Novel residual service life prediction model of double-module and variable sliding module based on deep learning

Abstract

The invention discloses a novel deep learning-based dual-module and variable sliding module residual service life prediction model, and relates to the technical field of predictive maintenance of mechanical equipment, in particular to a novel deep learning-based dual-module and variable sliding module residual service life prediction model. The model dynamically samples a full-life-cycle vibration signal through a variable sliding window technology, and the window length is flexibly expanded from a failure starting point to capture a global degradation track; a dual-module mechanism is adopted, a health module extracts health stage features, and a degradation module achieves early remaining service life prediction based on an interaction formula. The model integrates feature extraction, life calculation and hyper-parameter optimization, and training is carried out through an RMSE loss function. The method solves the problems of fixed window information fragmentation and long sequence dependence loss, improves the prediction precision and the working condition generalization capability, and is suitable for multi-stage degradation equipment such as bearings and engines.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance technology for mechanical equipment, specifically a novel deep learning-based model for predicting the remaining service life of a dual-module and variable sliding module. Background Technology

[0002] Industrial equipment maintenance strategies have evolved from reactive maintenance and preventative maintenance to predictive maintenance. Predictive maintenance uses real-time monitoring of equipment data and machine learning methods to predict the remaining lifespan of the equipment, thereby scheduling maintenance tasks at the optimal time. Studies have shown that predictive maintenance can reduce maintenance costs by 20-30% and reduce equipment downtime by 20-50%. However, its application rate in practice is less than 20%, mainly due to the limited accuracy of existing predictive models and their insufficient generalization ability under different operating conditions.

[0003] Remaining useful life prediction methods are mainly divided into direct mapping methods and indirect mapping methods:

[0004] Direct mapping methods: These methods utilize models such as random forests and extreme learning machines to directly establish a mapping relationship between signal data and RUL (Relative Limiting Parameters). However, this method struggles to accurately reflect equipment degradation characteristics when dealing with noisy, nonlinear, and non-stationary signals.

[0005] Indirect mapping methods: These methods extract features related to the degradation process (such as health indicators) and then use a predictor to estimate the RUL. This method has attracted more attention due to its high prediction accuracy, but its performance is highly dependent on the design of the feature extractor.

[0006] In indirect mapping methods, feature extraction is a crucial step. Early methods approximate degradation trajectories using statistical models (such as nonlinear Wiener processes), but these struggle to characterize complex degradation processes. In recent years, deep learning methods, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), have been widely applied to model time series data. However, RNNs are prone to gradient vanishing or exploding problems when processing long-term series, leading to the loss of long-range dependency information; while LSTMs partially improve this problem, they still have limitations.

[0007] Equipment lifecycle data is typically a long-term sequence containing multiple stages. LSTM is prone to losing early information and has difficulty learning the complete degradation trajectory.

[0008] Lifetime data contains both temporal and spatial features, which are difficult to extract simultaneously using a single LSTM model. Summary of the Invention

[0009] The purpose of this invention is to provide a novel deep learning-based dual-module and variable sliding module remaining service life prediction model, which more accurately depicts the degradation process of equipment through a dynamic data processing method and a modular network structure.

[0010] To achieve the above objectives, this invention provides a novel deep learning-based dual-module and variable sliding window (DM-SVW) model for predicting remaining useful life. This model includes acquiring full-lifecycle vibration signals, sampling using a variable sliding window technique, and feature extraction and prediction using a dual-module mechanism. Existing LSTM models are prone to losing early information and struggle to learn complete degradation trajectories. This invention addresses this problem through a dual-module DM-SVW mechanism. The variable sliding window technique dynamically expands the window length (e.g., t2 = (FPT, ..., i)) from the failure point (FPT) of the full-lifecycle data, avoiding information loss in a fixed window and thus capturing the global degradation trajectory. In the dual-module mechanism, the health module extracts features from the health stage (HS) signal, and the degradation module uses these features to predict early remaining useful life. This design embodies the mapping relationship of deep learning models. It combines a hybrid model (TT-DAE) of temporal Transformer (TT) and denoising autoencoder (DAE), which can extract spatial and temporal features simultaneously and improve prediction accuracy.

[0011] Furthermore, regarding the specific implementation of the variable sliding window technique, the entire lifetime data is divided into variable-length segments, and the sampling formula is as follows:

[0012] Where P is the partitioning parameter and S is the total lifespan, the SVW gradually expands from the starting point P to the current point i to form input samples, avoiding information fragmentation caused by the fixed-length truncation of SW. This design enables the model to learn the overall degradation trajectory, rather than local segments, and is especially suitable for multi-stage degradation equipment (such as turbofan engines). The variable length of SVW allows the model to adapt to different operating conditions.

[0013] Furthermore, the interaction of the dual-module mechanism is predicted by the following formula:

[0014] in The interaction between the Health Module (HS) and the Degradation Module (DS) is a key innovation of this invention, which uses the health phase signal as a reference. The health module extracts prior features (such as the initial state of the device) from the HS signal and inputs them into the degradation module to guide the learning process of the DS phase. The dual-module mechanism (DM) combines the HS and DS signals, enabling the model to more accurately identify the degradation initiation point (FPT) and trajectory. The health features serve as a benchmark, and the degradation module calculates the RUL based on them. The interaction in the formula ensures feature fusion and reduces prediction errors.

[0015] Furthermore, the equipment lifespan calculation steps use the formula...

[0016] Where L 10 Here, C represents the rated life, P represents the basic dynamic load, n represents the rotational speed, and a represents the life exponent. Bearing degradation is accelerated by applying a radial force. Parameters C and P in the formula need to be adjusted according to operating conditions. n is adjusted by the frequency converter, and P is measured by a force sensor. The life exponent a is taken as 3 for ball bearings. This formula ensures the reliability of life calculations and provides a benchmark for RUL prediction. This formula is used for turbofan engine life estimation, and combined with a deep learning model, it improves the predictive practicality.

[0017] Furthermore, the formula for calculating the equivalent dynamic load P is as follows:

[0018] P = f p F r , where f p F is the load factor. r For radial force, an electric linear actuator is used to apply the force, and the voltage is adjusted via a laboratory power supply to control the force magnitude. This design ensures the accuracy of the P-value, which in turn affects the lifespan formula L. 10 Reliability.

[0019] Furthermore, the peak acceleration calculation in feature extraction is given by the following formula:

[0020] A peak =max(a n ), 1 < n < N, where a n The acceleration sequence, where N is the number of sampling points, is one of the key features of the vibration signal. Collected by the QM30VT2 sensor, it is used to detect early equipment failures. Peak acceleration reflects the extreme values ​​of the vibration signal. Combined with other features (such as kurtosis and peak factor), it can identify signs of degradation. Peak acceleration, as one of the input features, is denoised using a DAE model to extract spatial features, improving feature quality. The calculation of the maximum value in the formula ensures the capture of sudden vibrations, especially in the early stages of bearing degradation, where a rise in peak value indicates the onset of failure.

[0021] Furthermore, the frequency analysis step uses the Discrete Fourier Transform (DFT) formula. Where Un is the velocity sequence, k>0Hz, and DFT is used to calculate the peak velocity component frequency, revealing the frequency domain characteristics of the vibration signal and helping to identify resonance or fault frequencies. In bearing monitoring, frequency domain features can distinguish between inner and outer ring defects, and combined with time domain features, improve the degradation characterization capability. DFT is applied in the preprocessing stage, and the extracted features are input into the TT-DAE model to reduce noise interference. Complex number calculations in the formula ensure accurate extraction of frequency components, providing rich input for the model.

[0022] Furthermore, the loss function used for model training is the RMSE error:

[0023] in For predicted values, RUL i The true value is used to optimize the parameters of the deep learning model. It improves accuracy by minimizing the difference between the predicted and the true RUL. RMSE is used as the core evaluation metric to ensure the model's generalization ability. The square root calculation in the formula avoids error amplification and is suitable for long-term series prediction.

[0024] Furthermore, the location coding step uses the formula

[0025] Where w j =1 / 10000 2jd Positional encoding, used to embed sequence order information, addresses the lack of natural order awareness in Transformer models. It encodes the position of time step t using sine and cosine functions, enabling the model to capture long-term dependencies. The frequency parameter w in the formula... j It adjusts with dimension d to adapt to different sequence lengths.

[0026] Furthermore, the hyperparameter optimization employs the Hyperband algorithm to minimize the reconstruction.

[0027] in To reconstruct the signal, the feature dimension is reduced after optimization. The hyperparameter combination (such as learning rate and number of network layers) is evaluated by a stepwise halving strategy. Poor-performing configurations are eliminated first to save computing resources. The reconstruction loss measures the reconstruction capability of the denoising autoencoder. Minimizing this loss ensures that the extracted features are highly representative. The feature dimension is reduced after optimization, which improves the efficiency of the model.

[0028] This invention provides a novel deep learning-based model for predicting the remaining useful life of a dual-module and variable sliding module, which has the following advantages:

[0029] 1. Variable sliding window (SW) technology samples vibration signals throughout the entire lifecycle of equipment using a variable-length window. The window length can be dynamically adjusted based on data characteristics, rather than remaining fixed. Existing SW technology divides the entire lifecycle data into fixed-length segments, causing the model to be unable to learn the complete degradation trajectory, especially in long-term series where early information is easily lost. This invention uses SVW technology to represent the input signal as... The time point can be flexibly extended from the failure point of time (FPT) to the current point, thereby capturing global dependencies from health to failure. Specifically, Where P is the partitioning parameter and i is the dynamic endpoint, this allows the model to handle variable-length data at different degradation stages, avoiding the loss of key degradation information due to fixed window cutting. Equipment lifecycle data usually includes multiple stages (such as health period, degradation period, and failure period). SVW technology ensures that the model learns the overall degradation pattern through adaptive window length, thereby improving the accuracy and stability of RUL prediction. This technology also enhances the model's generalization ability under different operating conditions, because the variable window can better adapt to changes in data distribution and reduce the risk of overfitting.

[0030] 2. The dual-module mechanism includes a health module and a degradation module. The health module specifically extracts features of the device's health stage (HS), while the degradation module predicts the remaining lifespan based on health features. The two modules work together through an interactive formula. Information fusion is achieved. Section 3.2 of Document 2 points out that existing indirect mapping methods often ignore health stage information or fail to effectively integrate it into degradation learning, making it difficult for the model to capture the continuity of the degradation trajectory. This invention, through the DM mechanism, first uses the health module to extract prior features (such as features of the stable operating state) from the HS signal, and then uses these features as input to the degradation module, enabling the model to compare the differences between the healthy and degradation states, thereby more accurately identifying the degradation start point and trend. This mechanism is particularly suitable for multi-stage degradation scenarios. The combination of the dual-module mechanism and SVW technology (DM-SVW) further enhances global trajectory learning, enabling the model to make high-precision predictions in the early stages of degradation, reducing maintenance costs and equipment downtime.

[0031] 3. Feature extraction techniques, including frequency analysis (such as Discrete Fourier Transform). ) and location encoding (e.g. Existing methods, such as LSTM or statistical models, struggle to simultaneously handle temporal and spatial features and are susceptible to noise interference. This invention employs multi-dimensional feature extraction, combining time-domain (e.g., peak acceleration, root mean square value) and frequency-domain (e.g., peak velocity component frequency) analysis to comprehensively characterize equipment degradation. A denoising autoencoder (DAE), as part of the feature extractor, introduces a noise layer to reconstruct the signal, reducing the impact of white noise and achieving a reconstruction loss as low as 0.029, thus improving feature quality. This comprehensive approach not only enhances the discriminative power of features but also strengthens the model's adaptability to non-stationary signals, providing reliable input for RUL prediction.

[0032] 4. Model training details, including the loss function (e.g., ... Existing deep learning models often suffer from large prediction errors due to training instability or overfitting. This invention directly minimizes the deviation between predicted and true values ​​using the RMSE loss function, ensuring consistency between training objectives and evaluation metrics. Simultaneously, it employs the Hyperband algorithm for hyperparameter optimization, efficiently searching for optimal configurations (such as learning rate and number of network layers) through a stepwise halving strategy, avoiding the high computational cost of grid search. Normalization further stabilizes the training process, enhancing the model's generalization ability across different datasets. This optimization strategy not only accelerates convergence but also reduces dependence on hardware resources, making it suitable for real-time industrial prediction scenarios.

[0033] 5. The overall architecture (a synthesis of claims 1-10) embodies an end-to-end RUL prediction solution, combining data acquisition, preprocessing, feature extraction, model training, and optimization. Existing methods, such as single LSTM or Transformer models, struggle to simultaneously consider spatiotemporal features and rely on precise stage division. This invention constructs a hybrid model TT-DAE by integrating a variable sliding window, dual modules, DAE, and a Transformer encoder-decoder (temporal Transformer, TT), achieving the fusion of spatial features (extracted by DAE) and temporal features (extracted by TT). This innovative architecture not only overcomes the limitations of existing technologies (such as vanishing gradients or missing spatial features) but also improves the model's practicality in complex industrial environments, potentially breaking through the 20% application bottleneck in predictive maintenance and reducing maintenance costs by 20-30%. Attached Figure Description

[0034] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of the indirect mapping method framework for deep learning in this invention;

[0036] Figure 2 This is a schematic diagram of the timing Transformer of the present invention;

[0037] Figure 3 This is a cross-sectional view of the bearing of the present invention;

[0038] Figure 4 This is a schematic diagram of the aircraft engine sensor data of the present invention;

[0039] Figure 5 This is a schematic diagram of the five-fold cross-validation method of the present invention. Detailed Implementation

[0040] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0042] Example 1: Implementation of a Remaining Useful Life Prediction Model Based on Variable Sliding Window Technology and Data Sampling

[0043] This embodiment mainly corresponds to claims 1 and 2, and relates to the specific application of variable sliding window technology. Variable sliding window technology is used to sample vibration signals throughout the entire life cycle of equipment, avoiding information loss caused by a fixed window, thereby capturing the global degradation trajectory. In implementation, firstly, the entire life cycle data of the equipment (such as a bearing or aero-engine) is collected, including vibration signals, temperature signals, etc. The data acquisition frequency is set to record once every 5 seconds, for a total of 10240 sampling points, ensuring coverage of the complete trajectory from the healthy stage to the failure stage. The input signal of the variable sliding window is represented as t2 = (FPT,...,i), where FPT is the failure start point and i is the current time point. The window length is variable, depending on the equipment... The degradation rate is dynamically adjusted: for example, in the early stages of degradation, the window is shorter to capture subtle changes; in the later stages of degradation, the window is longer to integrate global information.

[0044] The output is the RUL value, calculated using the following formula:

[0045] The implementation steps include data preprocessing: first, the raw signal is cleaned to remove constant signals; then, maximum-minimum normalization is used to scale the data to the 0-1 range. Data sampling employs a sliding variable-length window (SVW) technique, dividing the entire lifetime data into variable-length segments. Where t4 = (P, P+1, ..., i), P is the partitioning parameter, and S is the total lifespan. For example, in a bearing dataset, P can be set as the failure starting point FPT, and the window can be gradually expanded from FPT to the current point i, avoiding the limitations of a fixed window.

[0046] The advantage of this embodiment lies in its ability to effectively learn complete degenerate trajectories, thereby improving prediction accuracy. Compared to the traditional sliding window (SW) method, the SVW technique reduces RMSE error by approximately 36%, performing particularly well when processing long-term series. During implementation, attention should be paid to hyperparameter optimization, such as adjusting the window length P using the Hyperband algorithm to minimize reconstruction loss.

[0047] Example 2: Implementation of the interaction between health and degradation characteristics of the dual-module mechanism

[0048] This embodiment corresponds to claim 3, focusing on the implementation of the dual-module mechanism, including the interaction between the health module and the degradation module. The health module is responsible for extracting features of the device's health stage (HS), and the degradation module predicts the remaining lifespan based on these health features. In implementation, the device's entire lifespan data is first segmented: health stage signals... (where t3 = (1, 2, ..., FPT-1)) and degradation stage signal (Where t4 = (P, P+1, ..., i)). The health module uses a denoising autoencoder (DAE) to extract HS features, eliminate noise, and reduce dimensionality; the degradation module uses a temporal Transformer (TT) to learn the long-term dependencies of DS features. The prediction formula is as follows: Among them, health characteristics are used as prior information input into the degradation module.

[0049] The implementation steps include: after data preprocessing, feature extraction is performed on the HS signal, for example, using the DAE model, whose reconstruction loss calculation formula is as follows: Hyperparameters are optimized using Hyperband (e.g., 3 hidden layers and 180 neurons). Then, healthy features and degenerative signals are fused and input into the TT model for encoding-decoding. Positional encoding is used to embed sequence order information, with the following formula: Where w j =1 / 10000 2jd .

[0050] The TT-DAE model, combined with DM-SVW technology, reduces the RMSE in bearing prediction to 0.158, a 4.8% improvement compared to a single module. During implementation, attention must be paid to the accuracy of health phase data, such as determining the failure initiation point (FPT) (based on acceleration RMS values ​​reaching the 3σ standard). The advantage lies in the dual-module approach, which avoids early information loss and is particularly suitable for equipment undergoing multi-stage degradation.

[0051] Example 3: Equipment Life Calculation and Implementation of Equivalent Dynamic Load

[0052] This embodiment, based on claims 4 and 5, describes the specific application of the equipment life calculation steps and the equivalent dynamic load formula. The equipment life calculation uses the following formula. Where L 10 Where C is the rated life (90% reliability), P is the basic dynamic load, n is the rotational speed, and a is the life index (3 for ball bearings). The formula for calculating the equivalent dynamic load P is P = f p F r , where f p F is the load factor (1.0-1.2 for no impact). r It is a radial force.

[0053] In practice, taking a bearing as an example: First, obtain the parameters from the specification sheet, such as the SKF 61702-2RS1 bearing, with a bore diameter d = 15mm, outer dimensions D = 21mm, and basic dynamic load C = 0.527kN. Then, apply a radial force through the test bench and measure the force using a force sensor. r And adjust f according to working conditions p For example, in the accelerated degradation experiment, the rotational speed n was set to 1500 r / min, and the radial force F... r If the value is 0.5 kN, then P is calculated as P = 1.1 × 0.5 = 0.55 kN. Substituting this into the lifespan formula, we get approximately 95 hours, which can be used as a lifespan baseline.

[0054] The implementation steps include data acquisition: recording vibration signals using a QM30VT2 sensor and extracting peak acceleration A. peak =max(a n ) and peak speed V peak =max(u n This is used to verify the impact of load. The calculated lifespan matches the actual degradation curve with an error of less than 10%. The advantage of this embodiment is that it combines theoretical lifespan with real-time data, providing a calibration basis for the prediction model. During implementation, attention should be paid to changes in operating conditions, such as speed or load fluctuations, and the accuracy should be ensured by dynamically adjusting the P-value.

[0055] Example 4: Implementation of Feature Extraction and Model Training Loss Function

[0056] This embodiment corresponds to claims 7 and 8, and relates to the specific application of feature extraction (such as frequency analysis) and model training loss functions. Feature extraction includes time-domain and frequency-domain analysis: frequency components are calculated using the Discrete Fourier Transform (DFT), as shown in the formula... (Where Un is the velocity sequence), used to identify the frequency of peak components. Simultaneously, statistical features such as kurtosis, peak factor, and RMS value are extracted; for example, the kurtosis formula is... Peak factor is A peak / A rms .

[0057] Implementation steps: First, the vibration signal is preprocessed, and features are screened using a hybrid evaluation metric (monotonicity, trend, predictability), eliminating low-scoring features (such as F5-F8). Then, DAE is applied for spatial feature extraction, reducing the dimensionality to representative features. Model training uses a loss function. in For predicted values, RUL i These are the true values. Training was performed using the Adam optimizer with a learning rate of 1e-3, a batch size of 512, and an early stopping mechanism to prevent overfitting.

[0058] On aero-engine data, the TT-DAE model achieved an RMSE of 9.16, outperforming LSTM and Transformer variants. In implementation, hyperparameter optimization employed the Hyperband algorithm, randomly searching for 500 combinations, reducing computation time by 66.6%. Its advantage lies in effective noise reduction during feature extraction, enhancing model robustness.

[0059] Example 5: Implementation of Position Encoding and Hyperparameter Optimization

[0060] This embodiment, based on claims 9 and 10, describes the detailed steps of positional encoding and hyperparameter optimization. Positional encoding is used to embed sequence order information, and the formula is as follows: Where w j =1 / 10000 2jd d represents the encoding dimension. In implementation, the position encoding vector Pt∈R is... d With input signal This combination ensures that the Transformer model retains temporal characteristics when processing time series data.

[0061] Hyperparameter optimization employs the Hyperband algorithm, with the goal of minimizing the reconstruction loss. Optimized parameters included learning rate, batch size, number of network layers, and number of attention heads. The hyperparameters of the DAE were configured using Hyperband search: the number of hidden layers was optimized to 3, the number of neurons to 180, the dropout rate to 0.3, and the learning rate to 1e-3. After optimization, the reconstruction loss decreased from 0.465 to 0.047, the feature dimensionality was reduced, and training efficiency was improved.

[0062] Implementation steps: First, k-fold cross-validation (k=5) was performed, with the data divided into training, validation, and test sets in a 60:20:20 ratio. Then, a Hyperband stepwise halving strategy was used to evaluate hyperparameter combinations, allocating more resources to the best-performing configurations. After optimization, TT-DAE achieved an RMSE of 0.158 on the bearing data, a 4.8% improvement over the unoptimized model. The advantage lies in balancing computational efficiency and performance, and avoiding local optima.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A novel deep learning-based dual-module and variable sliding module remaining useful life prediction model, characterized in that: The method comprises the following steps: Collecting vibration signals of the equipment in the whole life cycle; Using a variable sliding window technology to sample the signals, the window length being variable to capture the global degradation trajectory; Using a double module mechanism, a health module extracts health stage features, and a degradation module predicts the remaining life based on the health features, so that early prediction of the remaining useful life can be realized; Wherein, the input signal of the variable sliding window is expressed as FPT is the failure starting point, the output is the RUL value, and the calculation formula is:

2. The novel deep learning-based dual-module and variable sliding module remaining useful life prediction model according to claim 1, characterized in that: In the variable sliding window technology, the whole life data is divided into variable length sections to avoid information loss caused by fixed windows, and the sampling formula is: Where S is the total life cycle, and P is the division parameter.

3. The novel deep learning-based dual-module and variable sliding module remaining useful life prediction model according to claim 1, characterized in that: In the double module mechanism, the health module and the degradation module interact, and the prediction formula is: wherein, is a healthy phase signal.

4. The novel deep learning-based dual-module and variable sliding module remaining useful life prediction model according to claim 1, characterized in that: The method further comprises a step of using a life calculation formula in the design of an accelerated test platform, and the formula is: where L 10 is the rated life, C is the basic dynamic load, P is the equivalent dynamic load, n is the rotational speed, and a is the life exponent. This formula is used to estimate the reference life of a bearing or engine to guide the experimental parameter settings.

5. The novel deep learning-based dual-module and variable sliding module remaining useful life prediction model according to claim 4, characterized in that: The calculation formula of the equivalent dynamic load P is: P = f p F r where f p is the load factor, F r is the radial force, which calculation is used to determine the load conditions in the acceleration test platform to simulate the actual working conditions to provide input for the life formula.

6. The novel deep learning-based dual-module and variable sliding module remaining useful life prediction model according to claim 1, characterized in that: The feature extraction further comprises peak acceleration calculation, and the formula is: A peak = max(a n ), 1 < n < N where a n is the acceleration sequence, N is the number of sampling points, and this calculation is used to extract key time-domain features from the vibration signal to identify signs of device degradation.

7. The novel deep learning-based dual-module and variable sliding module remaining useful life prediction model according to claim 1, characterized in that: The feature extraction further comprises frequency analysis, and the discrete Fourier transform is used: Where Un is the speed sequence.

8. The novel deep learning-based dual-module and variable sliding module remaining useful life prediction model according to claim 1, characterized in that: The model training uses a loss function: wherein, RUL i RUL i is the true value.

9. The novel deep learning-based dual-module and variable sliding module remaining useful life prediction model according to claim 1, characterized in that: The method further comprises a position encoding step, and the formula is: where w j = 1 / 10000 2jd , for embedding sequence order information.

10. The novel deep learning-based dual-module and variable sliding module remaining useful life prediction model according to claim 1, characterized in that: The hyperparameter optimization uses a Hyperband algorithm to minimize the reconstruction loss: wherein For the reconstructed signal, the optimized feature dimension is reduced.

Images