A bearing residual life prediction method and system oriented to degradation alignment

Abstract

The application discloses a bearing residual life prediction method and system for degradation alignment, and relates to the technical field of bearing fault prediction.The method is characterized in that: firstly, a self-attention driven degradation starting point detection technology is used to automatically divide the health stage and the degradation stage; secondly, a modal alignment module is designed to map the X-axis vibration signal into a time sequence embedding, and meanwhile, the Y-axis vibration signal is used to construct a structured prompt word and generate a prompt word embedding by a large language model, so that a unified representation of the numerical modal and the semantic modal is realized; finally, the large language model performs self-recursive future fragment generation based on the fused sequence, and is trained by a mean square error loss.The health and degradation models are independently trained, and combined prediction is performed according to the stage division during reasoning.The application effectively solves the modal difference problem between the continuous numerical signal and the large language model text Token, and realizes the structured modeling and high-precision residual life prediction of the bearing degradation process.

Description

Technical Field

[0001] This invention relates to the field of bearing failure prediction technology, and in particular to a method and system for predicting the remaining life of bearings for degradation alignment. Background Technology

[0002] As the core support and transmission component of rotating machinery, the stability and reliability of rolling bearings' health status directly affect the operational safety and service life of the entire system. Statistics show that approximately 45% of mechanical failures in rotating machinery originate from rolling bearing failure. Accurately predicting the remaining useful life (RUL) of bearings can provide crucial decision-making basis for condition-based maintenance and predictive maintenance, significantly reducing the risk of unplanned downtime and the total life-cycle maintenance costs. Therefore, bearing RUL prediction has become one of the research hotspots in the field of prognostic and health management (PHM).

[0003] Existing RUL prediction methods can be broadly categorized into three types: physical model-based methods, data-driven methods, and hybrid methods. Physical model-based methods establish state evolution equations based on failure physics mechanisms, offering good interpretability. However, rolling bearings experience complex conditions such as time-varying speeds, varying loads, and boundary lubrication during actual service. Their internal wear, fatigue spalling, and corrosion degradation mechanisms exhibit strong coupling characteristics, making it difficult to construct accurate analytical models of damage evolution. Therefore, the applicability of such methods in engineering applications is limited. Data-driven methods directly extract the mapping relationship between degradation characteristics and lifespan from monitoring data, without relying on prior physical models. In recent years, with the development of sensing technology and deep learning, they have received widespread attention.

[0004] The introduction of deep learning methods has significantly improved the adaptability of RUL prediction. Gated recurrent units, long short-term memory networks, and their variants have demonstrated good performance in capturing the degradation trends of sensor data due to their temporal modeling capabilities. The Transformer architecture, with its self-attention mechanism, has made breakthrough progress in sequence-to-sequence modeling and has been widely used in multivariate time series prediction tasks, resulting in a series of representative works such as iTransformer and PatchTST. However, such models still face several fundamental limitations in practical industrial applications.

[0005] First, the vibration signals of rolling bearings throughout their entire lifecycle exhibit significant non-stationarity and multi-scale degradation characteristics. Bearings typically undergo a continuous service process of several hours to hundreds of hours from healthy operation to failure. The vibration signals show vastly different statistical characteristics at different stages—the initial stage has stable amplitudes dominated by random noise; the middle stage introduces periodic impact components; and the near-failure stage is characterized by a rapidly accelerating degradation rate accompanied by abrupt changes such as amplitude jumps. Existing deep learning models mostly employ a single network structure to uniformly model the entire lifecycle, failing to explicitly distinguish between the stable evolution of the healthy stage and the accelerated deterioration process of the degradation stage, making it difficult for the model to maintain stable prediction accuracy during the rapid degradation stage.

[0006] Secondly, existing models generally model multivariate time-series signals independently in the numerical domain, making it difficult to effectively capture the semantic relationships and cross-variable dependencies between variables from different sensors. Previous research has shown that self-attention mechanisms are inherently permutation-invariant, treating input points as an unordered set and relying solely on positional encoding to inject weakly ordered information. While this structural feature is effective in natural language processing, it presents a potential structural mismatch for time-series data in bearing vibration signals, which have clear causal order and continuity requirements. More critically, the predictive performance of existing deep learning methods largely depends on the stationarity assumption of the benchmark dataset, making them prone to error accumulation and prediction drift in non-stationary real-world operating environments.

[0007] In recent years, Large Language Models (LLMs) have demonstrated powerful contextual understanding and autoregressive generation capabilities in sequence modeling and generation tasks, providing a new technical approach for the analysis of complex industrial time series. Researchers have begun to explore the application of pre-trained LLMs to time series prediction tasks. Representative frameworks such as LLM4TS adapt LLMs to time series prediction scenarios through a two-stage fine-tuning strategy (time alignment stage + prediction fine-tuning stage), fully leveraging the transfer advantages of pre-trained models in cross-domain representation learning. In experiments, it achieved the best ranking in small-sample scenarios. In the fields of fault prediction and health management, the application of LLMs is gradually emerging. Researchers are attempting to use the sequence modeling capabilities of LLMs to automatically extract complex temporal dependencies in sensor data, overcoming the high dependence of traditional deep learning on manual feature engineering and its insufficient generalization ability under multiple operating conditions.

[0008] However, directly applying large language models to bearing RUL prediction still faces significant technical challenges: First, there is the issue of modal discrepancies. Large language models are natively designed for processing discrete text tokens, and there is an inherent modal difference between their word embedding layers and continuous vibration signals. The raw output of a vibration sensor is a continuous sample value with a clear physical dimension and numerical distribution range. Directly quantifying this value and inputting it into a large language model will cause the numerical distribution of the training domain to deviate significantly from the statistical characteristics of the word distribution in the pre-trained text domain, making it difficult for the model to generalize effectively.

[0009] Second, there is the issue of adaptability to non-stationary degradation processes. Bearing degradation trajectories involve multiple stages, including a healthy and stable phase, an initial degradation phase, and an accelerated failure phase. The requirements for prediction accuracy and the focus of attention differ at each stage. Existing methods applying large language models to time series forecasting typically use a uniform autoregressive target to fit the entire sequence, failing to differentiate modeling based on the prediction priorities of the equipment at different operational stages. This can easily lead to insufficient response to key trend changes during the rapid degradation phase.

[0010] Third, there is the issue of the structural organization of the information provided. Bearing monitoring data typically contains multi-channel vibration signals in both horizontal and vertical directions, with clear physical correlations between different channels—for example, amplitude changes in one direction can predict subsequent trend changes in another direction. However, existing time-series large language model methods mostly use a single numerical sequence as input, failing to fully utilize cross-channel semantic guidance information to enhance the model's structured understanding of the degradation process, leading to a rapid amplification of prediction errors during long-term extrapolation.

[0011] In summary, existing bearing RUL prediction methods still have significant shortcomings in terms of characterization of non-stationary degradation processes, modeling depth of cross-variable semantic associations, and modal alignment between large language models and continuous vibration signals. There is an urgent need to develop a new time-series large model prediction framework to achieve structured modeling and high-precision RUL prediction of bearing degradation processes. Summary of the Invention

[0012] To overcome the shortcomings of the prior art, this invention proposes a bearing remaining life prediction method oriented towards degradation alignment.

[0013] To achieve the above objectives, the present invention employs the following technical solution: a method for predicting the remaining life of bearings for degradation alignment, comprising: S1: In the degradation point detection module, the degradation start point is detected in the X-axis vibration signal sequence of the acquired original bearing vibration signal to obtain the final degradation start point C. And based on the final degradation starting point C The entire lifecycle of the original vibration signal sequence is divided into a healthy stage and a degradation stage; S2: In the modal alignment module, the X-axis vibration signal sequence and Y-axis vibration signal sequence in the original vibration signal of the bearing are mapped and inferred respectively to obtain time series embedding and prompt word embedding. Then, modal alignment is performed to obtain the fused multimodal sequence embedding. S3: In the RUL prediction module, RUL prediction is performed based on the fused multimodal sequence embeddings corresponding to the healthy and degenerate stages to obtain the RUL prediction results.

[0014] Preferably, step S1 specifically includes: S11: Smooth the original X-axis vibration signal sequence to obtain a smoothed sequence; S12: Divide the smooth sequence into multiple sliding windows of fixed length, and use a self-attention network to calculate the attention matrix for each sliding window; S13: Extract attention variance, row entropy, sparsity, concentration, and drift distance from the previous sliding window attention matrix based on the attention matrix, and linearly combine the above features to obtain the attention score for each sliding window; S14: Normalize the attention scores of all sliding windows to form an attention score sequence; S15: A strategy combining adaptive thresholding and slope detection is adopted to mine a set of candidate degradation initiation points C from the attention score sequence; S16: From the candidate degradation initiation point set C, determine the final degradation initiation point C based on the selection rules of temporal semantics and engineering heuristics. And based on the final degradation starting point C The entire lifecycle of the original vibration signal sequence is divided into a healthy phase and a degradation phase.

[0015] Preferably, in step S13, the attention score of the sliding window is calculated as follows:

[0016]

[0017]

[0018] Where t is the index of the sliding window, 1≤t≤T; T is the total number of windows, that is, the total number of time points; The first window; This is the T-th window; This refers to the t-th window, which corresponds to the t-th time point. This indicates that a sliding window will be split. This is the original X-axis vibration signal sequence; (.) represents the self-attention mechanism; Let be the attention matrix at time point t; , , as well as These are the first, second, third, and fourth hyperparameters, respectively. (.) represents the variance function; (.) represents the entropy function; (.) represents the concentration function; (.) represents the offset function; Let t be the attention score for the t-th sliding window.

[0019] Preferably, in step S15, mining the candidate degradation initiation point set C from the attention score sequence includes: S151: Attention score sequence Perform sliding window division, constructing a time length of [missing information] at each time point. The attention score history windows are calculated, resulting in a total of Tw attention score history windows. S152: For each historical window of attention scores, construct an adaptive threshold based on a linear combination of the mean and standard deviation within the historical window of attention scores. :

[0020] Where i is the historical window number of the attention score; Let be the mean of the historical attention score for the i-th window; Let be the standard deviation of the i-th attention score history window; , These are the weight parameters; This is the preset minimum value; The adaptive threshold for the i-th attention score history window; S153: Combining the smoothed slope sequence Identify segments with increasing slope; these points are considered candidate degenerate initiation points. Let be the slope at the t-th time point after smoothing the attention score sequence; S154: Subsequently, candidate degradation start points are mined, and a set C of candidate degradation start points is constructed from points that meet the following conditions:

[0021] in, The significance ratio of the slope; is the slope at the i-th time point after smoothing the attention score sequence; d is the minimum distance; The overall standard deviation of the slope; This is the position of the previous candidate point.

[0022] Preferably, the final degradation initiation point C The filtering rules are as follows:

[0023] Where r is the rate of increase; j is the time point number in the candidate degradation start point set C; The current time point in the set of candidate degradation initiation points C; The mean attention score in the next window of length w; This represents the average attention score in the previous window of length w.

[0024] Preferably, step S2 specifically includes: Based on the final degradation initiation point C The X-axis vibration signal sequence is divided into a healthy stage and a deterioration stage. For each stage, it is divided into continuous time segments according to a fixed window, and then mapped to the corresponding time series embedding TE by a time segment encoder. The Y-axis vibration signal sequence is divided into a healthy stage and a degradation stage, which are then embedded into the task instruction template and converted into corresponding prompt words embedded in the PE through a large language model. The time series embedding (TE) generated by the X-axis vibration signal sequence and the Y-axis vibration signal sequence are fused with the prompt word embedding (PE) in a unified semantic space to obtain the fused multimodal sequence embedding.

[0025] Preferably, step S3 specifically includes: The token sequence obtained by fusing the modalities corresponding to the healthy and degenerate stages is used as context and input into the frozen large language model LLM; LLM generates the prediction result embedding for the next time segment through the probability distribution of the next token, and obtains the prediction result embedding for multiple time segments through an autoregressive recursive method. By using time-series projection, the prediction results of multiple time segments are embedded and decoded into the corresponding bearing remaining life prediction results; Based on the bearing remaining life prediction results of multiple time segments and their corresponding real vibration sequences, the bearing remaining life prediction model is trained using mean square error as the supervised loss function.

[0026] Preferably, after obtaining the final degradation starting point in step S1, the method further includes training and testing steps: Training steps: The degradation point detection module automatically divides each full life cycle bearing data into healthy and degradation stages; Sliding window samples were extracted from the healthy stage and the degenerative stage respectively, forming two sets of data subsets with different dynamic characteristics; K-fold cross-validation was performed on the two sets of data subsets respectively, and the health stage prediction model and the degradation stage prediction model were trained independently. Test steps: The degradation start point is detected in the original vibration signal sequence of the bearing under test, and the healthy interval and degradation interval are automatically determined. Input the data within the health interval into the health stage prediction model to obtain the health prediction results; The data within the degradation range is input into the degradation stage prediction model to obtain the degradation prediction results; By combining the health prediction results with the degradation prediction results in the time dimension, a complete future sequence prediction result is formed. Align the complete future sequence prediction results with the actual sequence, and calculate the mean squared error and mean absolute error of the healthy phase, the degenerate phase, and the overall sequence to evaluate the model performance.

[0027] A degradation-aligned bearing remaining life prediction system, used to implement the degradation-aligned bearing remaining life prediction method, includes: The degradation point detection module is used to detect the degradation start point in the X-axis vibration signal sequence of the acquired original bearing vibration signal and obtain the final degradation start point C. ; The modal alignment module is used to perform modal alignment on the X-axis vibration signal sequence and the Y-axis vibration signal sequence in the original vibration signal of the bearing to obtain a fused multimodal sequence embedding. The RUL prediction module is used to perform RUL prediction based on the fused multimodal sequence embedding to obtain the RUL prediction results.

[0028] A readable storage medium having a computer program stored thereon, which, when executed, implements the aforementioned method for predicting the remaining life of a bearing oriented towards degradation alignment.

[0029] The advantages of this invention are: (1) This invention proposes a RUL prediction framework based on Large Language Models for TimeSeries (LLM4TS), which achieves structured modeling and high-precision prediction of bearing degradation processes through degradation initiation point detection, cross-modal alignment, and autoregressive future segment generation. First, a self-attention-driven degradation initiation point detection is constructed, which automatically divides the healthy and degradation stages through dynamic thresholds and trend consistency constraints to obtain stable and reliable stage boundaries. Second, a modal alignment module is designed to map the peak-to-peak values ​​in the x-axis direction to time series embeddings, while using the peak-to-peak values ​​in the y-axis direction with leading-lag information to construct structured cue words, and LLM generates cue word embeddings to achieve a unified representation of numerical and semantic modalities. Finally, LLM performs autoregressive future segment generation based on the fused sequence and is trained using MSE loss. The healthy model and the degradation model are trained independently and combined for prediction during the inference stage according to the detected stage division. Experimental results based on the IEEE PHM Challenge 2012 bearing dataset show that the proposed method significantly outperforms traditional methods, mainstream time series prediction models, and the latest time series large model methods in terms of overall RUL prediction accuracy in single-point prediction and long-term prediction tasks. It effectively avoids the inherent problem that traditional single models cannot simultaneously fit the stable healthy segment and the accelerated degradation segment, and in particular solves the problem of prediction drift and sharp drop in accuracy during the rapid degradation stage.

[0030] (2) This invention proposes an automatic identification method for bearing degradation initiation points based on a self-attention mechanism, achieving adaptive division between healthy and degradation stages. Unlike traditional degradation stage definition methods that rely on thresholds or human experience, this invention uses a lightweight self-attention model to perform attention response analysis on local segments of vibration signals, and combines dynamic thresholds, slope changes, and trend consistency constraints to reliably detect degradation inflection points. This method can automatically distinguish between healthy and degradation stages on a full lifecycle scale, effectively alleviating the modeling challenges of non-stationarity and abrupt changes within the degradation interval, and establishing a stable data foundation for staged prediction.

[0031] (3) This invention constructs a modal alignment framework that integrates time series embedding and prompt word embedding, and utilizes a large language model to achieve generative remaining effective lifetime prediction. Addressing the modal differences between continuous numerical sequences and the discrete token semantic space of the LLM, this invention designs a Time Series Embedding and Prompt Embedding fusion mechanism to achieve unified semantic alignment across modalities. Based on this, the LLM can perform autoregressive future segment generation and project it back into the numerical space for RUL prediction. This framework effectively combines the leading-lag information of covariates, the semantic prior of task instructions, and the powerful contextual modeling capabilities of the LLM, fully leveraging the advantages of large language models in complex temporal reasoning and significantly improving trend prediction performance during the degradation phase.

[0032] (4) This invention transforms the prediction of future time segments into the task of generating semantic tokens. The large language model performs autoregressive next-to-toe token prediction based on the fused multimodal sequence and obtains multi-step future segments through recursive expansion. Subsequently, the temporal projection module can accurately decode the generated semantic vectors into an interpretable vibration amplitude sequence. The entire process only requires the mean square error to constrain the difference between the model output and the real sequence, without the need to design a complex combined loss function. Experimental results show that the overall RUL prediction accuracy of this method is significantly better than traditional methods, mainstream temporal prediction models, and the latest temporal large language model methods, and the reduction in mean absolute error is particularly significant in the rapid degradation stage.

[0033] (5) This invention explicitly embeds the Y-axis vibration signal (covariate) into the prompt word template, requiring the large language model to combine the trend changes of the covariate to correct the prediction direction of the main variable during inference. Traditional methods often process multi-channel signals independently or simply splice them together, ignoring the inherent leading-lag physical relationship between vibration signals in different directions. This design not only improves the model's response sensitivity to sudden changes in operating conditions, but also makes the prediction process more engineering interpretable—that is, the model can "explain" why the subsequent degradation behavior of the X-axis is inferred from the Y-axis leading trend, providing a more reliable decision-making basis for condition-based maintenance. Attached Figure Description

[0034] Figure 1 This is a diagram of the overall architecture of the present invention; Figure 2 This is a schematic diagram of the training and testing process of the present invention; Figure 3 Schematic diagram of a bearing life accelerated aging test platform; Figure 4 This is the result of the short-term experimental life prediction for bearings; Figure 5 A schematic diagram illustrating the differences in the distribution of mean absolute error for predicting the long-term remaining service life of bearings. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Example 1

[0037] Large Language Models (LLMs) have demonstrated strong contextual understanding and autoregressive capabilities in sequence modeling and generation tasks, providing new possibilities for complex time series prediction. However, directly applying LLMs to RUL (remaining bearing life) prediction still faces challenges such as significant modal differences between continuous numerical signals and text tokens, and strong non-stationarity of the degradation process.

[0038] Therefore, as Figures 1-5 As shown, this invention proposes a bearing residual life prediction method oriented towards degradation alignment. The overall model architecture includes: a degradation point detection module, a modal alignment module, and a RUL prediction module; the method specifically is as follows: First, the original vibration signal sequence of the bearing is obtained. The original vibration signal sequence of the bearing includes the peak-to-peak value of the X-axis vibration signal (hereinafter referred to as the X-axis vibration signal sequence) and the peak-to-peak value of the Y-axis vibration signal (hereinafter referred to as the Y-axis vibration signal sequence).

[0039] S1: In the degradation point detection module, the degradation start point is detected in the X-axis vibration signal sequence to obtain the final degradation start point. ; Throughout the entire lifecycle of a bearing, from operation to scrap, its vibration signals typically exhibit a degradation pattern of "long-term stability in the early stage and rapid deterioration in the later stage." However, in actual tasks involving predicting remaining effective service life, the degradation stage often displays significant non-stationarity, abrupt changes, and inconsistent acceleration variations. This makes traditional fixed-window strategies (such as fixed input length and fixed prediction step size) difficult to adapt to the dynamic feature distribution of different stages, often resulting in prediction drift or failure near abrupt change points. To address this issue, this invention introduces degradation point detection based on an attention mechanism, which can automatically identify the inflection points between the healthy and degradation stages on a full lifecycle scale, enabling staged prediction modeling.

[0040] In the task of predicting the remaining effective service life of bearings, given a set of vibration data and Health Indicator (HI) features as input... The goal of the lifetime prediction task is to predict data from the next P steps. .

[0041] The present invention then proposes using a large time-series model to predict the remaining service life of bearings, with the goal of training a mapping function based on a large language model. Predicted data of length P is obtained based on historical data of length h. Representative and Alignment covariates.

[0042]

[0043] in, This represents the learnable parameters during training; P is the number of time steps for prediction; h is the number of historical time steps. This represents the predicted remaining life of the bearing over p future time steps.

[0044] Specifically, it includes: First, the original X-axis vibration signal sequence was analyzed. Optional differential and moving average smoothing processes are performed to obtain a smoothed sequence, which reduces DC components and high-frequency noise and highlights dynamic changes; The smooth sequence is then divided into several sliding windows of fixed length (seq_len), and an attention matrix is ​​calculated for each sliding window using a lightweight self-attention network. .

[0045] Then, various statistical features are extracted from the attention matrix, including attention variance (var), row entropy, sparsity, concentration, and drift distance (offset) from the previous sliding window attention matrix; and the attention score for each sliding window is obtained by linearly combining them according to empirical weights. This compresses the complex attention structure into a single temporal response metric, which is beneficial for subsequent degradation initiation point detection.

[0046]

[0047]

[0048]

[0049] Where t is the index of the corresponding sliding window (1≤t≤T); T is the total number of windows, that is, the total number of time points; The first window; This is the T-th window; This refers to the t-th window, which corresponds to the t-th time point. This indicates that a sliding window will be split. This is the original X-axis vibration signal sequence; (.) represents the self-attention mechanism; Let be the attention matrix at time point t; , , as well as These are the first, second, third, and fourth hyperparameters, respectively. (.) represents the variance function; (.) represents the entropy function; (.) represents the concentration function; (.) represents the offset function; This is the attention score for the t-th sliding window (the window at the t-th time point), which is also the attention score at the t-th time point.

[0050] Then, the attention scores of all sliding windows are normalized. Based on the attention score sequence formed by the normalized attention scores at all time points, a strategy combining adaptive thresholding and slope detection is used to mine candidate degradation initiation points, i.e., the attention score sequence... The sliding window is divided again, and a time length of [value] is constructed at each time point. We calculate a total of Tw attention score history windows. For each attention score history window, we construct an adaptive threshold based on a linear combination of the mean and standard deviation within that window. :

[0051] Where i is the historical window number of the attention score; Let be the mean of the historical attention score for the i-th window; Let be the standard deviation of the i-th attention score history window; , These are the weight parameters; This is the preset minimum value; is the adaptive threshold for the i-th attention score history window.

[0052] And combined with the smoothed slope sequence Look for significant upward segments; these points on upward-sloping segments are considered candidate degenerate initiation points. Let be the slope at the t-th time point after smoothing the attention score sequence.

[0053] Subsequently, local gradient analysis was performed on the candidate degradation initiation points to fine-tune their positions (finding the points where the actual ascent begins within the left window of the candidate points), and the minimum distance was used as the criterion. Constraints prevent the detection of overly dense repeated triggers, thereby improving the stability and interpretability of the detection. The joint triggering condition for the final candidate degradation initiation point is shown in Equation (7):

[0054] in, The significance ratio of the slope; is the slope at the i-th time point after smoothing the attention score sequence; d is the minimum distance; The overall standard deviation of the slope; This is the position of the previous candidate point; This represents the set of all candidate degeneracy initiation points.

[0055] In the set of candidate degradation initiation points In this study, a selection rule based on temporal semantics and engineering heuristics is used to determine the final degradation starting point. Calculate the average energy change by taking a certain window forward and backward after each candidate point, requiring the mean of the later segment to have a sufficient increase compared to the earlier segment. Only if the strategy fails will the application be accepted; if all strategies fail, the application will revert to the earliest point in time in the set as a fallback.

[0056] Where r is the rate of increase; j is the time point number in the candidate degradation start point set; The current time point in the set of candidate degradation initiation points; The mean attention score in the next window of length w; This represents the average attention score in the previous window of length w.

[0057] Finally, based on the final degradation starting point The entire lifecycle of the original vibration signal sequence is divided into two stages: health and degradation, and the data is output for downstream modal alignment and LLM prediction.

[0058] Overall, this detection design significantly improves the robustness of degradation initiation point localization in scenarios with noise interference, abrupt degradation, and different degradation rates by using multi-source attention signal fusion, adaptive thresholding and local gradient fine-tuning, and a candidate degradation initiation point priority strategy for engineering scenarios. This provides a more stable and interpretable input basis for subsequent staged modeling and generative RUL prediction.

[0059] S2: In the modal alignment module, the multidimensional vibration signal sequence is modally aligned to obtain the fused multimodal sequence embedding; After identifying the final degradation initiation point, deep semantic modeling of the vibration signal sequence during the degradation stage is required. Traditional neural networks typically perform regression predictions directly in the original numerical space, while the input structure of large language models (LLMs) is naturally geared towards discrete token sequences, resulting in a significant modal gap between their semantic space and continuous time-series signals. Without appropriate modal mapping and semantic alignment, LLMs cannot effectively utilize their powerful contextual understanding and generation capabilities. Therefore, modal alignment was designed to establish a unified semantic representation space between time-series modalities and LLM token modalities. Specifically, this includes: Two information streams are constructed based on the X-axis vibration signal sequence (X-axis vibration signal peak-to-peak value) and the Y-axis vibration signal sequence (Y-axis vibration signal peak-to-peak value).

[0060] The first path uses the X-axis vibration signal to represent the main (degradation) variable, reflecting the energy evolution and structural changes of the bearing during the degradation stage. Based on the final degradation initiation point, the X-axis vibration signal sequence is divided into a healthy stage and a degradation stage. Then, for each stage, it is divided into continuous time segments according to a fixed window, and mapped to a high-dimensional representation space consistent with LLM embedding through a time-series encoder, forming a time-series embedding (TE) that can be directly processed by LLM.

[0061] The second approach uses the Y-axis vibration signal sequence as a covariate (degradation) variable, which typically exhibits a lead-lag effect in bearing degradation, meaning the covariate's change may precede the main variable's trend breakout, reflecting the main variable's impending trend. To leverage this cross-variable dynamic relationship, the Y-axis vibration signal sequence is divided into healthy and degrading phases, which are then embedded into task instruction templates. This is converted into structured semantic cue embeddings (PEs) using an LLM (Limited Language Management) system, explicitly injecting covariate information and task priors into the LLM during the input phase. It's worth emphasizing that the cue templates are specifically designed to enable the LLM to simultaneously understand the input sequence, covariate characteristics, and prediction requirements, such as... Figure 1 As shown in the prompt template, the prompt words not only explicitly include the Y-axis vibration signal sequence, but also explain to the model the time scale (long-term or short-term) of the current prediction task, the trend attributes to be focused on (stability or trend), and require correction by combining professional domain knowledge, thereby guiding LLM to perform structured reasoning in the input stage.

[0062]

[0063]

[0064] Where h is the time segment number, and a time segment contains several time points; The x-axis vibration signal for the h-th time segment; This represents the Y-axis vibration signal of the h-th time segment; (.) indicates the time segment encoding operation; (.) represents large language model inference; (.) is used for prompt word templates.

[0065] Subsequently, during the modal alignment process, the time series obtained from the X-axis vibration signal sequence was embedded. Embedded with cue words constructed from Y-axis vibration signal sequences The tokens are fused in a unified semantic space to obtain the fused token sequence (the fused multimodal sequence embedding):

[0066] in, This is the token value for the h-th time segment; This represents the total amount of time segments.

[0067] This enables LLMs to simultaneously perceive the numerical dynamics from degenerate signals, leading-lag information from covariates, and semantic constraints from task instructions when receiving input. In this way, continuous numerical time series are reorganized into a "language-like" token sequence, allowing LLMs to capture long-short dependencies, infer trend evolution, and provide a high-quality semantic foundation for the subsequent generation of future segments, leveraging their expertise in autoregressive generation.

[0068] The introduction of the modality alignment module builds a natural bridge between time series modalities and language modalities, effectively utilizes the capabilities of LLM in structured prediction, cross-variable relational reasoning and multi-scale trend modeling, and provides a unified, stable and task-aware semantic input representation for generative RUL prediction.

[0069] S3: In the RUL prediction module, RUL prediction is performed based on the fused multimodal sequence embedding to obtain the RUL prediction result.

[0070] After completing the modal fusion of time series embedding and cue word embedding, an autoregressive next token generation is performed through a large language model, thereby achieving multi-step prediction of future degenerate behavior in a unified semantic space. Unlike traditional neural networks that directly perform regression calculations in the numerical space, this invention follows the generative paradigm of language models: future time segments are regarded as "semantic tokens" to be generated, and the model uses the output of the previous segment as the input for the next step, allowing the prediction process to unfold naturally along the time axis.

[0071] Specifically, the token sequence after modal fusion As context, it is fed into the frozen large language model LLM. Based on its deep Transformer structure, the model generates the next segment prediction embedding through the probability distribution of the next token. The prediction result of the H+1th time segment is embedded into the token sequence after modal fusion. The data is recursively fed back into the model to obtain the prediction embedding for the H+2th time segment; this process is repeated to complete the multi-step forward unfolding. This autoregressive mechanism enables LLM to continuously accumulate states and adjust trend inferences over long time spans, naturally capturing the long-term dependencies, acceleration changes, and cross-stage dynamics inherent in the degenerate sequence.

[0072] Repeat the above steps p times to obtain the prediction results embedding for the next p time segments. .

[0073] To map the high-dimensional semantic tokens output by the LLM back to actual interpretable time-series amplitudes, this invention further introduces time series projection, decoding the generated semantic vectors into bearing remaining life predictions for the corresponding time segments. This projection process establishes an accurate mapping from the linguistic space back to the numerical space, enabling the entire framework to output physically interpretable vibration sequences while maintaining the powerful semantic generation capabilities of the language model.

[0074] This is the predicted bearing remaining life for time segment H+1.

[0075] Based on the bearing remaining life prediction results generated by the model for multiple time segments, the model is aligned point by point with the actual vibration sequence of the corresponding time segment. The mean squared error (MSE) is used as the only supervised loss function to constrain the difference between the future time segments generated by the model and the actual vibration sequence, thereby obtaining the bearing remaining life prediction model.

[0076] In summary, the RUL prediction module, by leveraging a unified semantic time series representation and the next token generation paradigm of LLM, transforms numerical degradation behavior into a generative, inferable, and interpretable token sequence, thereby constructing a generative remaining effective useful life prediction framework that is cross-modal, cross-stage, and has strong generalization capabilities.

[0077] After dividing the model into a healthy stage and a deterioration stage, the model is trained on each stage to obtain prediction results for both stages, thus obtaining prediction models for both the healthy stage and the deterioration stage.

[0078] Degradation point detection is performed on the bearing under test. When degradation point is present, the remaining life is predicted using the degradation stage prediction model; when no degradation point is detected, the remaining life is predicted using the health stage prediction model.

[0079] In summary, in the degradation point detection module of this invention, the X-axis vibration signal is divided into multiple time segments through a sliding window, and its trend and significance features are extracted through a self-attention mechanism to obtain an attention score as a feature representation. The model automatically identifies candidate degradation points based on the changing patterns of the attention response and the overall evolution trend of the health state, and further filters these candidate points using trend consistency constraints, thereby obtaining a stable and reliable degradation initiation position.

[0080] Based on this location, the entire lifecycle signal is divided into two subsequences with different statistical characteristics: a health phase and a degradation phase, and corresponding prediction tasks are constructed for each. The health phase is used to predict the stable evolution trend of the equipment during normal operation; while the degradation phase is used to characterize the accelerated changes during rapid deterioration. By predicting the two phases separately, the model can adaptively learn the corresponding dynamic patterns in different state spaces, thereby achieving continuous and structured prediction of the entire lifecycle.

[0081] Subsequently, in the modal alignment module, the multidimensional peak-to-peak sequence from the original vibration signal is first mapped to the token embedding space of the large language model through the time series encoder to achieve cross-modal projection from numerical time series to semantic representation, thus obtaining the time series embedding.

[0082] Simultaneously, structured prompt embeddings are generated by combining task configuration information and prompt word templates, and prior information such as prediction step size, trend focus, and stage features are encoded in text style and injected into the model. By fusing time segment embeddings and prompt word embeddings in a unified semantic space, a fused multimodal sequence embedding is obtained. The model can simultaneously perceive data-driven degradation dynamics and task-driven prior constraints, thus fully utilizing the strong representational and contextual modeling capabilities of LLM, laying a solid foundation for subsequent generative RUL prediction.

[0083] Finally, in the RUL prediction module, LLM takes the fused multimodal sequence embedding as input and performs autoregressive multi-step generation through its multi-layer self-attention structure to obtain the prediction results for the next step, thereby capturing the long-term dependencies within the degradation stage and its accelerated evolution patterns. The future time segment representation output by LLM is then mapped back to the original numerical space through a time-series projector, achieving interpretable reconstruction from semantic tokens to actual signal amplitudes.

[0084] Based on the gradual evolution trend of the generated sequence, the remaining time required for the device to reach the fault threshold can be inferred, thus obtaining the final RUL estimate. Relying on the unified semantic space provided by modal alignment and the strong sequence generation capability of LLM, it can maintain higher prediction stability and accuracy when facing non-stationary degradation and abrupt signals.

[0085] Overall, the framework fully integrates three key steps: degradation point detection, cross-modal alignment, and RUL prediction, to achieve more stable, reliable, and interpretable RUL prediction of the bearing degradation process.

[0086] Model training and testing

[0087] Based on the overall framework comprised of the above three modules, this invention employs a stage-based dual-model training and testing process to fully leverage the differences between the healthy and deteriorating stages in terms of statistical characteristics, dynamic changes, and trend structures. For example... Figure 2 As shown, firstly, the degradation point detection module is used to process each full-lifecycle bearing data, automatically dividing it into two parts: a healthy stage and a degradation stage. Then, for all sequences in the training set, sliding window samples are extracted from both the healthy and degradation stages, forming two data subsets with different dynamic characteristics. The healthy stage samples reflect a stable, low-noise operating state, while the degradation stage samples contain non-stationary degradation patterns with rapidly increasing energy. To improve generalization performance, K-fold cross-validation is performed on both data subsets, and the healthy and degradation models are trained independently, enabling each model to learn the optimal time-dependent structure and prediction mechanism in its corresponding state space.

[0088] During the testing phase, the process also begins with degradation point detection. The model first automatically determines the locations of healthy and degraded intervals in the test sequence. Then, based on the stage division, the healthy portion is input into the healthy model to obtain healthy prediction results, and the degraded portion is input into the degraded model to obtain degraded prediction results. The two prediction results are concatenated over time to form a complete future sequence prediction. Finally, this complete predicted sequence is aligned with the actual sequence, and the mean squared error (MSE) and mean absolute error (MAE) are calculated for each stage and the overall sequence. This evaluation method not only measures the overall prediction performance but also observes the model's sensitivity in the accelerated degradation interval and its stability in the healthy phase.

[0089] By designing a framework that "identifies stages → segments modeling → dual-model prediction → splicing evaluation", this framework fully utilizes the differences between stages while avoiding the problem of prediction drift or failure that occurs in non-stationary intervals across stages in traditional single models, thereby achieving more reliable prediction of degradation behavior throughout the entire life cycle.

[0090] Experimental verification

[0091] (I) Experiment Description

[0092] This section's experiments are based on the IEEE PHM Challenge 2012 bearing dataset, which consists of bearing life-cycle degradation experimental data collected on the PRONOSTIA test bench under three different operating conditions. For example... Figure 3As shown, the NI DAQcard is the NI data acquisition card, the Pressure regulator is the pressure regulator, the Cylinder Pressure is the cylinder pressure, the Force sensor is the force sensor, the Bearing tested is the bearing under test, the Accelerometers are accelerometers (vibration sensors), the AC Motor is the AC motor, the Speed ​​sensor is the speed sensor, the Speed ​​sensor is the reducer, the Torquemeter is the torque sensor, the Coupling is the coupling, and the Thermocouple is the thermocouple (temperature sensor). The PRONOSTIA test bench consists of a rotation unit, a loading unit, and a measurement unit. The rotation unit consists of an asynchronous motor, a gearbox, and two drive shafts, each driven by a 250W motor to achieve stable and controllable rotational motion. The loading unit accelerates the bearing aging process by applying radial force, while the measurement unit uses sensors to collect vibration and temperature data in real time to characterize the health status of the bearing. The accelerometer has a sampling frequency of 25.6kHz, recording a vibration signal every 10 seconds, with each acquisition lasting 0.1 seconds, resulting in a total of 2560 sampling points. When the vibration signal amplitude reaches 20g, the experiment is considered to have reached the bearing failure point and is terminated.

[0093] Table 1 provides complete information on the IEEE PHM Challenge 2012 bearing dataset. The training and test set partitioning method of this invention is consistent with the official challenge settings, i.e., the first two bearings under each working condition are selected as training data, and the remaining bearings are used for testing. In the experiment, peak-to-peak feature sequences are first extracted from the raw vibration signals collected each time, and these sequences are used as the core input for subsequent degradation point detection, modal alignment, and predictive modeling. To quantitatively evaluate the predictive performance of the model, this invention uses mean squared error (MSE) and mean absolute error (MAE) as the main indicators, comparing the peak-to-peak prediction sequences generated by the model with the actual sequences point by point to comprehensively measure its fitting accuracy and error level.

[0094] Furthermore, this invention, based on different prediction needs and model module designs, conducted three types of experiments to comprehensively verify the effectiveness of the proposed method: (1) Short-term prediction comparison experiment: The prediction setting of "12 points predict 1 point" was used to evaluate the performance of the model in terms of local trend capture and short-term dynamic response; (2) Long-term prediction comparison experiment: The prediction setting of "42 points predicting 6 points" was adopted to test the stability and accuracy of the model in handling long-sequence dependencies and cross-scale trend inference; (3) Ablation experiment: By removing the degradation point detection module and the prompt word embedding module, the contribution of each module to the overall prediction performance is analyzed, thereby verifying the necessity and effectiveness of the proposed framework design.

[0095] Table 1. IEEE PHM Challenge 2012 Bearing Dataset

[0096] (II) Short-term forecast

[0097] In the short-term prediction experiment, the historical input length was set to seq_len=12, the prediction step size was set to predict_len=1, and the time segment length was token_len=1. As a comparison method, this invention selected several representative models, including traditional recurrent neural network models (GRU, LSTM), mainstream deep time series prediction models (iTransformer, PatchTST), and large time series model methods proposed in recent years (Time-LLM, Chronos). In the comparative experiment, the mean squared error (MSE) and mean absolute error (MAE) between the predicted results and the true values ​​were calculated for each test bearing, and the average of the results for each bearing was taken as the final performance index. The prediction fitting results are as follows: Figure 4 As shown, Figure 4 In this context, (a) and (b) refer to the short-term prediction fitting effect of the method of the present invention on the bearing1_3 and bearing2_3 samples on the PHM2012 public dataset, respectively. The groundtruth represents the true life value (i.e. the bearing life measured in actual experiments), and the prediction represents the predicted life value.

[0098] Table 2. Short-term forecast comparison results 1 (MSE)

[0099] Table 3. Short-term forecast comparison results 2 (MSE)

[0100] As shown in Tables 2 and 3, under the short-term prediction experimental settings, the proposed method achieved optimal prediction performance on most tested bearings. In terms of overall average results, compared with the suboptimal model, the method of this invention reduced the mean square error (MSE) by 18.32% and the mean absolute error (MAE) by 8.31%. These results demonstrate that the proposed method has stronger modeling ability and higher prediction accuracy for local variations and instantaneous trends of vibration signals on a short-term scale.

[0101] (III) Long-term forecast

[0102] In the long-term prediction experiment, the historical input length was set to seq_len=42, the prediction step size was set to predict_len=6, and the time segment length was token_len=6. To address the higher requirements of long-term prediction tasks on the model's long-term dependency modeling capabilities, the comparative methods selected in this invention are all representative models designed for long-term time series prediction. This invention selects several representative models, including mainstream deep time series prediction models (iTransformer, PatchTST, Informer, TimesNet), as well as large-scale time series models proposed in recent years (Time-LLM, Chronos). In the comparative experiment, the mean squared error (MSE) and mean absolute error (MAE) between the predicted results and the true values ​​were calculated for each test bearing, and the average of the results for each bearing was taken as the final performance index. Specific results are summarized in Tables 4 and 5.

[0103] Table 4. Long-term forecast comparison results 1 (MSE)

[0104] Table 5. Long-term forecast comparison results 2 (MAE)

[0105] In long-term prediction experiments, the proposed method also demonstrated significant advantages. As shown in the table, the method of this invention outperforms the comparative model in predicting most tested bearings. In terms of overall average results, compared to the suboptimal model, the proposed method reduced the mean squared error (MSE) by 15.03% and the mean absolute error (MAE) by 10.16%. To more intuitively compare long-term prediction performance, this experiment performed kernel density estimation (KDE) analysis on the window-by-window mean absolute error (MAE) distribution of the proposed method and the suboptimal model PatchTST for typical bearings (bearing1_3, bearing2_3) under three operating conditions, highlighting the MAE corresponding to the peak density, as shown below. Figure 5 As shown, Figure 5In the figures (a) and (b), the error distributions of the long-term prediction performance of the proposed method on the bearing1_3 and bearing2_3 samples of the PHM2012 public dataset are shown, respectively. Density represents the probability density, Peak PatchTST represents the peak value of the prediction error distribution of the PatchTST model, and Peak Ours represents the peak value of the prediction error distribution of the proposed method. It can be seen that compared to the suboptimal PatchTST model, the proposed method exhibits higher density at lower MAE values, and the MAE value corresponding to the peak density is also lower. The experimental results demonstrate that the proposed method can more effectively capture the long-term dependencies and accelerated change trends in the bearing degradation process in multi-step prediction scenarios, maintaining stable and accurate prediction performance over long time scales.

[0106] (iv) Ablation test

[0107] To further verify the role and effectiveness of each key module in the proposed method, this invention conducted systematic ablation experiments under both short-term and long-term prediction experimental settings. Specifically, by removing the degradation point partitioning module and the cue word embedding module respectively, different simplified model variants were constructed, and comparative analyses were performed under the same data partitioning and prediction settings. The specific results are summarized in Table 6. The removal of the degradation point partitioning module was used to evaluate the contribution of the staged modeling strategy in addressing degradation non-stationarity; the removal of the cue word embedding module was used to analyze the impact of cross-modal semantic guidance and covariate information on LLM prediction performance.

[0108] Table 6 Ablation Experiment Results

[0109] Ablation experiments further validated the effectiveness of the key modules proposed in this invention. In short-term prediction experiments, compared with the model that removed the cue word embedding module, the complete model reduced the mean squared error (MSE) and mean absolute error (MAE) by 14.49% and 13.61%, respectively; compared with the model that removed the two-stage degradation partitioning module, the MSE and MAE were reduced by 4.69% and 2.54%, respectively. These results indicate that cue word embedding has a more significant effect on improving the model's ability to characterize local dynamic changes in short-term prediction, while the stage partitioning module further enhances the stability of the prediction.

[0110] In long-term prediction experiments, the importance of each module became even more prominent. Compared with the model without cue word embeddings, the complete model achieved performance improvements of 23.13% and 7.96% in MSE and MAE, respectively; compared with the model that removed the two-stage partitioning module, MSE and MAE decreased by 17.66% and 8.94%, respectively. These results demonstrate that in multi-step prediction scenarios, degradation stage partitioning and semantic cue guidance play a crucial role in mitigating long-term error accumulation and capturing accelerated degradation trends. Overall, the ablation experiments fully demonstrate the complementarity and necessity of degradation point partitioning mechanisms and cue word embedding design in short-term and long-term RUL prediction.

[0111] Experimental results based on the IEEE PHM Challenge 2012 bearing dataset show that the method of this invention significantly outperforms traditional methods, mainstream time series prediction models, and the latest time series large model methods in terms of overall RUL prediction accuracy in both single-point prediction and long-term prediction tasks.

[0112] Of course, those skilled in the art will recognize that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0113] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0114] The technologies, shapes, and structures not described in detail in this invention are all known technologies.

Claims

1. A method for predicting the remaining life of bearings for degradation alignment, characterized in that, include: S1: In the degradation point detection module, the degradation start point is detected in the X-axis vibration signal sequence of the acquired original bearing vibration signal to obtain the final degradation start point C. And based on the final degradation starting point C The entire lifecycle of the original vibration signal sequence is divided into a healthy phase and a degradation phase. S2: In the modal alignment module, the X-axis vibration signal sequence and Y-axis vibration signal sequence in the original vibration signal of the bearing are mapped and inferred respectively to obtain time series embedding and prompt word embedding. Then, modal alignment is performed to obtain the fused multimodal sequence embedding. S3: In the RUL prediction module, RUL prediction is performed based on the fused multimodal sequence embeddings corresponding to the healthy and degenerate stages to obtain the RUL prediction results.

2. The bearing remaining life prediction method for degradation alignment as described in claim 1, characterized in that, Step S1 specifically includes: S11: Smooth the original X-axis vibration signal sequence to obtain a smoothed sequence; S12: Divide the smooth sequence into multiple sliding windows of fixed length, and use a self-attention network to calculate the attention matrix for each sliding window; S13: Extract attention variance, row entropy, sparsity, concentration, and drift distance from the previous sliding window attention matrix based on the attention matrix, and linearly combine the above features to obtain the attention score for each sliding window; S14: Normalize the attention scores of all sliding windows to form an attention score sequence; S15: A strategy combining adaptive thresholding and slope detection is used to mine a set of candidate degradation initiation points C from the attention score sequence; S16: From the candidate degradation initiation point set C, determine the final degradation initiation point C based on the selection rules of temporal semantics and engineering heuristics. And based on the final degradation starting point C The entire lifecycle of the original vibration signal sequence is divided into a healthy phase and a degradation phase.

3. The bearing remaining life prediction method for degradation alignment as described in claim 2, characterized in that, In step S13, the attention score of the sliding window is calculated as follows: Where t is the index of the sliding window, 1≤t≤T; T is the total number of windows, that is, the total number of time points; The first window; This is the T-th window; This refers to the t-th window, which corresponds to the t-th time point. This indicates that a sliding window will be split. This is the original X-axis vibration signal sequence; (.) represents the self-attention mechanism; Let be the attention matrix at time point t; , , as well as These are the first, second, third, and fourth hyperparameters, respectively. (.) represents the variance function; (.) represents the entropy function; (.) represents the concentration function; (.) represents the offset function; Let t be the attention score for the t-th sliding window.

4. The bearing remaining life prediction method for degradation alignment as described in claim 3, characterized in that, In step S15, a set C of candidate degradation initiation points is mined from the attention score sequence, including: S151: Attention score sequence Perform sliding window division, constructing a time length of [missing information] at each time point. The attention score history windows are calculated, resulting in a total of Tw attention score history windows. S152: For each historical window of attention scores, construct an adaptive threshold based on a linear combination of the mean and standard deviation within the historical window of attention scores. : Where i is the historical window number of attention score; Let be the mean of the historical attention score for the i-th window; Let be the standard deviation of the i-th attention score history window; , These are the weight parameters; This is the preset minimum value; The adaptive threshold for the i-th attention score history window; S153: Combining the smoothed slope sequence Identify segments with increasing slope; these points are considered candidate degenerate initiation points. Let be the slope at the t-th time point after smoothing the attention score sequence; S154: Subsequently, candidate degradation start points are mined, and a set C of candidate degradation start points is constructed from points that meet the following conditions: in, The significance ratio of the slope; is the slope at the i-th time point after smoothing the attention score sequence; d is the minimum distance; The overall standard deviation of the slope; This is the position of the previous candidate point.

5. The bearing remaining life prediction method for degradation alignment as described in claim 4, characterized in that, Final degradation starting point C The filtering rules are as follows: Where r is the rate of increase; j is the time point number in the candidate degradation start point set C; The current time point in the set of candidate degradation initiation points C; The mean attention score in the next window of length w; This represents the average attention score in the previous window of length w.

6. The bearing remaining life prediction method for degradation alignment as described in claim 1, characterized in that, Step S2 specifically includes: Based on the final degradation initiation point C The X-axis vibration signal sequence is divided into a healthy stage and a deterioration stage. For each stage, it is divided into continuous time segments according to a fixed window, and then mapped to the corresponding time series embedding TE by a time segment encoder. The Y-axis vibration signal sequence is divided into a healthy stage and a degradation stage, which are then embedded into the task instruction template and converted into corresponding prompt words embedded in the PE through a large language model. The time series embedding (TE) generated by the X-axis vibration signal sequence and the Y-axis vibration signal sequence are fused with the prompt word embedding (PE) in a unified semantic space to obtain the fused multimodal sequence embedding.

7. The bearing remaining life prediction method for degradation alignment as described in claim 1, characterized in that, Step S3 specifically includes: The token sequence obtained by fusing the modalities corresponding to the healthy and degenerate stages is used as context and input into the frozen large language model LLM; LLM generates the prediction result embedding for the next time segment through the probability distribution of the next token, and obtains the prediction result embedding for multiple time segments through an autoregressive recursive method. By using time-series projection, the prediction results of multiple time segments are embedded and decoded into the corresponding bearing remaining life prediction results; Based on the bearing remaining life prediction results of multiple time segments and their corresponding real vibration sequences, the bearing remaining life prediction model is trained using mean square error as the supervised loss function.

8. The bearing remaining life prediction method for degradation alignment as described in claim 1, characterized in that, After obtaining the final degradation starting point in step S1, the training and testing steps are also included: Training steps: The degradation point detection module automatically divides each full life cycle bearing data into healthy and degradation stages; Sliding window samples were extracted from the healthy stage and the degenerative stage respectively, forming two sets of data subsets with different dynamic characteristics; K-fold cross-validation was performed on the two sets of data subsets respectively, and the health stage prediction model and the degradation stage prediction model were trained independently. Test steps: The degradation start point is detected in the original vibration signal sequence of the bearing under test, and the healthy interval and degradation interval are automatically determined. Input the data within the health interval into the health stage prediction model to obtain the health prediction results; The data within the degradation range is input into the degradation stage prediction model to obtain the degradation prediction results; By combining the health prediction results with the degradation prediction results in the time dimension, a complete future sequence prediction result is formed. Align the complete future sequence prediction results with the actual sequence, and calculate the mean squared error and mean absolute error of the healthy phase, the degenerate phase, and the overall sequence to evaluate the model performance.

9. A bearing remaining life prediction system for degradation alignment, characterized in that, A method for predicting the remaining life of a bearing oriented towards degradation alignment as described in any one of claims 1-8, comprising: The degradation point detection module is used to detect the degradation start point in the X-axis vibration signal sequence of the acquired original bearing vibration signal and obtain the final degradation start point C. ; The modal alignment module is used to perform modal alignment on the X-axis vibration signal sequence and the Y-axis vibration signal sequence in the original vibration signal of the bearing to obtain a fused multimodal sequence embedding. The RUL prediction module is used to perform RUL prediction based on the fused multimodal sequence embedding to obtain the RUL prediction results.

10. A readable storage medium, characterized in that, It stores a computer program that, when executed, implements a method for predicting the remaining life of a bearing oriented towards degradation alignment as described in any one of claims 1-8.

Images