A Method and System for Predicting the Remaining Service Life of Aero-engines Based on Layered Latent Space Diffusion Neural Field

Abstract

This invention discloses a method and system for predicting the remaining service life of aero-engines based on a hierarchical latent space diffusion neural field, belonging to the field of fault prediction technology. The method includes: acquiring and preprocessing engine observation data and operating condition information; extracting latent characteristics and hierarchical stage and detail codes through a multimodal encoder; generating multiple sets of future latent variables reflecting different degradation modes in the latent space using a hierarchical diffusion model; decoding the latent variables into continuous-time health indicators using a continuous neural field decoder, forming multiple degradation trajectories that conform to physical priors; and finally, calculating the time when the trajectory first crosses the failure threshold, statistically obtaining the probability distribution of the remaining service life, and performing uncertainty decomposition. This invention can generate multimodal, high-fidelity future degradation scenarios, providing rigorous and interpretable probabilistic predictions, and providing accurate decision support for condition-based maintenance and health management of aero-engines.

Description

Technical Field

[0001] This invention relates to the field of aero-engine health management and remaining service life estimation technology, and in particular to a method and system for predicting the remaining service life of aero-engines based on a hierarchical latent space diffusion neural field. Background Technology

[0002] As high-value and complex equipment, the accurate prediction of the Remaining Useful Life (RUL) of aero engines is crucial for ensuring flight safety and reducing maintenance costs. RUL prediction aims to forecast the remaining time an engine has left from the present moment until a failure occurs, based on current and historical monitoring data. Accurate RUL prediction provides critical information for decisions such as condition-based maintenance and spare parts management, preventing major safety accidents and economic losses.

[0003] Existing RUL prediction methods for aero-engines are mainly divided into physical model-based methods and data-driven methods. Among them, data-driven methods, represented by deep learning, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and their variants, are widely used due to their powerful feature learning capabilities. However, most existing methods treat RUL prediction as a regression problem, directly outputting a point estimate or a simple confidence interval. These methods have the following inherent drawbacks: First, they struggle to model and represent the multimodal paths that may exist during engine degradation (e.g., different scenarios such as rapid or slow degradation due to different failure mechanisms); second, their prediction results often lack rigorous and intuitive probabilistic interpretations, failing to provide decision-makers with a complete picture of "multiple possible degradation scenarios and their probabilities"; finally, many methods use autoregressive approaches for long-term prediction, leading to error accumulation problems, and the generated results may not conform to the physical laws of the degradation process (e.g., health indicators should decrease monotonically).

[0004] In recent years, generative models such as diffusion models have achieved great success in image and speech synthesis. Some studies have attempted to apply diffusion models to time series prediction, but this is typically done in high-dimensional raw data spaces, resulting in high computational costs, and the long-term consistency and physical plausibility of the generated sequences are difficult to guarantee. How to combine the powerful expressive capabilities of generative models with the probabilistic and interpretable requirements of RUL prediction, especially in complex and high-value equipment such as aero-engines, has become a pressing technical problem. Summary of the Invention

[0005] The purpose of this invention is to at least solve one of the technical problems existing in the prior art, and to provide a method and system for predicting the remaining service life of aero-engines based on a hierarchical latent space diffusion neural field that can generate multimodal degradation trajectories, provide rigorous probabilistic predictions and conform to physical laws.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, this invention proposes a method for predicting the remaining service life of an aero-engine based on a layered latent space diffusion neural field, comprising the following steps:

[0008] Acquire the observation sequence and operating condition information of the aero-engine, and preprocess the observation sequence;

[0009] The preprocessed observation sequence and operating condition information are input into the multimodal encoder to obtain the latent characterization, and the stage code representing the macroscopic degradation mode and the detail code representing the microscopic fluctuation are extracted.

[0010] Using the stage code and detail code as conditions, hierarchical diffusion generation is performed in the latent space to obtain multiple sets of future latent variable pairs that correspond to the current state;

[0011] The future latent variable pairs, latent representations, working condition information and continuous future time coordinates are input into the continuous neural field decoder to decode and generate multiple future health indicator degradation trajectories.

[0012] Based on a preset failure threshold, the time when the health index of each degradation trajectory first crosses the failure threshold is calculated to obtain a sample set of remaining lifespan.

[0013] Perform statistical analysis on the remaining lifespan sample set and output the probability distribution of the remaining lifespan.

[0014] Furthermore, after outputting the probability distribution, the method further includes:

[0015] The uncertainty of the probability distribution of the remaining useful life is decomposed to obtain data uncertainty, model uncertainty and mode uncertainty;

[0016] The data uncertainty is characterized by the variance of trajectories generated by multiple samplings under the same diffusion model; the model uncertainty is characterized by the prediction variance of different initialization models; and the mode uncertainty is characterized by the distribution of probability weights of each degradation stage mode corresponding to the stage code.

[0017] Secondly, the present invention provides an aero-engine remaining service life prediction system based on a hierarchical latent space diffusion neural field, comprising:

[0018] The data preprocessing module is used to acquire the observation sequence and operating condition information of the aero-engine, and to preprocess the observation sequence.

[0019] The multimodal coding module is used to encode the preprocessed observation sequence and operating condition information into latent representations, and extract the stage code representing macroscopic degradation patterns and the detail code representing microscopic fluctuations.

[0020] The hierarchical diffusion generation module is used to perform hierarchical diffusion generation in the latent space, based on the stage code and detail code, to obtain multiple sets of future latent variable pairs that correspond to the current state;

[0021] The continuous neural field decoding module is used to decode the future latent variable pairs, latent representations, working condition information and continuous future time coordinates to generate multiple future health indicator degradation trajectories;

[0022] The remaining service life inference module is used to calculate the time when the health index of each degradation trajectory first crosses the failure threshold based on a preset failure threshold, obtain a remaining service life sample set, perform statistical analysis on the remaining service life sample set, and determine the probability distribution of the remaining service life of the aero-engine.

[0023] Furthermore, the system also includes:

[0024] The uncertainty decomposition module is used to decompose the uncertainty of the probability distribution of the remaining useful life after the remaining useful life inference module outputs the probability distribution, and obtain the data uncertainty, model uncertainty and mode uncertainty.

[0025] Beneficial effects:

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] This invention, through the design of hierarchical latent variable stage codes and detail codes and their diffusion generation in the latent space, can naturally and efficiently generate multiple future degradation trajectories reflecting different failure modes, providing rich scenario analysis basis for decision-making. Each generated degradation trajectory corresponds to a possible future; by statistically analyzing the time when these trajectories first cross the failure threshold, a RUL distribution with clear probabilistic significance can be directly obtained. Simultaneously, the stage codes and other latent variables provide a direct explanation of the degradation modes, exhibiting rigorous probabilistic prediction and high interpretability.

[0028] This invention introduces structured priors such as monotonicity into the training of the continuous neural field decoder, ensuring that the generated degradation trajectory conforms to the fundamental physical laws of irreversible degradation of engine health, thus avoiding unreasonable oscillations. Diffusion generation in a low-dimensional latent space rather than a high-dimensional original data space significantly reduces computational complexity. The characteristics of the continuous neural field support predictions at arbitrary time resolutions and facilitate integration with systems such as online updates and digital twins. Furthermore, this invention can decompose prediction uncertainty, providing three-level uncertainty analysis of data, model, and mode, making the prediction results not only reliable but also deeply diagnostic, guiding data acquisition and model optimization decisions. Attached Figure Description

[0029] Figure 1 The overall flowchart of the aero-engine remaining service life prediction method based on hierarchical latent space diffusion neural field provided in the embodiments of the present invention;

[0030] Figure 2 This is a schematic diagram of coarse-fine sampling and multimodal trajectory generation of hierarchical latent variables (stage codes / detail codes) in an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram illustrating the decomposition of RUL prediction uncertainty in an embodiment of the present invention. Detailed Implementation

[0032] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0033] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. 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.

[0034] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar words used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components.

[0035] Example 1

[0036] This embodiment details the specific implementation steps of the aero-engine remaining service life prediction method based on a layered latent space diffusion neural field provided by the present invention. For example... Figure 1 As shown, it includes:

[0037] S101: Acquire the observation sequence and operating condition information of the aero-engine, and preprocess the observation sequence.

[0038] Specifically, the observation sequence of the aero-engine is acquired. This sequence may include, but is not limited to, vibration acceleration signals, lubricating oil spectral data, and gas path parameters (such as exhaust temperature and fuel flow rate), typically in the form of multi-channel time-series data. Simultaneously, operating condition information is acquired, including flight phases (takeoff, cruise, and landing), thrust settings, and environmental conditions. Then, the observation sequence is preprocessed. Preprocessing operations include data cleaning (removing outliers), normalization, and multi-scale time-frequency decomposition. Specifically, the multi-scale time-frequency decomposition includes: performing short-time Fourier transform (STFT) on the observation data using time windows of different lengths to obtain multi-scale time-frequency maps; and adaptively selecting and fusing the multi-scale time-frequency map information at the frequency band level through a learnable gating mechanism. This helps the model simultaneously capture both periodic and non-stationary impact components. For example, short time windows (to capture transient impacts) and long time windows (to capture steady-state features) can be used in parallel for STFT. The resulting time-frequency graph learns the importance weights of each frequency band at different scales through a lightweight neural network (gated network) and performs weighted fusion, thereby enhancing the representation ability of features.

[0039] S102: Input the preprocessed observation sequence and operating condition information into the multimodal encoder to obtain the latent characterization and extract the stage code and detail code.

[0040] Specifically, the multi-scale features and operating condition information preprocessed in step S101 are input into a multimodal encoder. This multimodal encoder can consist of multiple branch networks, such as a 1D CNN branch processing the original waveform, a 2D CNN or ViT branch processing the time-frequency graph, and an MLP branch processing statistical features. The features output by each branch are fused through cross-attention or gating fusion mechanisms, ultimately outputting a low-dimensional latent representation vector z. Simultaneously, the encoder predicts two key latent variables through fully connected layers: the stage code g and the detail code l. The stage code is a discrete variable used to index K different degradation stage modes (e.g., K=4, corresponding to healthy, early degradation, accelerated degradation, and near-failure); the detail code is a continuous vector used to represent short-term fluctuations caused by random shocks, noise, etc., within the same stage. The multimodal encoder can also output the probability weights of each degradation stage mode, reflecting the probability that the current observation belongs to different degradation modes. The stage code g captures macroscopic degradation modes, while the detail code l represents microscopic fluctuations caused by noise, random shocks, etc., within the same macroscopic mode.

[0041] S103: Using stage code and detail code as conditions, perform hierarchical diffusion generation in the latent space to obtain multiple sets of future latent variable pairs that correspond to the current state.

[0042] The core of this step is to train a hierarchical diffusion model that learns to generate pairs of future latent variables consistent with current observations and operating conditions from standard Gaussian noise. , Specifically, this step involves generation in a low-dimensional latent space to reduce computational complexity and improve generation quality. Hierarchical diffusion generation specifically includes defining independent forward denoising and reverse denoising processes for the stage code and the detail code, respectively. In the reverse denoising process, using latent representations, operating condition information, and another latent variable as conditions (e.g., the detail code as a condition when denoising the stage code), a denoising network (such as U-Net) is used to predict noise, and target latent variable pairs are generated by sampling from standard Gaussian noise through iterative denoising.

[0043] The training process of the hierarchical diffusion model is as follows:

[0044] Training data preparation: A large amount of historical operating sequence data of aero-engines is collected. Each sequence contains observation data, operating condition information, and its corresponding subsequent degradation stage labels. Through the processing in steps S101 and S102, the corresponding latent representation z, stage code g, and detail code l are obtained for each training sample.

[0045] Training objective: The training objective is to improve the denoising network. Learn to predict the noise added to the clean target latent variable pair (g0, l0) during the forward noise addition process. For a random sampling diffusion time step t Uniform(1,T), the forward process adds noise to the target latent variable pair according to the selected noise schedule (the first noise schedule is used for g, and the second noise schedule is used for l), to obtain the noise version. , ).

[0046] Loss function: The model is trained by minimizing the following loss function:

[0047]

[0048] in, , This is the actual noise added. and These are denoising networks for stage codes and detail codes, respectively, whose inputs are conditioned on another latent variable, latent representation z, and operating condition information u. Let L represent the expected value of the joint distribution of diffusion time step t, clean target latent variable pair (g0,l0), latent representation z, and operating condition information u; L is the loss function.

[0049] Training process: The network parameters are updated through numerous iterations using optimization algorithms such as stochastic gradient descent. This continues until the loss function converges. After training, a model is obtained that can sample and generate reasonable pairs of future latent variables from standard Gaussian noise, conditioned on current observations. , The hierarchical diffusion model.

[0050] To optimize generation quality, a first noise scheduler can be used for the stage code and a second noise scheduler for the detail code. The noise growth rate of the first noise scheduler is slower than that of the second noise scheduler to maintain the stability of the macroscopic degradation pattern during generation. This means that the noise addition process for the stage code is smoother, ensuring the stability of the generated macroscopic pattern structure; the noise addition process for the detail code is faster, to generate rich microscopic variations within the same pattern, thereby enhancing the diversity within that pattern.

[0051] Specifically, the core characteristic of the first noise scheduling is its slow noise growth rate; for example, during the diffusion process of the total number of steps T, its noise variance... It increases slowly with the number of steps t; for example, the first noise schedule can be a linear schedule with a noise variance of from =1e-4 slowly increases to =0.02. This allows the macroscopic degradation pattern information carried by the stage code to be well preserved in the early stages of reverse denoising generation, avoiding being overwhelmed by rapidly added noise, thus ensuring the stability and consistency of the generated trajectory in the overall trend.

[0052] The core characteristic of the second noise dispatching method is its rapid noise growth rate and its noise variance. It approaches 1 more quickly with increasing step number t; for example, the second noise scheduling can employ cosine scheduling. This allows the detail code to be mixed with Gaussian noise more quickly, thereby sampling a richer variety of micro-fluctuation details under a given macroscopic pattern during inverse denoising generation, enhancing the diversity of trajectories within the same degradation stage.

[0053] This differentiated noise scheduling design, which involves slow adjustment of macroscopic parameters and fast adjustment of microscopic parameters, enables the generative model to flexibly generate diverse specific implementations (determined by detail codes) while maintaining the overall degradation path (determined by stage codes) as reasonable and stable. This more realistically simulates the coexistence of deterministic macroscopic trends and random microscopic fluctuations in the actual degradation process, and is an important technical guarantee for generating multimodal trajectories that conform to physical laws and have diversity.

[0054] S104: Input the future latent variable pairs, latent representations, working condition information and continuous future time coordinates into the continuous neural field decoder to decode and generate multiple future health indicator degradation trajectories.

[0055] Specifically, the multiple sets of future latent variable pairs generated in step S103 ( , The current latent representation z, operating condition information u, and continuous future time coordinates τ are input to the continuous neural field decoder Fψ. The continuous neural field decoder is a continuous function with time coordinates as input. It directly outputs the health index HI(τ) at the future time τ. By calculating for a series of continuous τ values, a complete degradation trajectory can be obtained. Repeating steps S103 and S104 multiple times (e.g., M=100 times) can generate multiple (M) different degradation trajectories, demonstrating the possibility of multimodality, such as... Figure 2 As shown.

[0056] To make the generated trajectory more consistent with physical laws, prior constraints need to be introduced into the loss function of the continuous neural field decoder. Preferably, constraints are introduced on the monotonic non-decreasing nature of the health index over time, and / or a penalty term for the sparsity of stage switching points. For example, the monotonicity constraint can be achieved by calculating the difference of the health index at randomly sampled time points and penalizing negative values. Furthermore, the decoder network can employ Fourier feature maps to better fit high-frequency changes.

[0057] S105: Based on a preset failure threshold and multiple degradation trajectories, determine the probability distribution of the remaining service life of the aero-engine.

[0058] Specifically, a failure threshold h is set. For each generated degradation trajectory, its health index HI(τ) is calculated to determine when it first exceeds the preset failure threshold h. The moment T f This moment represents the remaining useful life (RUL) corresponding to that trajectory. The specific calculation is as follows: T f =inf{τ:HI(τ)≥h }. Calculate the remaining service life T for all M trajectories. f By obtaining the remaining useful life (RUL) sample set, an empirical distribution containing M samples can be formed. Based on this empirical distribution, statistical analysis of the sample set can output the mean, median, quantiles (e.g., P5, P50, P95), and probability density distribution of the RUL, thus completing a rigorous probability prediction. This completes the mapping from multimodal trajectories to a rigorous RUL probability distribution.

[0059] Figure 2 This illustrates the core processes of hierarchical latent variable diffusion generation and multimodal trajectory formation in this invention.

[0060] First, in the hierarchical latent space, based on the prior distribution of the stage code g and detail code l inferred from current observations, a "coarse-to-fine" sampling process is performed. As described in step S103, coarse-grained sampling is first performed on the stage code g, which represents the macroscopic degradation pattern, to determine a general degradation path pattern (e.g., slow degradation or rapid degradation). Subsequently, under the given stage code conditions, fine-grained sampling is performed on the detail code l, which represents microscopic fluctuations, to add specific fluctuation details to the selected macroscopic pattern. Then, as described in step S104, the future latent variable pairs obtained from each set of sampling ( , By inputting the data into the continuous neural field decoder, a trajectory of health indicator degradation pointing from the current moment to the future can be generated. By repeating the above sampling and decoding process, multiple degradation trajectories reflecting different probabilities of failure can be obtained, such as... Figure 2 As shown in the lower left and upper right sides, these trajectories form a bundle diagram of future degradation scenarios. Finally, as in step S105 and... Figure 2 As shown in the lower right corner, a common failure threshold is set for these tracks. The point at which the health metric of each track first exceeds this threshold is recorded as a potential failure event. By aggregating and statistically analyzing the remaining lifetimes corresponding to all trajectories, a probability distribution of the remaining lifetime (RUL) is formed. This distribution exhibits a multi-peaked shape, with different peaks directly corresponding to several macroscopic degradation modes with higher probabilities indexed by different stage codes g, thus preserving the interpretability of multimodal information in probabilistic prediction.

[0061] Furthermore, to improve the reliability and operability of the prediction results, after outputting the probability distribution, the method in this embodiment further includes the following step:

[0062] S106: Decompose the uncertainty of the probability distribution of the remaining useful life to obtain the data uncertainty, model uncertainty and mode uncertainty.

[0063] Data uncertainty is characterized by the variance of trajectories generated from multiple samplings under the same diffusion model, reflecting the inherent randomness of the degradation process. Model uncertainty is characterized by the prediction variance of different initialization models (deep ensemble) or techniques such as MC Dropout, reflecting the knowledge uncertainty of the model itself due to limited training data. Mode uncertainty is characterized by the distribution (e.g., entropy) of the probability weights of modes at each degradation stage corresponding to the stage code, reflecting the ambiguity of current observations regarding the macroscopic mode attribution.

[0064] like Figure 3As shown, the proportion of various uncertainties can be visualized (e.g., output as a pie chart) and diagnostic suggestions can be provided, offering deeper insights for maintenance decisions. For example, if the proportion of model uncertainty is high, it may suggest that additional data may be needed to improve the model's confidence.

[0065] Figure 3 This is an exemplary schematic diagram illustrating the decomposition of remaining lifetime prediction uncertainty provided in an embodiment of the present invention. By performing source tracing analysis on the total prediction uncertainty, it can be decomposed into the following four components:

[0066] Data uncertainty: accounting for 68%, is caused by the inherent randomness of the degradation process. It can be quantified by performing multiple forward samplings on the same trained diffusion model and calculating the variance of the generated multiple degradation trajectories. This uncertainty usually cannot be eliminated by improving the model.

[0067] Mode uncertainty: accounting for 19%, reflects the model's confidence in the current macroscopic degradation stage of the engine (indexed by stage code). This uncertainty can be directly characterized by the distribution (e.g., entropy value) of the probability weights of each degradation stage mode output by the multimodal encoder.

[0068] Model uncertainty, accounting for 13%, stems from cognitive uncertainty in model parameters due to limited training data, primarily manifested in the prediction of micro-variables (detail codes). This uncertainty can be quantified by calculating the variance of different models' predictions of future details using techniques such as deep ensemble (training multiple models with different initializations) or Monte Carlo Dropout.

[0069] By outputting as Figure 3 The uncertainty decomposition pie chart shown in this embodiment of the invention not only provides the probability distribution of remaining lifetime but also endows the prediction results with deeper interpretability. Decision-makers can use this to identify the main sources of uncertainty: if data uncertainty dominates, it means the degradation process is highly random, and a more conservative maintenance strategy is recommended; if mode or model uncertainty dominates, it indicates that the current model understanding is insufficient, and more data collection or model validation under different operating conditions should be considered, thus achieving closed-loop support from prediction to diagnosis and decision-making.

[0070] Example 2

[0071] This embodiment provides a system for predicting the remaining service life of an aero-engine based on a hierarchical latent space diffusion neural field, used to implement the method described in Embodiment 1. The system includes:

[0072] The data preprocessing module is used to execute step S101, which involves acquiring the observation sequence and operating condition information of the aero-engine and preprocessing the observation sequence. Its preprocessing function includes multi-scale time-frequency decomposition.

[0073] The multimodal coding module is used to perform step S102, which encodes the preprocessed observation sequence and operating condition information into latent representations, and extracts the stage code representing macroscopic degradation mode and the detail code representing microscopic fluctuations, as well as the probability weight of each mode.

[0074] The hierarchical diffusion generation module is used to execute step S103, which is used to perform hierarchical diffusion generation in the latent space, based on stage codes and detail codes, to obtain multiple sets of future latent variable pairs that correspond to the current state.

[0075] The continuous neural field decoding module, used to execute step S104, is a continuous-time function used to decode future latent variable pairs, latent representations, working condition information, and continuous future time coordinates to generate multiple future health indicator degradation trajectories. Its training process incorporates prior constraints such as monotonicity.

[0076] The remaining service life estimation module is used to execute step S105, which is used to calculate the time when the health index of each degradation trajectory first crosses the failure threshold based on a preset failure threshold, obtain the remaining service life sample set, perform statistical analysis on the remaining service life sample set, and determine the probability distribution of the remaining service life of the aero-engine.

[0077] Uncertainty decomposition module: used to execute step S106, which decomposes the uncertainty of the probability distribution of the remaining useful life after the remaining useful life inference module outputs the probability distribution, and obtains the data uncertainty, model uncertainty and mode uncertainty.

[0078] The method and system described in this invention, through the innovative technical approach of "hierarchical latent variable diffusion generation + continuous neural field decoding + event time statistics", effectively solves multiple challenges in RUL prediction of aero-engines, such as multimodal expression, probabilistic rigor, physical rationality and result interpretability, and provides a powerful tool for achieving highly reliable predictive maintenance.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of an aero-engine based on a hierarchical latent space diffusion neural field, characterized in that, Includes the following steps: Acquire the observation sequence and operating condition information of the aero-engine, and preprocess the observation sequence; The preprocessed observation sequence and operating condition information are input into a multimodal encoder to obtain latent representations, and stage codes representing macroscopic degradation patterns and detail codes representing microscopic fluctuations are extracted. The stage codes are discrete variables used to index K different degradation stage patterns, and the detail codes are continuous vectors used to represent short-term fluctuations under the same stage. Using the stage code and detail code as conditions, hierarchical diffusion generation is performed in the latent space to obtain multiple sets of future latent variable pairs that correspond to the current state; wherein, a first noise scheduling is used for the stage code and a second noise scheduling is used for the detail code, and the noise growth rate of the first noise scheduling is slower than that of the second noise scheduling, so as to maintain the stability of the macroscopic degradation mode during the generation process; The future latent variable pairs, latent representations, working condition information, and continuous future time coordinates are input into the continuous neural field decoder to decode and generate multiple future health indicator degradation trajectories. The continuous neural field decoder is a continuous function with time coordinates as input and outputs health indicators at future time coordinates. By calculating multiple continuous future time values, a complete degradation trajectory is obtained. By repeating the calculation multiple times, multiple future health indicator degradation trajectories are generated. Based on a preset failure threshold, the time when the health index of each degradation trajectory first crosses the failure threshold is calculated to obtain a sample set of remaining service life. Perform statistical analysis on the remaining lifespan sample set and output the probability distribution of the remaining lifespan.

2. The method according to claim 1, characterized in that, The preprocessing includes performing multi-scale time-frequency decomposition on the observation sequence. The multi-scale time-frequency decomposition includes: performing short-time Fourier transform on the observation data using time windows of different lengths to obtain multi-scale time-frequency maps; and adaptively selecting and fusing the multi-scale time-frequency maps at the frequency band level through a learnable gating mechanism.

3. The method according to claim 1, characterized in that, The stage code is a discrete variable used to index K different degradation stage modes; the detail code is a continuous vector; the multimodal encoder also outputs the probability weights of each degradation stage mode.

4. The method according to claim 1, characterized in that, The layered diffusion generation includes: Define independent forward noise addition and reverse noise reduction processes for the stage code and the detail code, respectively; In the reverse denoising process, the noise is predicted using the latent representation, operating condition information and another latent variable as conditions, and the target latent variable pair is generated by sampling from the standard Gaussian noise through iterative denoising.

5. The method according to claim 1, characterized in that, The continuous neural field decoder is a continuous function with time coordinates as input; in the loss function for training the continuous neural field decoder, a monotonically non-decreasing constraint on health indicators over time and / or a penalty term for the sparsity of stage switching points are introduced.

6. The method according to claim 1, characterized in that, For each generated degradation trajectory, the moment when its health index first exceeds the failure threshold is calculated as the remaining service life corresponding to that trajectory; the remaining service life corresponding to multiple degradation trajectories is statistically analyzed to form an empirical distribution, and the mean, quantiles and probability density of the remaining service life are output based on the empirical distribution.

7. The method according to claim 1, characterized in that, After outputting the probability distribution, the method further includes: The uncertainty of the probability distribution of the remaining useful life is decomposed to obtain data uncertainty, model uncertainty and mode uncertainty; The data uncertainty is characterized by the variance of trajectories generated by multiple samplings under the same diffusion model; the model uncertainty is characterized by the prediction variance of different initialization models; and the mode uncertainty is characterized by the distribution of probability weights of each degradation stage mode corresponding to the stage code.

8. A prediction system for the remaining service life of an aero-engine based on a hierarchical latent space diffusion neural field, characterized in that, include: The data preprocessing module is used to acquire the observation sequence and operating condition information of the aero-engine, and to preprocess the observation sequence. The multimodal coding module is used to encode the preprocessed observation sequence and operating condition information into latent representations, and extract stage codes representing macroscopic degradation patterns and detail codes representing microscopic fluctuations; wherein, the stage codes are discrete variables used to index K different degradation stage patterns, and the detail codes are continuous vectors used to represent short-term fluctuations under the same stage. The hierarchical diffusion generation module is used to perform hierarchical diffusion generation in the latent space, based on the stage code and detail code, to obtain multiple sets of future latent variable pairs that correspond to the current state; wherein, a first noise scheduling is used for the stage code and a second noise scheduling is used for the detail code, and the noise growth rate of the first noise scheduling is slower than that of the second noise scheduling, so as to maintain the stability of the macroscopic degradation mode during the generation process. The continuous neural field decoding module is used to decode the future latent variable pairs, latent representations, working condition information, and continuous future time coordinates to generate multiple future health indicator degradation trajectories. The continuous neural field decoding module is a continuous function with time coordinates as input and outputs health indicators at future time coordinates. By calculating multiple consecutive future time values, a complete degradation trajectory is obtained. By repeating the calculation multiple times, multiple future health indicator degradation trajectories are generated. The remaining service life inference module is used to calculate the time when the health index of each degradation trajectory first crosses the failure threshold based on a preset failure threshold, obtain a remaining service life sample set, perform statistical analysis on the remaining service life sample set, and determine the probability distribution of the remaining service life of the aero-engine.

9. The system according to claim 8, characterized in that, The system also includes: The uncertainty decomposition module is used to decompose the uncertainty of the probability distribution of the remaining useful life after the remaining useful life inference module outputs the probability distribution, and obtain the data uncertainty, model uncertainty and mode uncertainty.

Images