and standard deviation of the historical window, respectively.

[0055] Furthermore, in the IDP determination stage, an exponential function is used to fit the HI curve within the degradation interval, and the IDP is determined by finding the curvature extrema of the fitted HI curve.

[0056] The beneficial effects of this invention are:

[0057] This invention provides a rolling bearing degradation point detection method based on sparse time-frequency autoencoder adversarial mechanism. It innovatively integrates a triple noise reduction mechanism of time-frequency domain sparse coding, attention mechanism, and adversarial training, achieving effective suppression of signal noise and accurate preservation of fault features through collaborative optimization. Based on this, a robust rolling bearing IDP determination criterion is established using the reconstructed HI signal, combined with dynamic residual analysis and the curvature optimization principle. Experimental analysis is conducted on the PHM 2012 public dataset. The experimental results show that the proposed method significantly improves the accuracy and reliability of initial degradation point detection.

[0058] In addition, this invention introduces sparse regularization into the HI construction framework to construct a more reliable HI by suppressing noise interference, while retaining key degradation features. This effectively removes severe noise from the original vibration signal without losing important degradation characteristics. Attached Figure Description

[0059] Figure 1 The flowchart of a rolling bearing degradation point detection method based on sparse time-frequency autoencoder countermeasures provided by the present invention is shown.

[0060] Figure 2This is a diagram of the architecture of the Sparse Time-Frequency Autoencoder Adversarial Network (STFA-GAN) model.

[0061] Figure 3 Procedure for determining the final initial degradation range.

[0062] Figure 4 This is a comparison curve of the HI fitting slope of the previous window of the degradation interval identified during the HI consistency test.

[0063] Figure 5 The process of locating the initial degradation point. In the figure, (a) is the exponential fitting curve; (b) is the curvature. Curve showing how the time t changes.

[0064] Figure 6 The time-domain waveforms of the rolling bearings are shown in the figure. (a) Rolling bearing PB11; (b) Rolling bearing PB12.

[0065] Figure 7 The HI constructed for Task 1. In the figure, (a) RMS; (b) P-Entropy; (c) ISOMAP; (d) KPCA; (e) CEEMDAN; (f) VAE; (g) SSAE; (h) the present invention.

[0066] Figure 8 The HI constructed for Task 2. In the figure, (a) RMS; (b) P-Entropy; (c) ISOMAP; (d) KPCA; (e) CEEMDAN; (f) VAE; (g) SSAE; (h) the present invention.

[0067] Figure 9 The average performance metrics for HI were constructed using different methods in Task 1 and Task 2. In the figure, (a) Mon; (b) Tred; (c) Rob; (d) HS.

[0068] Figure 10 The signals and sequences of rolling bearing PB14 are shown in the comparative analysis for determining the initial degradation point. In the figure, (a) is the vibration signal of PB14; (b) is the HI spliced ​​sequence.

[0069] Figure 11 This describes the testing process for rolling bearing PB14.

[0070] Figure 12The IDP detection results were determined for different thresholding methods. In the figure, (a) bearing PB11; (b) bearing PB12; (c) bearing PB16; (d) bearing PB23; Figures (a.1)-(d.1) show the detection effect of different thresholding methods on the original vibration signal; Figures (a.2)-(d.2) present the detection results of the kurtosis2σ method on its degradation index kurtosis; Figures (a.3)-(d.3) compare the detection performance of the method proposed in this invention and the RMS2σ method on the RMS degradation index; Figures (a.4)-(d.4) show the detection effect of the RMS3σ method on the RMS degradation index.

[0071] Figure 13 IDP detection results were determined for different model classification methods. In the figure, (a) bearing PB11; (b) bearing PB12; (c) bearing PB13; (d) bearing PB14; (e) bearing PB15; (f) bearing PB16; (g) bearing PB17. Detailed Implementation

[0072] The present invention will be further described in detail below with reference to the accompanying drawings.

[0073] Because the original vibration signal of rolling bearings often contains strong background noise and interference components, traditional IDP detection methods are prone to misjudgment or missed detection under noise interference. To solve this critical problem, this invention provides a rolling bearing degradation point detection method based on sparse time-frequency autoencoder adversarial mechanism, namely, an asymptotic two-stage detection method that combines a HI construction model for reconstructing the vibration signal with dynamic threshold analysis. Its overall framework is as follows: Figure 1 As shown, it mainly includes two stages:

[0074] The first stage is the HI construction stage, which uses a sparse time-frequency autoencoder adversarial network (STFA-GAN) model to perform feature learning and data augmentation on the vibration signals collected under the healthy state of the rolling bearing. Through deep feature extraction, an HI curve that can sensitively reflect the performance degradation of the rolling bearing is constructed.

[0075] The second stage is the IDP determination stage. First, the HI sequence output by the Sparse Time-Frequency Autoencoder Adversarial Network (STFA-GAN) model is analyzed for local features using a dynamic sliding window technique. A dynamic threshold is constructed using linear residual analysis to determine potential initial degradation intervals. Then, within this initial degradation interval, an optimal curvature method is introduced to quantify the rate of change of the geometric features of the HI curve, ultimately pinpointing statistically significant initial degradation points. This invention, by integrating deep feature learning and dynamic signal processing techniques, achieves effective capture and accurate localization of early, subtle fault characteristics in rolling bearings.

[0076] likeFigure 1 As shown, the present invention provides a method for detecting degradation points in rolling bearings based on sparse time-frequency autoencoders, and its specific implementation process is as follows:

[0077] Step 1: Construct a Sparse Time-Frequency Autoencoder-GAN model;

[0078] Traditional signal denoising methods rely on preset thresholds or frequency band divisions, which may result in the loss of useful information. To achieve low-noise reconstruction of vibration signals, this invention proposes a sparse time-frequency autoencoder adversarial network (STFA-GAN) model, which combines a parallel time-frequency domain autoencoder with a generative adversarial network (GAN) model. The STFA-GAN network model aims to learn the data distribution of the health status of each rolling bearing in the training set.

[0079] like Figure 2 As shown, the Sparse Time-Frequency Autoencoder-GAN (STFA-GAN) model constructed in this invention mainly includes: a generator. and discriminator The generator By parallel time-frequency domain autoencoder (time-domain encoder) and frequency domain encoder The channel attention mechanism SENet, the sparse regularized network SRNet with sparse long short-term memory (SLSTM) layers, and the decoder Composition; discriminator It consists of convolutional layers and fully connected layers.

[0080] In the time-frequency domain, key features are particularly important. To extract these key features, the original vibration signal sequence... Input into generator In the middle, training generator This process generates spurious vibration signals and reconstructs the original vibration signals, while simultaneously optimizing the sparse coding in the space using sparse regularization. Both the original and reconstructed vibration signals are then simultaneously passed to the discriminator. This enables it to distinguish between generated signals and real signals, thereby driving the generator. Generate a high-quality reconstructed vibration signal that is closer to the original input.

[0081] Step 2: Construct an unsupervised health index (HI);

[0082] In the generator G, the model first adds the batch normalization function and the ReLU activation function to the convolutional layer of the encoder. (Original vibration signal) Simultaneously input to the time-domain encoder and frequency domain encoder In a parallel structure, the time-domain encoder... and frequency domain encoder Both have the same structural composition, consisting of two convolutional layers and two pooling layers, in the frequency domain encoder. Previously, a Fast Fourier Transform (FFT) was performed on the original vibration signal. Features extracted by a parallel time-frequency domain autoencoder were then input into the channel attention mechanism SENet to enhance the feature representation. The time-domain and frequency-domain features were concatenated and then passed to a sparse regularization network SRNet containing a sparse long short-term memory (SLSTM) layer for sparse regularization. This layer is applied to the latent states in the generator G to learn a complete dictionary. This generates sparse code. Decoder The generator G is responsible for outputting the reconstructed vibration signal. Its structure includes two deconvolutional layers and two upsampling layers, corresponding to the encoder's structure, thus forming the overall architecture of the generator G. The discriminator D contains five convolutional layers, all using the same activation function. Furthermore, the discriminator... It also includes a fully connected layer to help the generator G learn the distribution of the input data, thereby improving the quality of the reconstructed vibration signal.

[0083] (1) Total training loss;

[0084] During adversarial training, the training loss of the STFA-GAN network model consists of time-frequency reconstruction loss. Combating losses and sparsity loss Its composition, and its mathematical expression are as follows:

[0085] (1)

[0086] (2) Time-frequency reconstruction loss ;

[0087] To measure the time-frequency difference between the reconstructed vibration signal and the original vibration signal, the generator G is trained using batch size. Calculate the reconstruction loss:

[0088] (2)

[0089] (3)

[0090] in, For temporal reconstruction loss, For frequency domain reconstruction loss, The input is the original vibration signal. To reconstruct the vibration signal, For the frequency domain signal after Fast Fourier Transform, the time-frequency reconstruction loss is... The sum of the time-frequency domain losses, i.e.:

[0091] (4)

[0092] (5)

[0093] in, , All are weight coefficients, and the time-frequency domain weights can be automatically adjusted using the loss magnitude, as shown in Equation (5). The superscript (t) indicates time t, so as to ensure dynamic updates during each training round and ensure the balance of the loss magnitude.

[0094] (3) Combating losses ;

[0095] For combating losses The weights of the generator G and discriminator D are adjusted using Wasserstein distance with gradient penalty as the adversarial loss. (Adversarial loss) The definition is as follows:

[0096] (6)

[0097] (7)

[0098] (8)

[0099] in, For generator loss, For discriminator loss, For the discriminator input, These are the weighting coefficients. A signal created by randomly combining the original vibration signal and the reconstructed vibration signal. and As a Wasserstein distance label.

[0100] (4) Sparsity loss ;

[0101] To adjust the dictionary for sparse regularization sparse loss The definition is as follows:

[0102] (9)

[0103] in, Sparse code generated by sparse regularization The sparse code that serves as input to the generator. It is the Frobenius norm. for Norm, with the superscript T indicating transpose.

[0104] During adversarial training, the STFA-GAN network model learns the distribution of data under the healthy state of the rolling bearing. When the real-time monitored vibration signal is input into the STFA-GAN network model again, the discriminator D score and the residual loss generated during the iteration process are used as indicators of the degree of degradation. The mathematical expression is as follows:

[0105] (10)

[0106] in, These are the weighting coefficients; The residual loss is expressed mathematically as follows:

[0107] (11)

[0108] The STFA-GAN network model serves as a means to quantify the deviation between the distribution of real-time monitored vibration signals and the baseline health status data. During normal equipment operation, the HI output of the STFA-GAN network model exhibits a stable and consistent characteristic, indicating a high degree of consistency between the real-time monitored vibration signal distribution and the baseline health status data distribution. When equipment begins to degrade, the HI output of the STFA-GAN network model shows significant increased volatility and trend changes. This dynamic characteristic reflects the evolution of the real-time monitored vibration signal distribution gradually deviating from the baseline health status. Through this quantitative analysis based on data distribution differences, the subtle transition of equipment from a healthy state to a degraded state can be effectively captured, providing a reliable criterion for early fault detection.

[0109] Step 3: Dynamic threshold degradation point detection based on unsupervised health index;

[0110] To reduce the impact of noise on IDP detection, the unsupervised health index (HI) obtained based on the STFA-GAN network model is used to construct a dynamic threshold by residual analysis combined with a sliding window to determine the initial degradation interval. Within the determined initial degradation interval, the final degradation point is determined by curve fitting combined with the curvature optimization method.

[0111] S3.1: Determination of the initial degradation interval based on residual analysis;

[0112] Traditional IDP detection methods use a fixed threshold for over-threshold detection, ignoring individual differences under different operating conditions. To address the impact of individual differences under different operating conditions on detection, this invention employs residual analysis combined with a sliding window to construct a dynamic threshold to determine the initial degradation interval.

[0113] like Figure 3 As shown, the specific implementation process is as follows:

[0114] S3.1.1: Calculate the residual sequence within the window;

[0115] Window size is determined by the grid search method; residual sequence The specific calculation formula is as follows:

[0116] (12)

[0117] in, For unsupervised health indices, is the fitted value within the window, and t is the value at time t.

[0118] S3.1.2: Construct a dynamic threshold;

[0119] As the window slides across the time series, the window for different time series is calculated. mean of residuals within and standard deviation Construct dynamic threshold Its mathematical expression is as follows:

[0120] (13)

[0121] (14)

[0122] (15)

[0123] in, The window start time, The end time of the window; The sensitivity coefficient is set as follows: since the HI generated by the STFA-GAN network model has undergone multiple noise reduction steps, the sensitivity coefficient is set as follows: .

[0124] S3.1.3: Determine if the number of consecutive points exceeding the threshold N within the window is greater than or equal to 3; if not, return to step S3.1.1; if yes, then accumulate the degradation interval ( ). In the window In the middle, a threshold number is set. At that time, the first time point exceeding the threshold will be... The starting point of the degradation range is taken as the first time point that falls below the threshold. This allows for the identification of the first degradation region. Following this method, the first five degradation regions were identified sequentially. As a candidate set for the initial degenerate interval.

[0125] S3.1.4: HI Consistency Test;

[0126] In practical applications, residuals may exceed the limit within the initial degradation interval, but the HI (hysteresis index) may not change significantly. This could be due to model fitting bias rather than the actual degradation process. Therefore, the HI monotonicity test is performed sequentially on the five identified degradation intervals. By comparing the HI fitting slope of the previous window with the identified degradation interval, the initial degradation interval is selected. Specifically, as follows... Figure 4 As shown, This represents the fitting of historical window vibration signals. This indicates the fitting of the vibration signal in the current window. The slope of the historical window. The current window slope, Fit the intercept to the historical window. Fit the intercept for the current window.

[0127] The mathematical expression for the HI consistency test is as follows:

[0128] (16)

[0129] in, The current window slope, These represent the slope and standard deviation of the historical window, respectively.

[0130] S3.2: Initial degradation point location;

[0131] Considering that curvature analysis can effectively detect abrupt changes in curves, an exponential function is used to fit the HI curve within the degradation interval, and the IDP is determined by finding the curvature extrema of the fitted HI curve.

[0132] Since the HI (hysteresis index) of rolling bearings exhibits an early, stable, and gradually increasing trend as they transition from a healthy to a degraded state, the HI sequence within the degradation interval containing the IDP (independent degradation point) can be fitted with an H-number curve to further eliminate random noise interference. To more accurately capture the exponential degradation trend, the degradation interval containing the IDP is appropriately expanded to obtain a more representative exponential HI fitting curve.

[0133] like Figure 5 As shown in (a), the degradation interval where the IDP is located is obtained. The HI exponent fitting curve was obtained. nm-1 curvature circles were constructed on the HI exponent fitting curve. Among them, the two ends of the HI index fitting curve and Let O be a fixed point, and let O be a point on the fitted curve of the HI exponent. to Slip point, curvature It changes as the sliding point O moves. Figure 5 (b) demonstrates curvature A line showing changes over time.

[0134] To demonstrate the effectiveness of the proposed method, data from the publicly available dataset IEEE PHM 2012 were used to compare and analyze the HI construction performance of the STFA-GAN network model with other models under different operating conditions, as well as the IDP detection performance of different threshold and model classification methods. The experimental platform was an HP workstation equipped with an Intel Core i9 processor, a 14900 processor, and an NVIDIA GeForce RTX 4090 graphics card.

[0135] The IEEE PHM 2012 dataset used in the experiment (Nectoux P, Gouriveau R, Medjaher K, Ramasso E, Chebel-Morello B, Zerhouni N, Varnier C. PRONOSTIA: An experimental platform for bearings accelerated degradation tests. In: IEEE international conference on prognostics and health management. pHM'12, IEEE Catalog Number: CPF12PHM-CDR; 2012, p. 1–8.) was collected from the PRONOSTIA platform. (Type: NSK6804DD) The motor power was 250W, and the maximum speed was 2830rpm, ensuring the second shaft speed was 2000rpm. Degradation datasets were collected by measuring the vibration signals of the rolling bearings in the horizontal and vertical directions every 10s using two miniature accelerometers at a sampling frequency of 25.6 kHz. Each sample contained 2560 points. Because of the influence of radial load, the horizontal vibration signal can better reflect the condition of the rolling bearing, so the horizontal vibration signal is used for the test.

[0136] (1) Comparative analysis of HI construction results;

[0137] To verify the HI construction capability of the proposed method under various operating conditions, two different research tasks were proposed as shown in Table 1. For Task 1, six rolling bearings (PB12, PB13, PB14, PB15, PB16, PB17) were used as the training set, and the rolling bearing (PB11) was tested under the first operating condition: a load of 4000N and a speed of 1800rpm. For Task 2, the rolling bearing training set was consistent with that of Task 1, and the rolling bearing (PB21) was tested under the second operating condition: a load of 4200N and a speed of 1650rpm. The full-cycle time-domain waveform of the tested rolling bearing is shown in Table 1. Figure 6 As shown.

[0138] Table 1 Research Tasks for HI Construction

[0139]

[0140] To quantitatively evaluate the performance of the HI construction method, four key evaluation metrics were selected in the experiment: monotonicity (Mon), trendability (Tred), robustness (Rob), and hybrid scale (HS). Higher values ​​for these metrics indicate superior performance of the HI construction method. The monotonicity metric is specifically used to assess whether HI accurately reflects the unidirectional trend of equipment degradation, based on the irreversible nature of mechanical equipment degradation in industrial environments. The mathematical expression for the monotonicity metric is as follows:

[0141] (17)

[0142] Where K is the total number of data points in the health indicator sequence, and y is the input sequence. ( ) is the index function; as well as This represents the trend value of HI at time points calculated using a smoothing method according to the traditional method in HI performance evaluation (Lei Y, Li N, Guo L, Li N, Yan T, Lin J. Machinery health prognostics: A systematic review from data acquisition to RUL prediction. Mech Syst Signal Process 2018;104:799–834.).

[0143] A trend indicator was chosen to assess whether the trend of HI was related to operating time, as rolling bearings are more likely to degrade gradually with increasing operating time. The mathematical expression for the trend indicator is as follows:

[0144] (18)

[0145] in, This represents the operation time corresponding to the k-th data point.

[0146] Considering the unavoidable nature of measurement noise, the randomness of the degradation process, and variations in operating conditions, a robustness index is chosen to evaluate whether HI is robust to disturbances and exhibits a smooth degradation trend. The mathematical expression for the robustness index is as follows:

[0147] (19)

[0148] To comprehensively evaluate the performance of HI, a mixed-scale index is proposed. The mathematical expression for the mixed-scale index is as follows:

[0149] (20)

[0150] To verify the effectiveness of the proposed method, this invention selected several mainstream unsupervised health index (HI) construction methods for comparative experiments, including root mean square (RMS), permutation entropy (P-Entropy), isometric mapping (ISOMAP), kernel principal component analysis (KPCA), fully adaptive ensemble empirical mode decomposition (CEEMDAN), variational autoencoder (VAE), and stacked sparse autoencoder (SSAE). These methods cover different technical routes from traditional signal processing and manifold learning to deep learning, and are widely representative. To ensure the fairness of the experiments, all comparative methods adopted a unified experimental setup: the network structure of VAE was consistent with the STFA-GAN network model structure in this invention; SSAE was implemented as a two-layer stacked linear network, with specific structures of 1280-400-1280 and 400-1-400, respectively. Through this systematic comparative experimental design, the performance advantages of the proposed method in the HI construction task can be comprehensively evaluated.

[0151] Original vibration signal such as Figure 6 As shown, the degradation score obtained by inputting into the STFA-GAN network model is as shown in formula (10). The root mean square of the degradation score is used to quantify the degree of degradation to obtain the final HI. The results are as follows: Figure 7 and Figure 8As shown in the table, by comparing the HI curves constructed by different methods, it can be intuitively seen that the method proposed in this invention can more accurately characterize the degradation process of rolling bearings. The quantitative analysis results in Table 2 further demonstrate that the method proposed in this invention achieves the best performance on the vast majority of evaluation metrics in all construction tasks. Specifically, in Task 1 and Task 2, the method proposed in this invention achieves the best performance on the three key metrics of monotonicity (Mon), robustness (Rob), and mixed scale (HS). Although the KPCA method is slightly better than the method proposed in this invention on the trend (Tred) metric, ... Figure 7 As shown in (d), the HI curve constructed by KPCA exhibits a clear linear growth trend with significant fluctuations after step 1146, which deviates significantly from the actual degradation process. In contrast, the HI curve constructed by the method proposed in this invention not only maintains good trend characteristics but also more accurately reflects the actual degradation state of the rolling bearing. These results fully demonstrate that, compared with existing methods, the method proposed in this invention has superior early fault detection capability and degradation trend characterization capability.

[0152] Table 2 Comparison of evaluation indicators between the present invention and other HI construction methods

[0153]

[0154] A comprehensive analysis of the experimental results further validates the superiority of the method proposed in this invention. For example... Figure 9 As shown, by comparing the average performance of HIs constructed by different methods, it is clear that the method proposed in this invention outperforms all the comparative methods. In particular, the method proposed in this invention demonstrates significant advantages in the two key indicators characterizing the rolling bearing degradation process: monotonicity (Mon) and trend (Tred). Statistical results show that the method proposed in this invention achieves the highest score in all evaluation indicators, and this excellent overall performance fully demonstrates the strong generalization ability of the method proposed in this invention across different operating conditions. Compared with other comparative methods, the HI constructed by the method proposed in this invention exhibits a higher correlation with the actual degradation process of rolling bearings, not only maintaining excellent monotonically increasing characteristics but also accurately reflecting the degradation trend, while possessing excellent robustness. These characteristics make the method proposed in this invention of significant application value in the health monitoring of actual industrial equipment.

[0155] (2) Comparative analysis of initial degradation point determination;

[0156] The following comparison of three thresholding methods with three commonly used model classification methods verifies the detection performance of the proposed rolling bearing degradation point detection method based on sparse time-frequency autoencoder adversarial methods under different individual differences.

[0157] Taking the rolling bearing PB14 as an example, the original vibration signal is as follows:Figure 10 As shown in (a), it can be represented as HI splicing sequence such as Figure 10 As shown in (b), it can be represented as .

[0158] like Figure 11 As shown, the first degradation interval exceeding the threshold for Historical Window The slopes of the two windows are respectively , The HI consistency test was performed on the two windows (Formula (16)), and the test result was as follows: It meets the inspection standards, that is The initial degradation range for rolling bearing PB14 was determined. After expansion, exponential fitting is performed, and the fitting function is: The optimal curvature of the fitted curve is found to be 0.0798. That is, 1081 is the initial degradation point of the rolling bearing PB14.

[0159] Table 3. IDP detection results determined by different threshold methods

[0160]

[0161] For four sets of rolling bearing data (PB11, PB12, PB16, and PB23) in the PHM dataset, the performance of three over-threshold detection methods (IDP) was compared experimentally. As shown in Table 3, the detection results of the three over-threshold detection methods on the four sets of rolling bearing data show significant differences. For further analysis and comparison, Figure 12 The detection performance of different threshold methods for determining IDPs is systematically demonstrated.

[0162] Figure 12 The detection results are presented intuitively using a multi-subgraph format. Figures (a.1)-(d.1) show the detection performance of different thresholding methods on the original vibration signal. Figures (a.2)-(d.2) present the detection results of the kurtosis2σ method on its degradation index kurtosis. Figures (a.3)-(d.3) compare the detection performance of the proposed method and the RMS2σ method on the RMS degradation index. Figures (a.4)-(d.4) show the detection performance of the RMS3σ method on the RMS degradation index. This multi-level and multi-dimensional visualization method helps to comprehensively evaluate the detection performance of each method under different signal representations.

[0163] By comparing Table 3 and Figure 12Comparative analysis reveals the following conclusions: In IDP detection of four types of rolling bearing data (PB11, PB12, PB16, and PB23), different threshold-breaking methods exhibit significant performance differences. While the kurtosis2σ method achieved accurate detection on rolling bearing PB11, it resulted in false detections on the other three. The RMS2σ method produced false detections on all tested rolling bearings. Although the RMS3σ method had a prediction lag compared to the RMS2σ method, its IDP detection results for the fourth group of rolling bearings were earlier. In contrast, the method proposed in this invention accurately located the degradation initiation point of the rolling bearing in all test cases, demonstrating significant superiority.

[0164] In-depth analysis of the results reveals that the RMS3σ method, due to its wider threshold range, exhibits better tolerance for signal noise, resulting in a predictive lag compared to the RMS2σ method. However, the presence of noise in the RMS threshold leads to false detections in both thresholding methods. Particularly noteworthy is the kurtosis2σ method, which successfully overcomes noise interference in the early stages (t=1~1000) of rolling bearing PB11 through a mechanism involving the determination of two consecutive kurtosis values, explaining its superior performance on PB11. However, in cases with more severe noise, such as rolling bearing PB12, the kurtosis2σ method still struggles to avoid false detections. These experimental results not only confirm the significant impact of noise interference on traditional over-threshold detection methods but also verify the effectiveness of the proposed method in noise suppression.

[0165] Table 4. IDP detection results determined by different model classification methods

[0166]

[0167] To further verify the effectiveness of the method proposed in this invention, a comparative analysis was conducted on all rolling bearings under the first working condition using three model classification methods and the method proposed in this invention. The comparison results are shown in Table 4. Figure 11 This paper demonstrates different model classification methods for determining IDP detection results. The three model classification methods are: the Hidden Markov Model (HMM) method; the LSTM-RNN method (ZHANG B, ZHANG S, LI W. Bearing performance degradation assessment using long short-term memory recurrent network[J]. Computers in Industry, 2019, 106: 14-29.); and the CNN (Convolutional Neural Network) method.

[0168] According to Table 4 and Figure 13 Comparative analysis of experimental results shows that the method proposed in this invention can achieve accurate prediction of IDP on all seven sets of rolling bearing datasets, while the other three model classification methods (LSTM-RNN, CNN and HMM) have different degrees of false detection and prediction bias.

[0169] pass Figure 13 Vibration signal feature analysis of rolling bearings PB11 and PB13 in (a) and (c) reveals that these two bearings exhibit a slow upward trend in vibration amplitude after degradation begins, with minimal difference in data characteristics between their early degradation stage and healthy state. This subtle degradation characteristic can easily lead to IDP detection lag in supervised learning-based classification models (such as LSTM-RNN and CNN methods). In contrast, unsupervised learning-based HMM methods improve IDP prediction performance to some extent by mining the temporal correlation features and state transition patterns of vibration signals.

[0170] And for such Figure 13 (d) illustrates the degradation process of rolling bearing PB14. During the healthy phase, the vibration signal of PB14 remains stable. However, upon entering the degradation phase, the vibration amplitude rapidly increases, exhibiting typical rapid degradation characteristics. This significant state transition creates a substantial data difference between the healthy and degradation phases, providing favorable conditions for accurate classification by various model classification methods. The proposed method and the three model classification methods all achieved good prediction results in the IDP detection of this rolling bearing PB14.

[0171] from Figure 13 As shown in (b) of the vibration signal of the rolling bearing PB12, although there is significant random noise interference during its healthy phase, no obvious false detections were observed in the proposed method and the three model classification methods. This result indicates that the model classification-based IDP detection method and the proposed method have strong noise robustness and can effectively overcome the influence of random noise on the detection results.

[0172] Further analysis Figure 13 (f) It can be observed that the LSTM-RNN and CNN methods exhibited false positives on the rolling bearing PB16 data. This is because the rolling bearing PB16 exhibits significant local fluctuations during its healthy phase, causing interference to the supervised learning-based classification model during training. Figure 13The health stage signals of rolling bearings PB15, PB16, and PB17 in (e), (f), and (g) show that the health status of these rolling bearings PB15, PB16, and PB17 is not completely stable, but exhibits a certain overall trend of fluctuation. This fluctuation makes it difficult for the HMM method, which relies on unsupervised state modeling, to accurately distinguish between healthy and deteriorating states, thus resulting in false positives.

[0173] In summary, the analysis above shows that model-based IDP detection methods (including LSTM-RNN, CNN, and HMM) are susceptible to local fluctuations or overall trend changes in the health phase. Furthermore, supervised learning methods such as LSTM-RNN and CNN require pre-training with samples from both normal and severely faulty states, increasing data acquisition costs. In contrast, the method proposed in this invention, based on a dynamic over-threshold strategy, directly performs IDP detection on the HI generated by the STFA-GAN network model, eliminating the need for additional training samples and reducing data requirements.

[0174] This invention addresses the false detection problem caused by vibration signal noise interference in rolling bearing IDP detection, proposing a rolling bearing degradation point detection method based on sparse time-frequency autoencoder adversarial network (STFA-GAN). The original vibration signal is denoised and reconstructed using a sparse time-frequency autoencoder adversarial network (STFA-GAN) model. Based on the reconstructed HI (hysteresis indices), a progressive IDP detection strategy is adopted. Specifically, an initial degradation interval is determined using a dynamic threshold method, and then the final degradation point is accurately located within this initial degradation interval using an optimal curvature method. Two sets of comparative experiments verify the effectiveness of the proposed method, and the following important conclusions are drawn:

[0175] (1) In terms of signal denoising and reconstruction, experimental analysis of the PHM dataset shows that the STFA-GAN network model, compared with the currently popular HI construction methods based on machine learning and deep learning, exhibits superior trend, monotonicity, and robustness under both the same and different working conditions. Furthermore, the denoising and reconstruction operation of the STFA-GAN network model provides a more reliable data foundation for subsequent degradation point detection based on HI, significantly reducing false detections and false negatives caused by noise interference.

[0176] (2) In terms of IDP detection, experimental comparison and analysis show that the progressive IDP detection strategy proposed in this invention has obvious advantages over the traditional fixed threshold method and model classification method. It eliminates the problems of the traditional fixed threshold method causing the prediction point to be advanced due to the fixed threshold and the traditional model classification method being easily affected by local fluctuations or overall trend changes in the health stage.

[0177] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation, characterized in that, Comprising the following steps: HI construction stage: the sparse time-frequency auto-encoding adversarial network model is used to learn the features and enhance the data of the vibration signals collected under the health state of the rolling bearing, and the HI curve sensitive to the performance degradation of the rolling bearing is constructed through deep feature extraction; IDP determination stage: first, the HI sequence output by the sparse time-frequency auto-encoding adversarial network model is analyzed by using the dynamic sliding window technology, and the linear residual analysis method is combined to construct a dynamic threshold to determine the potential initial degradation interval; then the curvature optimal method is introduced in the initial degradation interval, and the change rate of the geometric features of the HI curve is quantified to finally accurately locate the initial degradation point with significant statistical significance.

2. The rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation according to claim 1, characterized in that, The sparse time-frequency auto-encoding adversarial network model comprises a generator and a discriminator; the generator is composed of a parallel time-frequency domain auto-encoder, a channel attention mechanism, a sparse regularization network with a sparse long short-term memory layer and a decoder; the parallel time-frequency domain auto-encoder comprises a time domain encoder and a frequency domain encoder, the structure of the time domain encoder and the frequency domain encoder is the same, and both are composed of two convolutional layers and two pooling layers, and a batch normalization function and a ReLU activation function are added to the convolutional layer of the encoder; the decoder is responsible for outputting the reconstructed vibration signal, and its structure comprises two deconvolutional layers and two up-sampling layers, corresponding to the structure of the encoder; the original vibration signal sequence is input into the generator, the generator is trained to generate false vibration signals and reconstruct the vibration signal, and the sparse coding in space is optimized by combining sparse regularization; The discriminator is composed of multiple convolutional layers and a fully connected layer, the multiple convolutional layers use the same activation function, and the fully connected layer is used to help the generator learn the distribution of the input data; the original vibration signal and the reconstructed vibration signal are simultaneously transmitted to the discriminator, so that it can distinguish between generated signals and real signals, thereby prompting the generator to generate high-quality reconstructed vibration signals closer to the original input.

3. The rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation according to claim 2, characterized in that, The original vibration signal is simultaneously input into the parallel structure of the time domain encoder and the frequency domain encoder; the original vibration signal needs to be subjected to fast Fourier transform before the frequency domain encoder; the features extracted by the parallel time-frequency domain auto-encoder are input into the channel attention mechanism to enhance the feature representation; after the time domain and frequency domain features are spliced, they are transmitted to the sparse regularization network with a sparse long short-term memory layer for sparse regularization processing, which is applied to the latent state in the generator to learn the over-complete dictionary and generate sparse codes.

4. The rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation according to claim 1, characterized in that, In the process of adversarial training, the training loss L of the sparse time-frequency self-encoding adversarial network model is composed of a time-frequency reconstruction loss , an adversarial loss and a sparsity loss . （1）； The time-frequency reconstruction loss The mathematical expression is: （4）； wherein, are weight coefficients, is a time domain reconstruction loss, is a frequency domain reconstruction loss; the adversarial loss The mathematical expression is: （6）； wherein, Lg is the generator loss, Ld is the discriminator loss; The sparse loss The mathematical expression is: （9）； wherein is a sparse regularized dictionary, is a sparse code generated by sparse regularization, is a sparse code input to the generator, is a Frobenius norm, is is a norm, with the upper index T denoting the transpose.

5. The rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation according to claim 4, characterized in that, The mathematical expressions of the time domain reconstruction loss and the frequency domain reconstruction loss are respectively: （2）； （3）； wherein, is the batch size, is the input raw vibration signal, is the reconstructed vibration signal, is the frequency domain signal after fast Fourier transform.

6. The rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation according to claim 4, characterized in that, The loss magnitude is automatically adjusted to ensure dynamic updating during each round of training and balance the loss magnitude, and its mathematical expression is: （5）。 7. The rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation according to claim 4, characterized in that, The discriminator loss And the mathematical expression of the generator loss is: （7）； （8）； wherein, is the discriminator input, is the reconstructed signal input to the discriminator, is the weight coefficient, is the signal created by randomly combining the original vibration signal and the reconstructed vibration signal, and as the Wasserstein distance label, is the original vibration signal input, is the reconstructed vibration signal, is the batch size.

8. The rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation according to claim 1, characterized in that, In the process of adversarial training, the sparse time-frequency auto-encoding adversarial network model learns the distribution of the data under the health state of the rolling bearing, and when the real-time monitored vibration signal is input again, the discriminator score and the residual loss generated in the iteration process are used as indicators of the degradation degree, and the mathematical expression is as follows: （10）； wherein, is a weight coefficient; is a residual loss, whose mathematical expression is as follows: （11）； wherein, is the input raw vibration signal, is the discriminator input, is the reconstructed vibration signal.

9. The rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation according to claim 1, characterized in that, In the IDP determination stage, the initial degradation interval is determined by constructing a dynamic threshold by using residual analysis combined with a sliding window, comprising the following steps: S3.1.1: Calculate the residual sequence in the window; The window size is determined by a grid search method; the calculation formula of the residual sequence is as follows: (12)； wherein, is the unsupervised health index, is the fit value within the window, t is the time instant; S3.1.2: Constructing dynamic threshold; As the window slides over the time series, the mean of the residuals within the different time series windows is computed and the standard deviation A dynamic threshold is constructed whose mathematical expression is:​ (13)； (14)； (15)； wherein, is a window start time, is a window end time; is a sensitivity coefficient; S3.1.3: judge whether the number of continuous over-threshold points N in the window is greater than or equal to 3; if not, return to step S3.1.1; if yes, accumulate the degradation interval; in the window, set the first over-threshold time point as the starting point of the degradation interval, and the ending point is the first time point below the threshold ;​​​​​ S3.1.4: HI consistency test; The HI monotonicity test is performed on the five identified degradation intervals in sequence, and the initial degradation interval is screened out by comparing the HI fitting slope of the previous window of the identified degradation interval. The mathematical expression of the HI consistency test is: (16)； wherein, is the current window slope, are the slope and standard deviation of the history window, respectively.

10. The rolling bearing degradation point detection method based on sparse time-frequency self-encoding confrontation according to claim 1, characterized in that, In the IDP determination stage, the HI curve in the degradation interval is fitted by an exponential function, and the IDP is determined by finding the curvature extreme point of the fitted HI curve.

Images