Bearing residual service life prediction method and system based on SSA and multi-model fusion

Abstract

The invention belongs to the technical field of bearing life prediction and state evaluation, and particularly relates to a bearing residual service life (RUL) interval prediction method based on SSA decomposition and multi-model fusion. According to the method, on the basis of a bearing vibration signal, firstly, original data are preprocessed, noise of different degrees is introduced through a forward process of a diffusion model to enhance data diversity, and the original signal is gradually denoised and recovered by using a U-Net network in a reverse process. Then, a singular spectrum analysis (SSA) algorithm is adopted to decompose the enhanced signal, key feature sequences such as a trend term and a periodic term are extracted, and noise components are filtered out; and respectively inputting the denoised effective components into a Pyraformer model and a TCN (Time Convolutional Network) model to carry out depth feature extraction and time sequence modeling, and fully utilizing the global modeling capability of the Pyraformer and the local dependence capture capability of the TCN. And finally, performing interval fusion prediction on output results of the two models by adopting a Bayesian inference mechanism, and generating a confidence interval of the residual life of the bearing, thereby realizing high-robustness, multi-scale and quantifiable prediction on the running state of the bearing. The method integrates signal enhancement, time-frequency decomposition, deep modeling and uncertainty estimation, has the advantages of high prediction precision, strong anti-noise capability, good interpretability and the like, and is suitable for industrial equipment health management and intelligent maintenance scenes under complex working conditions.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and equipment health management, and particularly to a data-driven method for predicting remaining useful life (RUL). Specifically, this invention relates to a method and system for predicting the remaining useful life of bearings, combining diffusion models, singular spectral analysis (SSA), and multi-model fusion technology, aiming to improve the accuracy and reliability of bearing RUL prediction. This method is widely used in the fields of health monitoring, fault prediction, and intelligent maintenance of industrial equipment, providing a scientific basis for equipment management and maintenance decisions. Background Technology

[0002] With the rapid development of industrial intelligence, equipment health management and predictive maintenance are becoming increasingly important in modern manufacturing. As a core component of mechanical equipment, the performance of bearings directly affects the operating efficiency and safety of the entire system. Bearing Remaining Useful Life (RUL) prediction technology aims to predict the timing of bearing failures by monitoring their operating status, thereby providing a basis for equipment maintenance, reducing downtime and maintenance costs, and improving production efficiency. However, existing bearing RUL prediction methods mainly rely on physical modeling methods or traditional data-driven methods. Physical modeling methods typically require a large amount of prior knowledge and assumptions, and have poor adaptability to complex operating conditions, making it difficult to accurately predict the remaining useful life of bearings. While data-driven methods can learn the degradation patterns of bearings through data, they still face challenges when dealing with complex industrial environments, especially when data noise is high and the sample size is insufficient, often resulting in prediction accuracy that fails to meet practical needs. Traditional data-driven methods mainly rely on extracted artificial features, ignoring the potential complex patterns in the signal. Furthermore, some existing deep learning models are easily affected by data scarcity and noise interference when processing long-term series data, resulting in poor model generalization ability and difficulty in handling complex multi-dimensional features and temporal dependencies. Therefore, how to effectively combine advanced data preprocessing methods and multi-model fusion technology to improve the accuracy and robustness of bearing RUL prediction has become a current technical challenge. Summary of the Invention

[0003] With the increasing demand for industrial intelligence and equipment health management, the prediction of bearing remaining service (RUL) has become particularly important. However, most existing bearing life prediction methods suffer from problems such as data noise interference, scarce samples, and high uncertainty in prediction results, resulting in low prediction accuracy. Therefore, the bearing life prediction industry urgently needs an intelligent prediction method with strong noise resistance, accurate feature extraction, and quantifiable results to overcome the shortcomings of existing technologies.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] This invention proposes a bearing remaining service (RUL) prediction method based on SSA and multi-model fusion. Through multiple data preprocessing, feature extraction, and interval prediction techniques, it can effectively improve the accuracy, robustness, and interpretability of bearing life prediction. The method includes the following steps:

[0006] Step 1: Use the forward process of the diffusion model to add noise perturbation to the original vibration data to enhance data diversity;

[0007] Step 2: Use the U-net neural network to denoise and reconstruct the perturbed data to obtain a high-quality bearing signal;

[0008] Step 3: Use Singular Spectrum Analysis (SSA) to decompose the denoised signal and extract trend and periodic components;

[0009] Step 4: Integrate Pyraformer and TCN models to extract multi-scale features from the signal;

[0010] Step 5: Use the Bayesian interval prediction method to output the bearing remaining life prediction results with confidence.

[0011] Step 1 includes:

[0012] This step uses a diffusion model to simulate the forward noise disturbance process. The original bearing vibration signal is gradually injected with Gaussian noise at preset time steps, creating multiple disturbance versions. The forward process of the diffusion model is as follows:

[0013]

[0014] In each step t, new data samples X are generated by adding noise. t We gradually generate data X from the original data X0. t Until the data is completely covered by noise. X t This is the new data sample generated in step t. α t It is a parameter that gradually decreases within the range [0,1], controlling the degree of noise addition. X t-1 This is the data sample from step t-1, i.e., the data sample from the previous step. ∈ t It is standard Gaussian noise, which follows a standard normal distribution, i.e., a normal distribution with a mean of 0 and a variance of 1. It is to take the data from the previous step X t-1 The scaling factor. It is to include noise ∈ t The scaling factor.

[0015] Diffusion models, while preserving the original structural information, increase data diversity, which helps alleviate the problem of insufficient samples and enhances the model's adaptability to anomalous inputs.

[0016] Furthermore, the U-net network reverse process steps in Step 2 include:

[0017] The generated noise sample X t Input is U-Net. The encoder part of U-Net extracts features from the input samples, performing convolution and downsampling layer by layer. In the reverse process, noise ∈ is predicted step by step through U-Net. θ (X t X is calculated using the reverse process formula. t-1 Stepwise recovery: For each time step from T to 1, noise is removed sequentially to obtain the recovered sample X0. The formula for the reverse process is as follows:

[0018]

[0019] Where X t This is the noise data from the current step. α t It is the attenuation factor, which controls the attenuation of noise. It is the cumulative product of the decay coefficients, representing the decay from time step 1 to t, ∈ θ (X t σ(t) is the current noise component estimated by the neural network. t is the standard deviation related to the noise, used to control for randomness in each sampling step. z is the noise sampled from a standard normal distribution.

[0020] This step inputs the diffused data into the U-net network for structure reconstruction. U-net, through a symmetrical architecture consisting of multi-layer convolutional encoders and decoders, combined with a skip connection mechanism, effectively preserves the local details of the original vibration signal and removes invalid noise, ultimately outputting a high-fidelity vibration sequence, providing an ideal basis for feature decomposition.

[0021] Furthermore, step S3 includes: based on the denoised signal, using Singular Spectral Analysis (SSA) to further decompose the signal to extract trend and periodic terms. SSA helps the model separate useful trend information and periodic features by decomposing the time series into multiple components. The main steps of SSA are as follows:

[0022] The embedding operation transforms the time series data into a trajectory matrix. In this step, the original one-dimensional time series X of length N is... N =(x1,x2,…,x N First, the window is segmented according to a suitable window length M, and then converted into k lagged variables of length M. Finally, a trajectory matrix is ​​constructed in Hankel form.

[0023] Singular value decomposition (SVD) is used to decompose the trajectory matrix into singular values, and the singular values ​​containing the main information are extracted. The specific formula is as follows:

[0024] X=UΣV T

[0025] Where U and V are the left and right singular vectors of the matrix, respectively, and Σ is the singular value.

[0026] In the signal reconstruction step, the trend and periodic terms of the signal are reconstructed based on the selected principal components, noise components are removed, and finally the denoised signal is obtained.

[0027] S4 includes: To better extract features from the bearing signal, Pyraformer and Temporal Convolutional Network (TCN) are used for multi-scale feature extraction. Pyraformer captures global long-term trend information through its pyramid attention mechanism, while TCN focuses on extracting local features. The combination of the two can more comprehensively model the global information and local fluctuations of the signal, as detailed below:

[0028] The input embedding layer in Pyraformer consists of three components: principal components, covariates, and positional encoding. Different features obtained from Singular Spectral Analysis (SSA) decomposition are used as principal components, and covariates (temperature, operating status, timestamps, and other background information) are used as auxiliary components. These are then fused, and after adding positional encoding, the embedding layer maps them to a high-dimensional space. This integrates multi-dimensional features into a high-dimensional space, generating a feature representation suitable for subsequent attention mechanisms. The formula is as follows:

[0029] X embed =Embedding(X input )

[0030] After embedding is complete, Pyraformer's core mechanism—the pyramid attention mechanism—will operate on the time dimension.

[0031] X out =LayerNorm(X+Attention(X))

[0032] After feature extraction via a multi-layer pyramid attention mechanism, the final output layer maps it back to the target space for time series prediction. Typically, the output layer is a linear layer.

[0033] X embed =Linear(X) input )

[0034] Causal convolution is one of the core features of TCN, ensuring that each time step of a time series depends only on past time steps and does not "leak" future information. It guarantees that the output at the current time step t depends only on t and the time steps before it, thus preserving temporal causality. Assuming a kernel size of k, for an input sequence x, causal convolution can be represented as:

[0035] y t =f(x) t ,x t-1 ,...,x t-k+1 )

[0036] Where y t This represents the output at time step t. f represents the convolution operation.

[0037] Dilated convolution expands the receptive field by inserting gaps between convolutional kernels, enabling TCNs to capture dependencies over longer periods. Dilated convolution essentially injects holes into a standard convolution to increase the receptive field. It adds a hyperparameter called the dilation rate, which refers to the number of kernel intervals (in standard CNNs, the dilation rate equals 1). The benefit of dilation is that it increases the receptive field without pooling, allowing each convolutional output to contain a larger range of information. The formula for dilated convolution is:

[0038]

[0039] Feature fusion: The outputs of Pyraformer and TCN are weighted and fused, using the following formula:

[0040] x fused =λ1x Pyraformer +λ2x TCN

[0041] Furthermore, S5 includes: in the final prediction stage, using a Bayesian interval prediction method to perform interval predictions with confidence levels for the remaining useful life (RUL) of the bearing. This method generates prediction results with confidence levels by calculating the posterior distribution, quantifying the uncertainty of the prediction results and providing a reliable basis for decision-making. The Bayesian interval prediction process is as follows:

[0042]

[0043] Where: p(θ) is the prior distribution of parameter θ. p(D|θ) is the likelihood function, representing the probability of observing data D under parameter θ. p(θ|D) is the posterior distribution, representing the probability distribution of parameter θ after observing data D.

[0044] The beneficial effects of this invention are as follows: Compared with the prior art, the improvements of this invention are:

[0045] This invention presents a multi-model fusion-based method for predicting the remaining useful life (RUL) of bearings. This method combines a diffusion model, singular spectral analysis (SSA), and a deep learning model. During the RUL prediction process, a feature extraction model enables deep learning of bearing signal features. Simultaneously, a Bayesian interval prediction method is used to quantify the uncertainty of the prediction results, providing predictions with confidence intervals.

[0046] Compared with existing methods, the present invention has the following advantages:

[0047] Significant results were achieved in data preprocessing: the combination of diffusion model and SSA effectively improved the quality of data, removed noise while retaining effective information, and provided higher quality input data for subsequent feature extraction.

[0048] Multi-model fusion enhances feature extraction capabilities: By fusing Pyraformer and TCN models, it is possible to simultaneously capture global long-term trends and local short-term fluctuations, thereby improving the model's accuracy and robustness.

[0049] The uncertainty of the prediction results can be quantified: By using the Bayesian interval prediction method, this invention can quantify the uncertainty of the prediction results, thereby enhancing the reliability of the prediction results;

[0050] High prediction accuracy and operational efficiency: Validation of the method in this invention on multiple bearing datasets shows that the prediction accuracy can reach 91-95%, while the running speed is fast, the training time is short, and the model calculation efficiency is high, making it suitable for real-time industrial prediction.

[0051] Therefore, this invention provides an efficient, reliable, and easily applicable bearing RUL prediction method that can significantly improve equipment health management, reduce equipment downtime, and lower maintenance costs. Attached Figure Description

[0052] Figure 1 This is an overall technical flowchart of an embodiment of the bearing remaining service life (RUL) range prediction method based on SSA and multi-model fusion of the present invention.

[0053] Figure 2 This is a flowchart of the U-net network structure of an embodiment of the bearing remaining service life (RUL) interval prediction method based on SSA and multi-model fusion according to the present invention.

[0054] Figure 3This is a flowchart of the pyramid model feature extraction process in an embodiment of the bearing remaining service life prediction method based on the improved Attention-GRU model of the present invention.

[0055] Figure 4 This is a flowchart of the feature extraction module of the TCN model in an embodiment of the bearing remaining service life (RUL) interval prediction method based on SSA and multi-model fusion of the present invention. Detailed Implementation

[0056] The following detailed description illustrates the specific implementation method:

[0057] Example 1:

[0058] A method and system for predicting the remaining service life of bearings based on SSA and multi-model fusion, such as Figure 1 As shown, it includes the following steps:

[0059] This embodiment provides a bearing remaining service (RUL) range prediction method based on SSA and multi-model fusion, aiming to improve the accuracy and reliability of bearing life prediction. The method uses the publicly available PHM2012 bearing dataset as the experimental data source and is implemented according to the following procedure:

[0060] First, the data collection phase involves extracting raw vibration signal data from the PHM2012 dataset. To address common issues in bearing life prediction such as insufficient sample size and signal noise interference, a diffusion model is employed to process the raw data. In the forward process of the diffusion model, varying degrees of Gaussian noise are progressively added to the bearing vibration signals, generating diverse training samples. This enhances the diversity of the dataset, improves the model's generalization ability, and alleviates the limitations imposed by practical sampling.

[0061] In the reverse process of the diffusion model, a denoising network based on the U-net structure is used to progressively denoise and restore the noisy data. The U-net network extracts sequence features through encoder and decoder structures and effectively preserves key temporal information by combining a skip connection mechanism, thereby maximizing the restoration of the original signal features and improving the signal quality after data preprocessing.

[0062] After initial denoising, Singular Spectrum Analysis (SSA) is further applied to decompose the denoised vibration signal. Through SSA embedding, singular value decomposition, and reconstruction steps, the signal is decomposed into trend, fluctuation, and noise components. Based on the energy distribution characteristics of the singular spectrum, high-frequency noise components are removed, retaining only the trend and main fluctuation components as input features for subsequent modeling, effectively highlighting the key variation patterns in the signal.

[0063] In the feature extraction stage, the effective feature sequences processed by SSA are input in parallel into two different deep learning models: Pyraformer (Pyramid Attention Model) and Temporal Convolutional Network (TCN). Pyraformer captures global trend information at different time scales through its pyramid attention mechanism, making it suitable for handling complex time-series data with long-term dependencies. TCN, on the other hand, extracts local short-term features through dilated causal convolution, capturing rapid change patterns in vibration signals. By modeling information at different scales, these two models effectively achieve multi-level feature fusion of time-series data.

[0064] After model training, the prediction results from Pyraformer and TCN are fused, and interval prediction is performed based on Bayesian inference. First, a prior distribution is set based on historical data and training output. Then, a posterior distribution is calculated by combining the model's observed output, and the predicted interval for the bearing's remaining life is derived from the posterior distribution. This method not only provides point prediction results but also gives an interval range with confidence (e.g., a 95% confidence interval), thereby effectively quantifying prediction uncertainty and enhancing risk assessment capabilities.

[0065] Finally, based on the actual bearing degradation curves and the predicted interval results, the prediction performance and model reliability are evaluated. Evaluation metrics include interval coverage, interval width, and point prediction error, to verify the effectiveness and robustness of the proposed method under complex working conditions.

[0066] The above methods, through diffusion model enhancement, SSA decomposition and noise reduction, multi-model feature extraction, and Bayesian interval prediction, effectively address the problems of insufficient samples, high noise interference, and prediction reliability of traditional methods, and significantly improve the accuracy and practicality of bearing remaining service life prediction.

[0067] As attached Figure 2 As shown, this embodiment further provides a diffusion process reverse denoising method based on U-net network to improve the quality of bearing vibration signal data and provide a high-quality feature basis for subsequent remaining lifetime (RUL) interval prediction.

[0068] First, in the data preprocessing stage, the original vibration signal is input into the diffusion model for forward diffusion. During this process, Gaussian noise is introduced to generate multiple noise samples at different diffusion steps, increasing data diversity and simulating complex environmental disturbances under actual working conditions.

[0069] When entering the reverse process, the following method is used: Figure 2 The U-net neural network structure shown performs stepwise denoising and reconstruction on diffused noise samples. The U-net network consists of two parts: an encoder and a decoder. It also introduces skip connections to preserve key feature information in the original signal and avoid prediction errors caused by information loss during the denoising process.

[0070] Specifically, the encoder consists of convolution (Conv), ReLU activation function, and max pooling operations to progressively extract deep features from noisy samples and reduce spatial resolution. The decoder, on the other hand, uses upsampling, convolution, and feature concatenation operations to progressively restore the original resolution and temporal scale of the signal, achieving fine-grained feature reconstruction. A skip connection mechanism directly connects the encoder and decoder feature maps between each corresponding layer, ensuring that key information is transmitted and supplemented across multiple scales.

[0071] At each time step t, the formula is reconstructed using the reverse process:

[0072]

[0073] Where X t This is the noise data from the current step. α t It is the attenuation factor, which controls the attenuation of noise. It is the cumulative product of the decay coefficients, representing the decay from time step 1 to t, ∈ θ (X t σ(t) is the current noise component estimated by the neural network. t is the standard deviation related to the noise, used to control for randomness in each sampling step. z is the noise sampled from a standard normal distribution.

[0074] Throughout the reverse diffusion process, the U-net network not only restored the overall shape of the vibration signal, but also accurately preserved key local features and fine-grained changes, effectively improving the stability and accuracy of subsequent feature extraction and modeling.

[0075] In summary, by introducing a back diffusion denoising mechanism based on the U-net network, this embodiment effectively eliminates noise interference while ensuring the diversity of vibration data, improves the robustness and accuracy of the prediction model under actual complex working conditions, and provides a solid data foundation for predicting the remaining service life (RUL) range of bearings.

[0076] As attached Figure 3 As shown, in this embodiment, after extracting key features from the SSA decomposition, the Pyraformer method is used to perform deep feature learning and remaining service life (RUL) prediction on the bearing vibration data.

[0077] The specific steps are as follows:

[0078] First, the trend and periodic terms obtained from SSA decomposition are used as the main input features, and covariates such as external environmental variables, device operating modes, and time steps are used as auxiliary inputs. These features and covariates are then concatenated and positional encoding is added to preserve the sequential structure of the time series. The embedding layer maps and transforms the fused features to obtain the encoded representation.

[0079] X embed =X + PE(X)

[0080] Where PE(X) represents the positional encoding function, and X is the concatenated input feature sequence.

[0081] Next, the coarse-grained structure building module (CSCM) extracts coarse-grained global features over a long time span:

[0082] X cscm =CSCM(X embed )

[0083] The output coarse features are fed into a multi-layer pyramid attention module for multi-scale modeling. Each layer contains a residual connection (Add & Norm), a feedforward network, and a pyramid attention mechanism (PAM) structure to extract temporal dependencies at different scales.

[0084]

[0085] Where l represents the level, PAM represents the pyramid attention layer, FF represents the feedforward layer, and AddNorm represents the normalized residual connection.

[0086] After summing the outputs of all layers, the results are input into the decoder. Based on the prediction requirements, the AttentionDecoder is selected for point prediction, forecasting the RUL value at the current time step. The calculation is as follows:

[0087] a t =softmax(W q ·tanh(W h h t +W e c t ))

[0088]

[0089] Where W is the learned linear weight, h t c represents the current hidden state of the decoder. t For context-aware vectors, This is the final predicted output.

[0090] By combining the global modeling capabilities of the pyramid attention structure with the key feature input after SSA decomposition, this embodiment enhances the model's ability to model complex time series data and improves its predictive stability, thereby increasing the accuracy of bearing remaining life prediction.

[0091] As attached Figure 4 This embodiment proposes a bearing remaining service life (RUL) modeling method based on Temporal Convolutional Network (TCN). This method leverages the powerful temporal modeling capabilities of the TCN network, employing multi-level convolutional feature extraction modules combined with causality, dilation, and residual connection mechanisms to complete deep learning modeling of the time-dependent structure in bearing vibration signals.

[0092] The specific process is as follows:

[0093] First, the input bearing time series features are fed into a series of stacked residual modules. Each module consists of two dilated causal convolution units. After each convolution unit, weight normalization and the ReLU activation function are connected in sequence. Finally, after the two sets of convolution operations, regularization is performed through the Dropout operation to prevent overfitting.

[0094] Causal convolution is used to ensure that the output of the current time step depends only on current and historical time information, without introducing features from future times, thus conforming to the unidirectional dependency structure of time series. Its expression is as follows:

[0095] y t =f(x) t-1,x t-2 ,...,x t-k )

[0096] Where y t Let x represent the output at time t. t-k is the input for a historical moment, and f is the convolution function.

[0097] To expand the model's receptive field, a dilated convolution mechanism is introduced into the structure. Holes are inserted between the convolution kernels to improve the modeling ability for longer time sequence dependencies. The formula is as follows:

[0098] y t =∑w i ·x t-i·d

[0099] Where d is the dilation rate and k is the kernel size, the effective receptive field can be expanded exponentially by increasing the dilation rate layer by layer (e.g., 1, 2, 4, 8).

[0100] Furthermore, to alleviate the gradient vanishing and performance degradation issues caused by deepening the network, a residual connection mechanism is introduced. The input of each residual block is directly added to the output of the convolutional path through a 1×1 convolution or identity mapping path, forming a cross-layer connection and enhancing the network's feature transfer capability and robustness. Its expression is as follows:

[0101] Output=Activation(X+Conv(X))

[0102] Where X is the input feature, Conv(X) is the output of the convolution path, and Output is the output of the fused residual block.

[0103] Through this multi-layered stacked structure, the TCN module can fully extract deep features of bearing vibration data at different time scales, providing a stable, accurate, and time-dependent feature foundation for subsequent life prediction tasks. This structure performs exceptionally well in handling long-sequence modeling tasks, and is particularly suitable for the complex remaining life prediction needs of industrial equipment.

[0104] Example 5: This example further proposes a bearing remaining life (RUL) interval prediction method based on Bayesian inference. On the basis of previous single-point prediction, by establishing a parameter distribution model, the uncertainty of the prediction results is quantified, thereby improving the reliability and robustness of the model.

[0105] This method utilizes Bayesian theory to construct a probability distribution of predicted values ​​within a specific interval by modeling the posterior distribution of parameters and combining it with observed data samples. The specific process is as follows:

[0106] (1) Establishing a prior distribution: In Bayesian statistics, a prior distribution is first established for the parameter, which represents prior knowledge about the parameter before observation data is available. For example, the parameter may follow a normal distribution or a uniform distribution.

[0107] (2) Updating the posterior distribution using data: Using observed data D (model predictions), we can update the posterior distribution of the parameters using Bayes' theorem. Given observed data D, we want to calculate the posterior distribution of the parameter θ using Bayes' theorem. The formula for Bayes' theorem is:

[0108]

[0109] Where: p(D|θ) is the prior distribution of parameter θ. is the likelihood function, representing the probability of observing data D under given parameters. p(θ|D) is the posterior distribution, representing the probability distribution of parameter θ after observing data D.

[0110] (3) Generating the prediction distribution: Once we have obtained the posterior distribution, we can further calculate the prediction distribution. The prediction distribution is the probability distribution of the new predicted value y given the input x and data D:

[0111] p(y|x,D)=∫p(y|x,θ)·p(θ|D)dθ

[0112] This formula represents the distribution of the new predicted value obtained by integrating over all possible values ​​of the parameter θ. Here, p(y|x,θ) is the conditional distribution of the predicted value y given the parameter θ. p(θ|D) is the posterior distribution of the parameter θ given the observed data D.

[0113] (4) Calculate the confidence interval: Based on the prediction distribution p(y|x,θ), we can calculate the confidence interval of the predicted value. For example, for a 95% confidence interval, we can obtain the confidence interval [y|x,θ] by calculating the 2.5% and 97.5% quantiles of the prediction distribution. 2.5％ ,y 97.5％ ].

[0114] This method not only provides single-point prediction results, but also provides the uncertainty range of the model prediction, which improves the stability, interpretability and practicality of the prediction results. It is suitable for data environments with uncertainty in industry, especially under conditions of scarce data or high noise, and can still output reasonable interval estimates, effectively enhancing the reliability of the model.

[0115] In summary, the bearing remaining life prediction method of this invention integrates multi-source information, multi-scale modeling, and uncertainty modeling techniques to form a comprehensive prediction framework that combines signal processing, feature extraction, deep learning, and statistical inference. First, a diffusion-based perturbation method is used to enhance the original vibration signal, followed by denoising and reconstruction using a U-Net neural network, improving data diversity and reconstruction quality. Second, singular spectral analysis (SSA) is introduced to perform trend decomposition on the denoised signal, extracting key feature components (trend and periodic terms), which, together with environmental variables, constitute a multi-source input structure. In the feature learning stage, a Pyraformer module is used for temporal global dependency modeling, while a Temporal Convolutional Network (TCN) is introduced for local feature supplementation extraction, combining the advantages of both to improve modeling accuracy and robustness. Finally, Bayesian inference is used for interval prediction, reasonably characterizing the uncertainty in the prediction results and outputting prediction intervals with confidence levels, providing reliable support for industrial intelligent prediction and operation and maintenance. The above method takes into account the multi-dimensional optimization of feature decomposition, network structure design and prediction strategy, and can effectively improve the accuracy and practicality of bearing remaining life prediction.

Claims

1. A method for predicting the remaining service life range of bearings based on SSA and multi-model fusion, characterized in that, Includes the following steps: S1: Collect the raw vibration signal data during the bearing operation as the input sequence x; S2: Add noise of varying degrees to the original input sequence through a diffusion model forward pass to create diverse noise samples. S3: Using the U-Net network as a denoising model, the diffusion model inverse process is executed to denoise the noisy samples and obtain the restored sample x0. S4: Apply the Singular Spectral Analysis (SSA) algorithm to the recovered sample to decompose it and extract the trend term and periodic term as effective features f. trend ,f periodic ; S5: Input the above key features into the Pyraformer and TCN models respectively, perform dual-model prediction and output the results; S6: The Bayesian interval estimation method is used to perform interval fusion of the prediction outputs of the two models, and finally output the interval prediction results of the remaining life of the bearing.

2. The method according to claim 1, characterized in that: The forward process of the diffusion model is as follows: Among them, X t For the noise sample at step t, α t For time step t, the control coefficient is ∈ θ For predicting the network, σ t is the variance term, and z is the standard Gaussian noise.

3. The method according to claim 1, characterized in that: The denoising module of the reverse process adopts a U-Net network structure, including an encoder, a decoder, and a skip connection module. The encoder extracts multi-scale features step by step through convolution and pooling. The decoder restores the original resolution by upsampling and stitching operations. Skip connections are used to fuse feature maps of different depths.

4. The method according to claim 1, characterized in that: The SSA decomposition module represents a time series x as the sum of multiple principal components: Where, x i The i-th component obtained from SSA decomposition is used to extract the trend term and periodic term as model inputs through reconstruction.

5. The method according to claim 1, characterized in that: The Pyraformer model employs coarse-grained construction and a multi-scale attention mechanism, and its output is calculated as follows: a t =softmax(W q ·tanh(W h h t +W c c t )) Among them, a t For attention weights, W is the output of the Lth layer of Pyraformer. q W h W c W o This is the weight matrix.

6. The method according to claim 1, characterized in that: The TCN model employs dilated convolution and residual connections, utilizing multiple residual blocks to extract local temporal features. Each residual block contains two layers of dilated causal convolution and one layer of 1×1 convolution, with the following residual connection form: Output layer =Activation(X input +Conv(X input )) 7. The method according to claim 1, characterized in that: The prediction results are obtained using Bayesian interval fusion, and the posterior distribution is calculated as follows: The distribution of the predicted output is generated based on the posterior distribution p(θ|D): p(y|x,D)=∫p(y|x,θ)·p(θ|D)dθ And calculate its confidence interval: [y 2.5％ ,y 97.5％ ] is used to express the range of uncertainty in the predicted output.

Images