Rolling bearing residual life prediction method and system based on generative data augmentation

Abstract

The application discloses a rolling bearing residual life prediction method and system based on generative data enhancement. First, the vibration signal of the rolling bearing is collected and preprocessed, then forward diffusion noise is added and reverse denoising based on time sequence consistency constraint is performed to reconstruct the supplementary samples of the key degradation stage; then, screening is performed based on the distribution difference between the real samples and the generated samples; then, the screened samples are processed by using the empirical mode decomposition and multi-dimensional time domain statistical features are extracted, then the double attention feature fusion network composed of the bidirectional gate recurrent unit and the Transformer encoder is input, the self-attention, the interaction attention and the channel splicing are used to fuse the local degradation features and the global long-range degradation trend, and finally the fusion features are input into the full connection layer to predict the residual service life of the rolling bearing. The application can alleviate the problems of the scarcity of key degradation samples and the uncontrollable quality of generated samples, and improve the residual life prediction accuracy and stability of the rolling bearing.

Description

Technical Field

[0001] This invention relates to the field of rolling bearing failure prediction and health management technology, specifically to a method and system for predicting the remaining life of rolling bearings based on generative data augmentation. Background Technology

[0002] Rolling bearings are critical components in rotating machinery that bear loads and transmit motion; their health status directly affects the operational stability and service performance of the power system. Monitoring the degradation state of rolling bearings and predicting their remaining service life is of significant engineering value for making maintenance decisions, avoiding sudden failures, and reducing operation and maintenance costs. With the continuous development of data-driven fault prediction and health management technologies, establishing remaining service life prediction models using full-life-cycle monitoring data collected by vibration sensors has become an important technical approach for condition-based maintenance of rolling bearings.

[0003] Data-driven methods such as convolutional neural networks and recurrent neural networks rely on large amounts of well-labeled training data for predictive performance. However, acquiring full-lifecycle failure data for rolling bearings is costly, and samples for key degradation stages are scarce and unevenly distributed, making it difficult for training data to fully cover the complete evolutionary process from healthy operation and slow degradation to rapid failure. Traditional generative adversarial networks are also prone to training instability and pattern collapse in long-term sequence generation tasks, making it difficult for generated samples to maintain the continuous trend of the real degradation process.

[0004] Furthermore, rolling bearing vibration signals are nonlinear time series, typically containing complex background noise, modulation components, and transient impacts. Directly inputting a single time-domain waveform into the prediction network makes it difficult to adequately separate fault information across different frequency bands; a single-path network also struggles to simultaneously capture local short-term impacts, medium-sequence evolution, and long-term trends throughout the entire lifespan. Therefore, a remaining life prediction method is needed that can supplement key missing data, perform quality screening on generated samples, and integrate multi-frequency band statistical characteristics, local temporal dependencies, and global degradation trends. Summary of the Invention

[0005] To address the challenges of scarce samples during critical degradation stages, uncontrollable sample quality, difficulty in fully extracting degradation features from complex vibration signals, and the inability of single-sequence models to simultaneously account for local short-term dependencies and global long-term trends, this invention proposes a method and system for predicting the remaining life of rolling bearings based on generative data augmentation. This method enables vibration signal supplementation, degradation feature extraction, and remaining life prediction under conditions of insufficient samples during critical degradation stages. The specific technical solution is as follows:

[0006] A method for predicting the remaining life of rolling bearings based on generative data augmentation includes:

[0007] Step 1: Acquire and preprocess the rolling bearing vibration signal;

[0008] Step 2: Perform forward diffusion noise addition and reverse denoising based on temporal consistency constraints on the preprocessed signal to reconstruct supplementary samples for the key degradation stage;

[0009] Step 3: Based on the KL divergence between real samples and generated samples and a preset threshold, the generated samples are filtered. If the threshold requirement is not met, the generation process in Step 2 is repeated.

[0010] Step 4: Process the screened samples using empirical mode decomposition and extract multidimensional time-domain statistical features;

[0011] Step 5: Construct a dual-attention feature fusion network consisting of a bidirectional gated recurrent unit and a Transformer encoder. Input the multidimensional temporal statistical features and fuse the local degradation features output by the bidirectional gated recurrent unit and the global long-range degradation trend output by the Transformer encoder through self-attention, interactive attention and channel splicing.

[0012] Step 6: Input the fused features into the fully connected prediction layer to complete model training and nonlinear mapping, and output the predicted value of the remaining service life of the rolling bearing.

[0013] Furthermore, the preprocessing in step 1 includes data cleaning, linear normalization, and dividing the dataset into training and test datasets using the leave-one-out-of-random method.

[0014] Furthermore, step 2 specifically includes:

[0015] Step 2.1: In the preprocessed signal, the key data target to be generated will be... The associated upstream and downstream process samples serve as conditional information, including upstream and downstream process observations, operating condition information, or degradation status characteristics;

[0016] Step 2.2: By adding noise through forward diffusion, the conditional information is gradually degraded into pure Gaussian noise. The forward diffusion noise addition process is modeled as a parameterized Markov diffusion chain with a total of K diffusion time steps. At each discrete time step k, 1 ≤ k ≤ K, based on the state of the previous time step... Sample to generate the noisy state at the current moment. Gaussian noise is gradually injected according to the variance scheduling parameters;

[0017] Step 2.3: By performing inverse denoising based on temporal consistency constraints, the pure Gaussian noise is gradually restored to reconstruct the noise that conforms to the time consistency constraints. The data distribution of features yields complete data containing missing information.

[0018] Furthermore, the network model used in the reverse denoising includes a long short-term memory network, a broadcast layer, multi-head self-attention, and one-dimensional convolution;

[0019] Modeling the evolutionary dependency of program sequences in long short-term memory networks;

[0020] By utilizing the location encoding of the broadcast layer to fuse the location embedding of the diffusion time step, and combining it with the process mechanism conditions, the location features and upstream and downstream process conditions jointly participate in the reverse denoising.

[0021] Multi-head self-attention is used to extract global contextual features and capture global long-range degradation relationships;

[0022] Local short-term impact features of adjacent sampling points are extracted by one-dimensional convolution, residual fusion is completed, and noise is removed from the vibration signal.

[0023] Furthermore, the Long Short-Term Memory (LSTM) network introduces a gating mechanism, utilizing the synergistic effect of the forget gate, input gate, and output gate to regulate the information flow. The forget gate, based on the hidden layer state of the previous time step and the current input, calculates the weight coefficients within the interval (0,1) using the Sigmoid activation function, determining whether historical information in the cell state of the previous time step should be discarded or retained. The input gate filters the new input at the current time step, generating candidate memory information through the hyperbolic tangent function, and quantifies the importance of new information to the current state by combining the gating signal of the Sigmoid function. The retention result of the forget gate is merged with the updated content of the input gate to complete the cell state refresh, and the integrated cell state is mapped to the current hidden layer output according to the control logic of the output gate.

[0024] Furthermore, the multi-head self-attention first maps the query vector set and the key-value pair vector set to different feature dimension subspaces through multiple independent linear projection matrices; then, it independently performs scaled dot product attention calculations in each subspace, measures the relevance of the query vector and the key vector by the dot product, introduces a scaling factor to alleviate the gradient minima problem caused by excessive dimensionality, and generates attention weights in the form of a probability distribution through the Softmax function; finally, it integrates features through a linear transformation layer to achieve effective integration and dimensionality restoration of information from multiple semantic subspaces.

[0025] Furthermore, the reverse denoising network model employs variational inference theory to construct a loss function during training. By maximizing the variational lower bound, the loss function is simplified to the mean square error between the predicted noise and the true Gaussian noise.

[0026]

[0027] in, This represents the expected value of the diffusion step number k. This represents network prediction noise; The key vibration signal sequence to be generated; That is, the cumulative retention coefficient is the product of the retention coefficients of each signal from step 1 to step k;

[0028] By updating the network parameters θ by minimizing the objective function, the model can extract effective rolling bearing vibration signal features from arbitrary noise, thus achieving the goal of resolving missing data. Filling and generation;

[0029] Further, step 4 specifically includes: based on the effective frequency band distribution characteristics of the rolling bearing vibration signal and prior experience in processing similar working condition data, using empirical mode decomposition to decompose the vibration signal into several intrinsic mode function components to cover the main fault characteristic frequency band and suppress the effects of endpoint effects and mode aliasing; each component separates the high-frequency noise, fault impact components and low-frequency trend components in the original signal; then extracts 11 types of time-domain statistical features from each intrinsic mode function component, including variance, energy, root mean square, mean, wavelength, median, peak factor, higher-order statistics, entropy, arctangent standard deviation and arcsine standard deviation, and forms a multidimensional degradation feature sequence according to the sampling time.

[0030] Furthermore, the feature extraction and fusion process in the dual attention feature fusion network specifically includes:

[0031] Step 5.1: The bidirectional gated recurrent unit processes the input multidimensional time-domain statistical feature sequence through two independent GRU networks in the forward and backward directions respectively: forward propagation captures the degradation and evolution trajectory from the past to the present, generating a forward hidden state that reflects historical information; backward propagation uses the observation data of subsequent time steps to assist the current inference, generating a backward hidden state that contains future context, and after obtaining the bidirectional information of each time step, the forward and backward hidden states are concatenated into a comprehensive hidden state;

[0032] Step 5.2: Use self-attention to weight the output of the bidirectional gated recurrent unit and form local features;

[0033] Step 5.3: The Transformer encoder injects absolute positional information into the parallel input feature sequence through positional encoding, making up for the lack of temporal positional sense in the pure attention mechanism. In the interactive attention layer, it uses the query vector in the global feature space and the key-value pairs in the local feature space to perform interactive calculations and output global features.

[0034] Step 5.4: The local and global features are fused at the channel dimension in the splicing layer. The fused feature vector contains both short-term fault impact information and vibration cycle degradation trend information.

[0035] A rolling bearing remaining life prediction system based on generative data augmentation includes:

[0036] Data preprocessing module: used to clean and normalize the acquired raw vibration signals, and to divide them into training and test sets;

[0037] The data generation module with temporal consistency constraints is used to perform forward diffusion noise addition and reverse denoising based on temporal consistency constraints on the preprocessed signal, and to supplement key missing samples.

[0038] Data evaluation module: Based on the KL divergence between real samples and generated samples and a preset threshold, the generated samples are screened.

[0039] Empirical Mode Decomposition Module: Used to decompose the screened samples into intrinsic mode function components and extract time-domain statistical features;

[0040] Dual Attention Sequence Feature Fusion Module: Employing a bidirectional gated recurrent unit and a Transformer encoder, it fuses local degradation features and global long-range degradation trend features through self-attention, interactive attention, and channel splicing.

[0041] Fully connected prediction module: Input fused features, output remaining lifetime prediction value.

[0042] The advantages and beneficial effects of this invention are as follows:

[0043] This invention utilizes a conditional diffusion mechanism and a temporal consistency constraint denoising network to supplement missing data in key degradation stages. It uses one-dimensional convolution, LSTM, and multi-head self-attention to jointly model local short-term impacts, medium-term sequence evolution, and global long-term degradation trends, ensuring that the generated vibration signal maintains continuity between adjacent time steps and the rationality of staged degradation. It also uses generated data based on KL divergence to evaluate closed-loop screening and remove candidate samples that do not conform to the distribution relationship between upstream and downstream processes, thereby improving the credibility of the enhanced data.

[0044] This invention utilizes empirical mode decomposition to decompose enhanced vibration signals into eigenmode function components of multiple different frequency bands and extracts multidimensional time-domain statistical features, which can reduce the interference of complex background noise and expand the dimensions of degradation features. By using a parallel interactive structure of bidirectional BiGRU branch and Transformer encoder branch, the local fault impact and the full life cycle degradation trend are deeply integrated, thereby improving the stability, accuracy and generalization ability of rolling bearing remaining service life prediction under the condition of insufficient key degradation samples. Attached Figure Description

[0045] Figure 1 This is an overall flowchart of a method for predicting the remaining life of rolling bearings based on generative data augmentation, according to an embodiment of the present invention.

[0046] Figure 2This is a diagram of the temporal consistency denoising network structure according to an embodiment of the present invention.

[0047] Figure 3 This is a flowchart of the data quality assessment module in an embodiment of the present invention.

[0048] Figure 4 This is the IMF component map obtained from the empirical mode decomposition of an embodiment of the present invention.

[0049] Figure 5 This is a structural diagram of the dual attention fusion module according to an embodiment of the present invention.

[0050] Figure 6 This is a comparison diagram of the generated features and the original features in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0052] like Figure 1 As shown, this embodiment provides a method for predicting the remaining life of rolling bearings based on generative data augmentation, including the following steps:

[0053] Step 1: Collect rolling bearing vibration signals, perform data preprocessing, and divide the data into training and testing sets; specifically including:

[0054] Step 1.1: Clean the collected raw rolling bearing vibration signals to eliminate the influence of abnormal data on subsequent degradation feature extraction.

[0055] Step 1.2: Perform linear normalization to unify the data distribution.

[0056] In engineering signal processing, features extracted from raw vibration signals are often distributed across different dimensions and numerical ranges. Directly feeding these features into a prediction model can weaken their comparability due to scale inconsistencies, thus affecting model training performance. To map these statistical features from vibration signals to the same numerical range, linear normalization is necessary. This reduces the impact of differences in variable magnitudes, facilitating subsequent calculations and parameter settings, while preserving the inherent physical meaning of the data and accelerating network convergence and improving prediction accuracy when predicting equipment remaining life. This invention employs a linear normalization strategy based on maximum and minimum values. Let the vibration amplitude sequence obtained from a single sampling be:

[0057]

[0058] in, Let be the k-th original vibration sample, k = 1, 2, ..., N. Let The minimum sample value in sequence S. Let the maximum sample value in sequence S be denoted as . Then for any sample, we have:

[0059]

[0060] in, This represents the normalized value corresponding to the k-th sample, whose value range is compressed to the interval [0,1], and is used as the input for the subsequent RUL prediction model.

[0061] Step 1.3: Use the leave-one-out-of-random method to split the training dataset and the test dataset.

[0062] This embodiment selects the operating condition 1 data from the PHM2012 Challenge public dataset and the XJTU-SY rolling bearing accelerated life test dataset as the benchmark verification objects. Both datasets are widely used in the study of rolling bearing degradation process and remaining life prediction, and have full life cycle monitoring data from normal operation to complete failure. In terms of data partitioning strategy, in order to closely resemble the application scenario of the model predicting the life of a single rolling bearing in actual engineering, this embodiment adopts the leave-one-out method: for a selected operating condition, all data of one rolling bearing is randomly selected from the samples of all rolling bearings under that condition as the test set, which is used to simulate its unknown degradation process and predict its remaining life; the data of all other rolling bearings under the same operating condition constitute the training set, which is used to build the degradation model. In order to effectively control the fluctuation of results caused by factors such as random model initialization and improve the stability and reliability of experimental conclusions, the training-test process is independently repeated 7 times. Finally, the average of the mean absolute error and the root mean square error are used as the comprehensive evaluation index of the model's remaining life prediction performance.

[0063] Step 2: Perform time-series consistency generative data augmentation to generate supplementary samples for key degradation stages. This specifically includes:

[0064] Step 2.1: Combine with the key data target to be generated Related upstream and downstream process samples , As conditional information, it is input into the conditional diffusion probability model;

[0065] Step 2.2: Perform forward diffusion and noise injection. During the forward diffusion process of the model, stepped Gaussian noise is injected to interfere with the original data features. After multiple rounds of diffusion operations, the original process data distribution gradually degenerates into irregular pure Gaussian noise.

[0066] Specifically, the forward diffusion process involves directing the original complete rolling bearing vibration signal sequence... Gaussian noise is gradually injected into the signal to simulate the entropy increase and information destruction process, which is then modeled as a parameterized Markov diffusion chain. Assuming a total of K diffusion steps, at each discrete time step k, 1 ≤ k ≤ K, the model is based on the state of the previous time step. Sample to generate the noisy state at the current moment. .set up Define the signal preservation coefficient to preset variance scheduling parameters. And the cumulative retention coefficient. Using the reparameterization technique, the noisy sample at step k... Can be directly from raw data Represented as:

[0067]

[0068] in, ~N(0,I) represents a random noise vector that follows a standard normal distribution. That is, the cumulative retention coefficient is the product of the signal retention coefficients from step 1 to step k. As k approaches K, the cumulative retention coefficient approaches 0, making... It approximately converges to a pure Gaussian noise distribution GT. This process not only disrupts the temporal distribution of the original data, but also provides a training target for the subsequent model to learn the inverse transform of recovering ordered signals from disordered noise.

[0069] Let the key vibration signal sequence to be generated be... The conditional information is c, where c consists of upstream and downstream process observations, operating condition information, or degradation state characteristics. The forward diffusion process is as follows:

[0070]

[0071] in, ;make The cumulative coefficient is the coefficient from step 1 to step t. The product of these values, as the number of diffusion steps t increases, gradually decreases the cumulative retention coefficient, indicating that the original signal information is progressively destroyed. It gradually approaches the standard Gaussian noise distribution.

[0072] Step 2.3: Perform reverse diffusion and recover missing data.

[0073] The process then transitions to a progressive inverse phase, using an iterative denoising mechanism to gradually restore the random Gaussian distribution and reconstruct a distribution that conforms to... The true data distribution of features.

[0074] The reverse generation process uses a parameterized denoising network to estimate the noise injected during the forward process, thereby gradually restoring the original signal distribution.

[0075] The reverse generation process aims to progressively recover complete data containing missing information from pure Gaussian noise ground plane. Because the true posterior distribution of the data is recovered from the noise. Since it is mathematically unsolvable, this invention uses a parameterized deep neural network to approximate this inverse process.

[0076] Step 2.4: Define temporal consistency constraints and use LSTM to extract the program sequence features of the degraded sequence.

[0077] To improve the temporal continuity and degradation stage consistency of the generated data, this invention integrates a denoising network based on temporal consistency constraints in the reverse denoising stage. Temporal consistency refers to the continuous and smooth evolution of the generated vibration signal between adjacent time steps, and the overall trend conforming to the staged degradation law of rolling bearings from healthy operation, slow degradation to rapid failure, avoiding abnormal jumps in the generated signal that do not match the actual degradation process. Specifically, one-dimensional convolution is used to capture local short-term fluctuations between adjacent sampling points, a Long Short-Term Memory (LSTM) network is used to maintain the mid-range evolution relationship of the degradation sequence, and multi-head self-attention is used to introduce global contextual information. By jointly utilizing local fluctuations, historical states, and global degradation trends, the model can more accurately estimate noise and reconstruct vibration signals with better temporal fidelity and degradation rationality. The denoising network, such as... Figure 2 As shown.

[0078] The backdiffusion process extracts high-dimensional temporal features of noise through a Long Short-Term Memory (LSTM) network. To address the vanishing gradient problem in long-sequence data, LSTM introduces a gating mechanism, utilizing the synergistic effect of the forget gate, input gate, and output gate to regulate information flow. The forget gate, based on the hidden layer state of the previous time step and the current input, calculates weight coefficients within the interval (0,1) using the Sigmoid activation function, determining whether historical information from the previous cell state should be discarded or retained. The input gate filters the new input at the current time step, generating candidate memory information using the hyperbolic tangent function and combining it with the gating signal of the Sigmoid function to quantify the importance of new information to the current state. The network integrates the retained results from the forget gate with the updated content from the input gate, completing the cell state refresh, and maps the integrated cell state to the current hidden layer output according to the control logic of the output gate. The state update and output process is as follows:

[0079]

[0080]

[0081] in, and These represent the cell states at the current and previous time points, respectively. This represents the output of the hidden layer at the current moment. This indicates the alternative update information at the current moment. , and ...

[0082] Step 2.5: Inject temporal position information into the denoising network using broadcast layer position coding.

[0083] In the conditional diffusion probability model, a broadcast layer position coding is introduced and combined with process mechanism conditions, so that position features and upstream and downstream process conditions jointly participate in reverse denoising.

[0084] Step 2.6: Extract global contextual features using multi-head self-attention.

[0085] Specifically, to enhance the ability of convolutional networks to dynamically capture key information in input data, this invention introduces a multi-head attention mechanism. The model first maps the query vector set and key-value pair vector set to different feature dimension spaces using multiple independent linear projection matrices. Then, it independently performs scaled dot product attention calculations within each subspace, measuring the relevance of the query vector and key vector by the dot product. A scaling factor is introduced to mitigate the gradient minima problem that may result from excessive dimensionality, and a probability distribution of attention weights is generated using the Softmax function. Finally, a linear transformation layer integrates the features, achieving effective integration and dimensionality restoration of multiple semantic subspace information. The fused features are output through a residual structure, and its core computational logic is as follows:

[0086]

[0087] Where Q, K, and V represent the query matrix, key matrix, and value matrix after linear transformation, respectively. The dimension of the key vector is represented by Softmax, which is a normalized exponential function. The multi-head attention mechanism performs differentiated linear transformations on the same input by stacking multiple attention layers in parallel. The outputs of each parallel head are concatenated and then fused using a final linear weight matrix.

[0088]

[0089] in, Let represent the calculation result of the i-th attention head, and h represent the total number of heads in the multi-head mechanism. This represents the linear weight matrix used for feature fusion in the output stage, and Concat represents the vector concatenation operation.

[0090] Step 2.7: Extract local short-term features using one-dimensional convolution and complete residual fusion.

[0091] Finally, one-dimensional convolution is used to extract noise features, thereby removing the noise.

[0092] From the perspective of rolling bearing degradation mechanism, bearing vibration signals simultaneously contain three types of information: local impact, intermediate degradation evolution, and long-term trend. Local impact typically originates from the periodic contact between rolling elements and raceway surface defects, manifesting as short-term high-frequency abrupt changes; intermediate degradation evolution reflects continuous processes such as crack propagation, wear accumulation, and changes in lubrication state; long-term trend reflects the overall degradation path of the equipment from healthy operation to final failure. Therefore, the denoising network is composed of one-dimensional convolution, LSTM, and multi-head self-attention. One-dimensional convolution is used to extract local short-term impact features, LSTM is used to model the evolutionary dependencies of the sequence, and multi-head self-attention is used to capture long-range global degradation relationships. This design enables the backdiffusion process to comprehensively utilize local fluctuations, historical states, and global contextual information to estimate noise and reconstruct the signal.

[0093] Step 2.8: Optimize the temporal consistency constraint denoising network based on variational lower bound and noise prediction loss.

[0094] To train a temporal consistency-aware network and enable it to perform accurate denoising, this invention employs variational inference theory to construct a loss function. By maximizing the variational lower bound, the loss function is simplified to the mean square error between the predicted noise and the actual Gaussian noise.

[0095]

[0096] in, This represents the expected value of the diffusion steps k. By minimizing this objective function and updating the network parameters θ, the model can extract effective rolling bearing vibration signal features from arbitrary noise, thus enabling the processing of missing data. The filling and generation.

[0097] Within the variational inference framework, the conditional diffusion model is trained by maximizing the variational lower bound of the data likelihood. When the variance of the inverse process is fixed, the optimization objective simplifies to the mean squared error between the predicted noise and the true noise.

[0098]

[0099] The objective function is non-negative and bounded by a lower bound of 0. During training, it is minimized... This invention enables the denoising network to progressively learn the inverse mapping to recover the original degraded signal from arbitrary noise levels. It employs batch normalization, residual connections, learning rate decay, and early stopping on the validation set to suppress gradient oscillations, thus stabilizing the empirical loss during training.

[0100] Step 2 of this invention addresses the problem of insufficient samples by constructing a data generation module with temporal consistency constraints. This module uses upstream and downstream processes associated with the key data target to be generated as conditional information, inputs it into a conditional diffusion probability model, maps the data to pure Gaussian noise through a forward diffusion process, and then uses a temporal consistency constraint denoising network to perform a reverse generation process to reconstruct the data.

[0101] Step 3: Evaluate the quality of generated data. Generate candidate data for upstream and downstream processes corresponding to the target process. Based on the KL divergence between the real data and the generated data and a preset threshold, the generated data is filtered. If the threshold requirements are not met, the generation process is re-executed to select high-quality samples for data augmentation. Specifically, this includes:

[0102] Step 3.1: To improve the credibility of the generated data, a generated data evaluation module is set up. The generated sample data is only output when the evaluation meets the requirements. The data will be generated by the network to create process sample data. The mechanistic conditions of process A and process C are input together into the conditional diffusion model, and the model is then used to generate upstream and downstream process data. , .

[0103] Step 3.2: Use real data from upstream and downstream processes. , With generated data , Perform KL divergence calculation to obtain the distance between the two sample distributions.

[0104]

[0105] in, , These represent the probability distributions of the real data and the generated data, respectively.

[0106] Step 3.3: Generate samples based on feedback and selection of KL divergence threshold.

[0107] Set a KL divergence threshold to determine the generated data. , The validity of the sample data is assessed. If the distance obtained exceeds the threshold, new sample data is generated repeatedly. The reverse generation process is then performed again to obtain... , If the result is not obtained, the calculation will be repeated; otherwise, it indicates that the generated data has been generated. , and raw data , If the samples are sufficiently similar, then the generated sample data will be output. The data evaluation process is as follows: Figure 3 As shown.

[0108] Step 4: Decompose the enhanced vibration signal and extract multidimensional time-domain statistical features to form a multidimensional degradation feature sequence. Specifically, this includes:

[0109] Step 4.1: Based on the local time scale of the vibration signal itself, use Empirical Mode Decomposition (EMD) to process the enhanced vibration signal.

[0110] Although the generation module with temporal consistency constraints effectively expands the rolling bearing vibration signal samples and solves the data missing problem, the generated vibration signals are still essentially nonlinear time series signals. The original vibration signals often contain complex background noise, and the degradation characteristics of rolling bearings typically manifest as modulated signals or transient impacts. Directly inputting the time-domain waveform into the subsequent prediction network makes it difficult to fully extract its deep degradation patterns. The low dimensionality of single-time-domain features limits the model's ability to capture effective information. Therefore, this invention introduces an empirical mode decomposition module to perform secondary processing on the enhanced signal to extract multi-band intrinsic features.

[0111] Empirical Mode Decomposition (EMD) is an adaptive time-frequency processing method for signals that does not require pre-defined basis functions but decomposes based on the local time-scale characteristics of the data itself. This algorithm assumes that a complex signal is composed of a superposition of several different intrinsic mode functions (EMFs). Through a selection process, it adaptively decomposes a complex rolling bearing vibration signal into a series of EMF components arranged from high to low frequency, and a residual component representing the overall trend of the signal. Each EMF component represents a local oscillation mode of the original signal at different time scales.

[0112] Step 4.2: Separate into 7 intrinsic mode function components with frequencies arranged from high to low and a residual component characterizing the overall trend.

[0113] In the application of empirical mode decomposition (EMD) algorithms, the number of modal components is a key parameter affecting the time-frequency representation of a signal. Too many decomposition layers not only increase computational redundancy but also easily introduce meaningless spurious modes in the low-frequency band; too few layers make it difficult to effectively separate complex background noise from weak fault impacts. Based on the effective frequency band distribution characteristics of rolling bearing vibration signals and prior experience in processing similar operating conditions, the vibration signal is decomposed into seven intrinsic mode function (EMF) components to cover the main fault characteristic frequency bands and suppress the effects of endpoint effects and mode aliasing. These components separate high-frequency noise, fault impact components, and low-frequency trend components from the original signal, such as... Figure 4 As shown.

[0114] Step 4.3: Extract 11 types of time-domain statistical features from each intrinsic mode function component, namely variance, energy, root mean square, mean, wavelength, median, peak factor, higher-order statistics, entropy, arctangent standard deviation, and arcsine standard deviation, and form a multidimensional degradation feature sequence according to the sampling time.

[0115] Temporal statistical features are extracted from the components obtained by empirical mode decomposition to expand the data dimension. The expanded data is used as the original feature set for rolling bearing RUL prediction.

[0116] The formulas for the specific time-domain statistical characteristics are as follows:

[0117] variance: ;

[0118] energy: ;

[0119] Root mean square: ;

[0120] Mean: ;

[0121] wavelength: ;

[0122] Median: ;

[0123] Peak factor: ;

[0124] Higher-order statistics: ;

[0125] Entropy value: ;

[0126] Arbitrary standard deviation: ;

[0127] Arcsine standard deviation: ;

[0128] in: This indicates sorting the sequence s. and This represents the standard deviation operation, where i represents the imaginary unit and T represents the sequence length. This represents the t-th sample value.

[0129] Step 5: Construct a dual-attention feature fusion network consisting of a bidirectional gated recurrent unit (BiGRU) branch and a Transformer encoder branch. Input the multidimensional temporal statistical features and fuse the local degradation features output by the bidirectional BiGRU branch and the global long-range degradation trend output by the Transformer encoder branch through self-attention, interactive attention, and channel concatenation. Specifically, this includes:

[0130] Step 5.1: Use BiGRU to extract forward and backward local temporal features.

[0131] The degradation process of rolling bearings exhibits a high degree of temporal correlation; the current state not only represents historical accumulation but also implies future evolution trends. A bidirectional gated recurrent unit (GRU) processes the input feature sequence via two independent GRU networks, one forward and one backward: forward propagation captures the degradation trajectory from the past to the present, generating a forward hidden state reflecting historical information; backward propagation utilizes observation data from subsequent time steps to assist current inference, generating a backward hidden state containing future context. After acquiring the bidirectional information at each time step, the forward and backward hidden states are merged through a concatenation operation.

[0132]

[0133] in, The term represents the integrated hidden state after fusion at time k. The forward hidden state and the reverse hidden state represent the hidden state vectors output by the forward GRU and the reverse GRU at time k, respectively. Concat represents the vector concatenation operation.

[0134] Step 5.2: Use self-attention to weight the BiGRU output and form local features.

[0135] To further enhance the feature weights of key time steps, a self-attention mechanism is introduced after the BiGRU outputs fused features. This mechanism identifies and focuses on the time nodes that contribute most to the remaining lifetime prediction based on the internal correlation of features, while suppressing irrelevant noise interference, and outputting a weighted local feature vector. .

[0136] Step 5.3: Use the Transformer encoder and interactive attention to form global features.

[0137] The Transformer encoder branch addresses the vanishing gradient and long-range forgetting problems that exist when recurrent neural networks process extremely long sequences. The Transformer encoder first injects absolute positional information into the parallel input feature sequences through positional encoding, compensating for the lack of temporal positional awareness in pure attention mechanisms. The interactive attention layer receives features from the BiGRU branch after self-attention weighting as auxiliary input, allowing the Transformer to reference the strong local features extracted by BiGRU when calculating global dependencies. The model utilizes query vectors in the global feature space. Key-value pairs in local feature space , Perform interactive computation. Assuming the mapping function is M, the interactive attention output... for:

[0138]

[0139] in, This is the scaling factor.

[0140] Step 5.4: Through cross-branch interaction, local high-frequency fluctuation features are integrated into the global long-range degradation trend. The local features and global features are then combined to obtain the comprehensive features. .

[0141] Features output by the bidirectional interactive BiGRU branch Features of Transformer branch output Channel-dimensional fusion is performed at the splicing layer. The fused feature vector Z contains both short-term fault impact information and vibration period degradation trend information.

[0142] .

[0143] As can be seen, this step aims to simultaneously capture both the local short-term dependencies and the global long-term degradation trend during the rolling bearing degradation process, employing a sequence feature fusion network based on dual attention. This network uses a dual-branch parallel interactive architecture, consisting of a bidirectional interactive BiGRU branch and a Transformer encoder branch, respectively. Deep coupling of the two feature paths is achieved through an interactive attention mechanism. The specific structure is as follows: Figure 5 As shown.

[0144] Step 6: Input the fused features into the fully connected prediction layer to complete model training and nonlinear mapping, outputting the predicted rolling bearing remaining service (RUL) value; evaluate the prediction results using mean absolute error and root mean square error, and conduct experimental analysis on data augmentation results, different data augmentation methods, and each component module. Specifically, this includes:

[0145] Step 6.1: Complete fully connected mapping, model training, and prediction of remaining rolling bearing life.

[0146] Comprehensive feature vector It is fed into a fully connected layer for nonlinear mapping, and the predicted value of the remaining service life of the rolling bearing is output.

[0147] The experimental computing platform was built on a Linux mini-server environment, equipped with an Intel Xeon Silver 4210R processor and an NVIDIA RTX3090 (24GB) graphics accelerator card; the software framework was based on Python 3.9 and PyTorch 2.5.1. The global iterations of model training were 300, the batch size was 128, the initial learning rate of the optimizer was 0.0001, and the network model structure parameters are shown in Table 1.

[0148] Table 1. Network parameter settings:

[0149]

[0150] Step 6.2: Evaluate the prediction results using mean absolute error and root mean square error.

[0151] To comprehensively evaluate the model's predictive performance, two metrics are introduced: Mean Absolute Error (MAO) and Root Mean Square Error (RMSE). Both MAO and RMSE measure the deviation between the model's predicted values ​​and the actual values. Both can prevent the cancellation of positive and negative errors; the smaller the MAO and RMSE values, the higher the model's predictive accuracy. The evaluation metrics are as follows:

[0152]

[0153]

[0154] Where n is the total number of test samples, The actual remaining lifetime of the i-th sample. These are predicted values.

[0155] Step 6.3: Analysis of data augmentation results.

[0156] To verify the performance of the feature generation module and evaluate the realism of the synthesized samples, this invention utilizes a t-distributed random neighborhood embedding algorithm for visualization studies, such as... Figure 6 As shown, t-SNE uses nonlinear dimensionality reduction techniques to map complex high-dimensional features to a two-dimensional plane, thus intuitively demonstrating the geometric similarity between the generated distribution and the true distribution. Figure 6The t-SNE visualization results and corresponding probability density curves for real and generated sample features under rolling bearing 1-1 are presented in detail. The results show that although the synthetic data exhibits unavoidable local biases relative to the original data, it maintains a high degree of approximation in the macroscopic distribution topology. This observation demonstrates that the TCCGM model not only accurately learns the regular operating patterns of time series data but also exhibits a high degree of approximation in generating data with deviations.

[0157] Furthermore, a comparison of the probability density curves reveals that the generated features are largely consistent with the real features in terms of peak position and peak width, indicating that the synthesized samples inherit the statistical regularities and trends during the degradation process well. Meanwhile, density differences still exist in a small number of tail regions, possibly due to insufficient modeling caused by the scarcity of samples from extreme operating conditions. Overall, this visualization provides intuitive evidence for the effectiveness of the feature generation module and lays a reliable foundation for sample expansion strategies in subsequent cross-condition migration experiments.

[0158] Step 6.4: Comparative experiment and analysis of different data augmentation methods.

[0159] In this embodiment, the operating condition 1 data from the IEEE PHM Challenge 2012 rolling bearing dataset and the XJTU-SY rolling bearing accelerated degradation dataset are selected as benchmark verification objects. Both datasets contain full-lifecycle vibration signals of rolling bearings from normal operation to complete failure, which are used to verify the effectiveness and generalization ability of the remaining service life prediction method of this invention.

[0160] The PHM2012 dataset was collected using the PRONOSTIA experimental platform. It contains 4 channels of vibration data and 1 channel of temperature signal. The vibration signal sampling frequency is 25.6kHz, and it is collected once every 10 seconds, with 2560 points collected each time.

[0161] The dataset contains three operating conditions: Condition 1 has a speed of 1800 rpm and a load of 4000 N, Condition 2 has a speed of 1650 rpm and a load of 4200 N, and Condition 3 has a speed of 1500 rpm and a load of 5000 N. Condition 1 contains the full life cycle data of seven rolling bearings, from Bearing1-1 to Bearing1-7.

[0162] The XJTU-SY dataset uses LDK UER204 rolling bearings, with single-axis accelerometers installed in the horizontal and vertical directions of the bearing housings under test. This dataset includes three operating conditions, testing five rolling bearings in each condition, for a total of 15 rolling bearings with full life-cycle data. The vibration signal sampling frequency is 25.6 kHz, the sampling interval is 1 min, and each sampling lasts 1.28 s. Each sample contains 32,768 discrete data points. The comparative experiment of different data augmentation methods in step 6.4 uses XJTU-SY condition 1 data, which has a rotational speed of 2100 rpm and a radial load of 12000 N, and includes five rolling bearings from Bearing1-1 to Bearing1-5.

[0163] To quantify the gain effect of introducing diffusion model-generated data on remaining lifetime prediction performance, a systematic comparative study was conducted on the XJTU-SY working condition 1 dataset. The experiment designed comparative verification of four models: data augmentation using GAN

[96] ; data augmentation using VAE-GAN

[97] ; data augmentation using WGAN

[98] ; and data augmentation using the method of this invention. Then, the intrinsic mode function components of the augmented data were extracted and fused with the original time series to construct a multidimensional feature vector, which was then input into the BiGRU-Transformer model for training. The experimental results of each group are shown in Table 2.

[0164] Table 2. Performance comparison of different data augmentation methods on the XJTU-SY dataset:

[0165]

[0166] Comparative experiments based on the XJTU dataset further confirm the advancement of the proposed method. As shown in Table 3, under the same experimental conditions, the data augmentation strategy proposed in this invention achieves the best results in all evaluation metrics. Experimental results show that the traditional GAN ​​model, due to its unstable training process and susceptibility to mode collapse, has the highest prediction error after augmentation, with mean absolute error (MAE) and root mean square error (RMSE) reaching 0.1625 and 0.1871, respectively. In contrast, the WGAN model, which incorporates Wasserstein distance, improves the generation quality to some extent. However, the method proposed in this invention, with its temporal consistency constraints and diffusion generation mechanism, achieves MAE and RMSE of 0.0312 and 0.0443, respectively, significantly outperforming WGAN's 0.0851 and 0.0982. This result confirms the effectiveness of the diffusion model in the data generation task of the critical degradation stage and also shows that by synthesizing high-fidelity vibration signals to complete the training data distribution, the limitation of model generalization ability caused by the scarcity of real failure samples can be effectively alleviated, thereby significantly improving the accuracy of remaining lifetime prediction.

[0167] Step 6.5: Ablation experiment and analysis.

[0168] The ablation experiment in step 6.5 uses the XJTU-SY operating condition 1 data described in step 6.4. Following the leave-one-out method described in step 1.3, all data from one rolling bearing are selected sequentially as the test set, and the data from the remaining four rolling bearings constitute the training set. The training and testing processes are independently repeated 7 times, and the mean of the mean absolute error and the root mean square error is used as the evaluation result of the predictive performance.

[0169] To explore in depth the specific contributions of the proposed TCCGM and generated data evaluation module to the prediction performance of rolling bearing remaining life, this section constructs a systematic ablation experiment scheme. The specific experimental settings are as follows: Scheme 1: Remove TCCGM and the generated data evaluation module, and rely only on the original small number of samples for model training as a performance benchmark; Scheme 2: Use a conventional generative adversarial network (GAN) to replace the TCCGM module, while retaining the generated data evaluation module to verify its advantages; Scheme 3: Use only TCCGM for data augmentation, but remove the generated data evaluation module and do not perform quality screening on the generated samples; The complete method of this invention integrates both TCCGM and the generated data evaluation module. The corresponding experimental results are statistically shown in Table 3. By comparing and analyzing the prediction indicators of the above four schemes, the independent role and synergistic gain effect of each sub-module in the feature enhancement process can be quantitatively analyzed, thereby verifying the rationality of each module.

[0170] Table 3. Performance comparison of ablation experiments on the XJTU-SY dataset:

[0171]

[0172] Experimental results show that Scheme 1, as the baseline model without any data augmentation strategy, has a mean absolute error (MAE) of 0.3455, indicating a significantly high prediction error. This suggests that with insufficient sample size during critical degradation stages, the original training data cannot adequately cover the degradation evolution characteristics of the bearing throughout its entire lifespan. The deep network cannot obtain sufficient effective sample support to learn stable temporal degradation patterns, leading to overfitting during training and ultimately reducing the accuracy of remaining life prediction. Secondly, in the comparison of generative models, Scheme 2, using GAN for data augmentation, shows improvement over the baseline, but its performance is still significantly weaker than Scheme 3, which only includes TCCGM. This is because GAN is prone to training instability and pattern collapse in long-term sequence generation tasks, making it difficult to fully characterize the continuous changing trends and detailed features of the bearing degradation process. TCCGM, with its diffusion mechanism, can more finely reconstruct temporal details, significantly reducing the MAE to 0.0919. Finally, the comparison between Scheme 3 and the method of this invention reveals the key value of the evaluation module. After introducing the generated data evaluation module, the model effectively eliminated inferior samples that did not conform to the mechanistic constraints, causing the root mean square error (RMSE) to plummet from 0.1126 to 0.0443. This strongly demonstrates that the dual strategy of "high-fidelity generation + refined screening" can maximize the benefits of data augmentation and achieve higher prediction accuracy even when data is missing during critical degradation stages. Overall, the complete method proposed in this invention significantly outperforms other variants in all metrics, validating the superiority of its architecture.

[0173] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of rolling bearings based on generative data augmentation, characterized in that, include: Step 1: Acquire and preprocess the rolling bearing vibration signal; Step 2: Perform forward diffusion noise addition and reverse denoising based on temporal consistency constraints on the preprocessed signal to reconstruct supplementary samples for the key degradation stage; Step 3: Based on the KL divergence between real samples and generated samples and a preset threshold, the generated samples are filtered. If the threshold requirement is not met, the generation process in Step 2 is repeated. Step 4: Process the screened samples using empirical mode decomposition and extract multidimensional time-domain statistical features; Step 5: Construct a dual-attention feature fusion network consisting of a bidirectional gated recurrent unit and a Transformer encoder. Input the multidimensional temporal statistical features and fuse the local degradation features output by the bidirectional gated recurrent unit and the global long-range degradation trend output by the Transformer encoder through self-attention, interactive attention and channel splicing. Step 6: Input the fused features into the fully connected prediction layer to complete model training and nonlinear mapping, and output the predicted value of the remaining service life of the rolling bearing.

2. The method for predicting the remaining life of rolling bearings as described in claim 1, characterized in that, The preprocessing in step 1 includes data cleaning, linear normalization, and dividing the dataset into training and test datasets using the leave-one-out-of-random method.

3. The method for predicting the remaining life of rolling bearings as described in claim 1, characterized in that, Step 2 specifically includes: Step 2.1: In the preprocessed signal, the key data target to be generated will be... The associated upstream and downstream process samples serve as conditional information, including upstream and downstream process observations, operating condition information, or degradation status characteristics; Step 2.2: By adding noise through forward diffusion, the conditional information is gradually degraded into pure Gaussian noise. The forward diffusion noise addition process is modeled as a parameterized Markov diffusion chain with a total of K diffusion time steps. At each discrete time step k, 1 ≤ k ≤ K, based on the state of the previous time step... Sample to generate the noisy state at the current moment. Gaussian noise is gradually injected according to the variance scheduling parameters; Step 2.3: By performing inverse denoising based on temporal consistency constraints, the pure Gaussian noise is gradually restored to reconstruct the noise that conforms to the time consistency constraints. The data distribution of features yields complete data containing missing information.

4. The method for predicting the remaining life of rolling bearings as described in claim 3, characterized in that, The network model used in the reverse denoising includes a long short-term memory network, a broadcast layer, multi-head self-attention, and one-dimensional convolution. Modeling the evolutionary dependency of program sequences in long short-term memory networks; By utilizing the location encoding of the broadcast layer to fuse the location embedding of the diffusion time step, and combining it with the process mechanism conditions, the location features and upstream and downstream process conditions jointly participate in the reverse denoising. Multi-head self-attention is used to extract global contextual features and capture global long-range degradation relationships; Local short-term impact features of adjacent sampling points are extracted by one-dimensional convolution, residual fusion is completed, and noise is removed from the vibration signal.

5. The method for predicting the remaining life of rolling bearings as described in claim 4, characterized in that, The Long Short-Term Memory (LSTM) network introduces a gating mechanism, utilizing the synergistic effect of the forget gate, input gate, and output gate to regulate the information flow. The forget gate, based on the hidden layer state of the previous time step and the current input, calculates the weight coefficients within the interval (0,1) using the Sigmoid activation function, determining whether historical information in the cell state of the previous time step should be discarded or retained. The input gate filters the new input at the current time step, generating candidate memory information through the hyperbolic tangent function, and quantifies the importance of the new information to the current state by combining the gating signal of the Sigmoid function. The retention result of the forget gate is merged with the updated content of the input gate to complete the cell state refresh, and according to the control logic of the output gate, the integrated cell state is mapped to the current hidden layer output.

6. The method for predicting the remaining life of rolling bearings as described in claim 4, characterized in that, The multi-head self-attention mechanism first maps the query vector set and the key-value pair vector set to different feature dimension subspaces using multiple independent linear projection matrices. Then, it independently performs scaled dot product attention calculations in each subspace, measuring the relevance of the query vector and key vector by the dot product, introducing a scaling factor to alleviate the gradient minima problem caused by excessive dimensionality, and generating attention weights in the form of a probability distribution through the Softmax function. Finally, it integrates features through a linear transformation layer to achieve effective integration and dimensionality restoration of information from multiple semantic subspaces.

7. The method for predicting the remaining life of rolling bearings as described in claim 4, characterized in that, The reverse denoising network model employs variational inference theory to construct a loss function during training. By maximizing the variational lower bound, the loss function is simplified to the mean square error between the predicted noise and the true Gaussian noise. in, This represents the expected value of the diffusion step number k. This represents network prediction noise; The key vibration signal sequence to be generated; That is, the cumulative retention coefficient is the product of the retention coefficients of each signal from step 1 to step k; By minimizing the objective function to update the network parameters θ, the model can extract effective rolling bearing vibration signal features from arbitrary noise, thus achieving the goal of resolving missing data. The filling and generation.

8. The method for predicting the remaining life of rolling bearings as described in claim 1, characterized in that, Step 4 specifically includes: based on the effective frequency band distribution characteristics of the rolling bearing vibration signal and prior experience in processing similar working condition data, using empirical mode decomposition to decompose the vibration signal into several intrinsic mode function components to cover the main fault characteristic frequency band and suppress the effects of endpoint effects and mode aliasing. The intrinsic mode functions separate the high-frequency noise, fault impact components and low-frequency trend components in the original signal. Then, 11 types of time-domain statistical features are extracted from each intrinsic mode function component, including variance, energy, root mean square, mean, wavelength, median, peak factor, higher-order statistics, entropy, arctangent standard deviation and arcsine standard deviation, and they are arranged into a multidimensional degradation feature sequence according to the sampling time.

9. The method for predicting the remaining life of rolling bearings as described in claim 1, characterized in that, The feature extraction and fusion process in the dual attention feature fusion network specifically includes: Step 5.1: The bidirectional gated recurrent unit processes the input multidimensional time-domain statistical feature sequence through two independent GRU networks in the forward and backward directions respectively: forward propagation captures the degradation and evolution trajectory from the past to the present, generating a forward hidden state that reflects historical information; backward propagation uses the observation data of subsequent time steps to assist the current inference, generating a backward hidden state that contains future context, and after obtaining the bidirectional information of each time step, the forward and backward hidden states are concatenated into a comprehensive hidden state; Step 5.2: Use self-attention to weight the output of the bidirectional gated recurrent unit and form local features; Step 5.3: The Transformer encoder injects absolute positional information into the parallel input feature sequence through positional encoding, making up for the lack of temporal positional sense in the pure attention mechanism. In the interactive attention layer, it uses the query vector in the global feature space and the key-value pairs in the local feature space to perform interactive calculations and output global features. Step 5.4: The local and global features are fused at the channel dimension in the splicing layer. The fused feature vector contains both short-term fault impact information and vibration cycle degradation trend information.

10. A rolling bearing remaining life prediction system based on generative data augmentation, employing the method described in any one of claims 1 to 9, characterized in that, include: Data preprocessing module: used to clean and normalize the acquired raw vibration signals, and to divide them into training and test sets; The data generation module with temporal consistency constraints is used to perform forward diffusion noise addition and reverse denoising based on temporal consistency constraints on the preprocessed signal, and to supplement key missing samples. Data evaluation module: Based on the KL divergence between real samples and generated samples and a preset threshold, the generated samples are screened. Empirical Mode Decomposition Module: Used to decompose the screened samples into intrinsic mode function components and extract time-domain statistical features; Dual Attention Sequence Feature Fusion Module: Employing a bidirectional gated recurrent unit and a Transformer encoder, it fuses local degradation features and global long-range degradation trend features through self-attention, interactive attention, and channel splicing. Fully connected prediction module: Input fused features, output remaining lifetime prediction value.

Images