Method and device for predicting residual life of aero-engine based on semantic injection of large language model and risk constraint

Abstract

The embodiment of the application provides a kind of based on big language model semantic injection and risk constraint's aero-engine residual life prediction method and device, it is related to prediction and health management technical field, wherein, the method comprises the following steps: the time series data of multivariate sensor and working condition data of aero-engine are intercepted to sliding window, generate health characterization vector, and trend feature of key sensor is extracted simultaneously;Based on trend feature and working condition data, generate trend abstract text input to first prompt word template and second prompt word template, first semantic score is mapped as first semantic coefficient, and second semantic score is mapped as second semantic coefficient;Risk constraint term is added to the objective function of risk constraint reinforcement learning model;Health characterization vector, first semantic coefficient and second semantic coefficient are input to the risk constraint reinforcement learning model trained as engine state.The scheme enhances the explainability and security of prediction result.

Description

Technical Field

[0001] This invention relates to the field of prediction and health management technology, and in particular to a method and apparatus for predicting the remaining life of aero-engines based on semantic injection of a large language model and risk constraints. Background Technology

[0002] As a complex and high-end piece of equipment operating in harsh environments, the operational safety and reliability of aero engines are of paramount importance. Remaining Useful Life (RUL) prediction is a core technology in the aero engine predictive health management (PHM) system. It aims to accurately estimate the remaining operating time of an engine from its current state to functional failure, providing crucial decision-making support for developing Condition-Based Maintenance (CBM) strategies, optimizing spare parts inventory, and fleet scheduling. The degradation process of aero engines is influenced by variable operating conditions and loads, exhibiting high nonlinearity, non-stationarity, and stochasticity.

[0003] Existing data-driven RUL prediction methods, such as those based on Long Short-Term Memory (LSTM), Temporal Convolutional Networks (TCN), or Transformer deep learning models, have demonstrated good performance in handling multivariate time-series data, but still face challenges in practical engineering applications. On the one hand, these methods are typically trained using risk-neutral metrics such as mean squared error (MSE), aiming to minimize the average prediction error, but are not sensitive enough to "catastrophic overestimation" (i.e., predicted remaining life far exceeding actual values) that could lead to serious safety accidents. On the other hand, deep learning models are often considered "black boxes," and their predictions are difficult to interpret. Engineers cannot intuitively connect the model's output with specific physical degradation patterns (such as compressor efficiency decline or abnormal combustion chamber temperatures), resulting in insufficient credibility and auditability of the predictions, making it difficult to establish a traceable basis for maintenance. Furthermore, relying solely on numerical features for modeling also limits the model's generalization ability under different operating conditions.

[0004] Therefore, existing technologies urgently need a new RUL prediction scheme that not only pursues the accuracy of prediction, but also needs to be able to explicitly control the risk of overestimation in the prediction, and provide interpretable and auditable prediction basis to meet the stringent requirements of reliability, interpretability and safety in safety-critical fields. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method for predicting the remaining service life of aero-engines based on semantic injection and risk constraints using a large language model, to solve the technical problem that existing remaining service life prediction methods struggle to balance risk control, interpretability, and prediction accuracy. The method includes:

[0006] Acquire multivariable sensor time-series data and operating condition data of aero-engines, perform preprocessing and sliding window truncation to generate window sample sequences, extract time-series features from the window sample sequences to generate health characterization vectors, and simultaneously extract trend features of key sensors.

[0007] Based on the trend features and the working condition data, a trend summary text is generated. The trend summary text is input into the first prompt word template and the second prompt word template respectively. Semantic reasoning is performed through a large language model to output a first semantic score and a second semantic score. The first semantic score is mapped to a first semantic coefficient and the second semantic score is mapped to a second semantic coefficient.

[0008] A training set is constructed based on the health representation vector, the first semantic coefficient, and the second semantic coefficient, and a risk constraint term is added to the objective function of the risk constraint reinforcement learning model to train the risk constraint reinforcement learning model.

[0009] The health representation vector, the first semantic coefficient, and the second semantic coefficient are used as engine state inputs to the trained risk-constrained reinforcement learning model, outputting the remaining life prediction distribution parameters, and modulating the remaining life prediction distribution parameters through the first semantic coefficient, outputting the modulated distribution parameters.

[0010] Based on the modulated distribution parameters, the point estimate, prediction range, and risk index of the remaining life of the aero-engine are calculated and output.

[0011] This invention also provides an aero-engine remaining life prediction device based on large language model semantic injection and risk constraints, to solve the technical problem that existing remaining life prediction methods struggle to balance risk control, interpretability, and prediction accuracy. The device includes:

[0012] A health representation vector module is constructed to acquire multivariable sensor time-series data and operating condition data of aero-engines, and to perform preprocessing and sliding window truncation to generate a window sample sequence. The time-series features of the window sample sequence are extracted to generate a health representation vector, and the trend features of key sensors are extracted simultaneously.

[0013] The semantic coefficient reasoning module is used to generate trend summary text based on the trend features and the working condition data, input the trend summary text into the first prompt word template and the second prompt word template respectively, perform semantic reasoning through the large language model, output the first semantic score and the second semantic score, and map the first semantic score to the first semantic coefficient and the second semantic score to the second semantic coefficient.

[0014] The model training module is used to construct a training set based on the health representation vector, the first semantic coefficient and the second semantic coefficient, and to add the risk constraint term to the objective function of the risk constraint reinforcement learning model to train the risk constraint reinforcement learning model.

[0015] The output distribution parameter module is used to input the health representation vector, the first semantic coefficient and the second semantic coefficient as engine state inputs to the trained risk constraint reinforcement learning model, output the remaining life prediction distribution parameter, and modulate the remaining life prediction distribution parameter through the first semantic coefficient to output the modulated distribution parameter.

[0016] The remaining life estimation module is used to calculate and output the point estimate, prediction range, and risk index of the remaining life of the aero-engine based on the modulated distribution parameters.

[0017] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least:

[0018] By introducing semantic reasoning and risk constraints, the risk of overestimation is reduced, the interpretability and security of the prediction results are enhanced, and a reliable basis for condition-based maintenance is provided. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a method for predicting the remaining life of an aero-engine based on semantic injection and risk constraints using a large language model, provided in an embodiment of the present invention.

[0021] Figure 2 This is a structural block diagram of an aero-engine remaining life prediction device based on semantic injection and risk constraints of a large language model, provided in an embodiment of the present invention. Detailed Implementation

[0022] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0023] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application 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 this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] In this embodiment of the invention, a method for predicting the remaining life of aero-engines based on semantic injection of a large language model and risk constraints is provided, such as... Figure 1 As shown, the method includes:

[0025] Step S101: Acquire multivariable sensor time-series data and operating condition data of the aero-engine, perform preprocessing and sliding window truncation to generate window sample sequence, extract time-series features from the window sample sequence to generate health characterization vector, and simultaneously extract trend features of key sensors.

[0026] Step S102: Based on the trend features and the working condition data, generate trend summary text, input the trend summary text into the first prompt word template and the second prompt word template respectively, perform semantic reasoning through the large language model, output the first semantic score and the second semantic score, and map the first semantic score to the first semantic coefficient and the second semantic score to the second semantic coefficient.

[0027] Step S103: Construct a training set based on the health representation vector, the first semantic coefficient, and the second semantic coefficient, and add the risk constraint term to the objective function of the risk constraint reinforcement learning model to train the risk constraint reinforcement learning model;

[0028] Step S104: Input the health representation vector, the first semantic coefficient and the second semantic coefficient as engine state input to the trained risk constraint reinforcement learning model, output the remaining life prediction distribution parameters, and modulate the remaining life prediction distribution parameters through the first semantic coefficient to output the modulated distribution parameters;

[0029] Step S105: Based on the modulated distribution parameters, calculate and output the point estimate, prediction interval, and risk index of the remaining life of the aero-engine.

[0030] In specific implementation, the following steps are used to extract time-series features from the window sample sequence, generate a health representation vector, and simultaneously extract trend features from key sensors:

[0031] The window sample sequence is input into a time-series encoder to extract a health representation vector. The time-series encoder is used to extract features representing the global degradation state of the engine from the multivariate sensor time-series data. Based on the physical correlation between the sensor and the engine failure mode and / or the mutual information index calculated based on the training set, key sensors are selected from all sensors installed on the aero-engine. Extract the key sensors from the window sample sequence. Trend characteristics ,in, , This is the slope calculation function. The standard deviation of the key sensor on the training set is given by [reference to a specific sensor]. For the current moment, The length of the sliding window.

[0032] In specific implementation, the following steps are used to generate trend summary text based on the trend characteristics and the operating condition data:

[0033] The direction and intensity of change in the trend features are mapped to symbolic representations according to a preset discretization threshold, wherein the symbolic representations include symbols for rising, falling, and remaining unchanged; the operating condition data are mapped to corresponding operating condition status labels; the symbolic representations and the operating condition status labels are combined according to a preset text template to generate the trend summary text in natural language form.

[0034] In specific implementation, the following steps are used to input the trend summary text into the first prompt word template and the second prompt word template respectively, perform semantic reasoning through a large language model, output a first semantic score and a second semantic score, and map the first semantic score to a first semantic coefficient and the second semantic score to a second semantic coefficient:

[0035] Input the trend summary text into the first prompt word template and output the first discrete semantic score. The first prompt word template is configured to guide the large language model to assess the current degradation trend from an engineering conservatism perspective, and the first discrete semantic score is positively correlated with the aggressiveness of the prediction strategy. The number of rating levels; input the trend summary text into the second prompt word template, and output the second discrete semantic score. The second prompt word template is configured to guide the large language model to assess the current degradation trend from a security risk perspective, and the second discrete semantic score is positively correlated with the risk of overestimating remaining lifetime. Given the number of risk levels; construct the first mapping function. Second mapping function , where the first mapping function and the second mapping function All are preset monotonic functions, the first mapping function The second mapping function is used to convert discrete semantic scores into action modulation mathematical coefficients. Used to convert discrete semantic scores into risk-weighted mathematical strengths; through the first mapping function The first discrete semantic score Mapped to a first basic adjustment value, and through the second mapping function The second discrete semantic score Mapped to the second basic adjustment amount; and set the first dynamic scheduling factor related to the training progress respectively. Second dynamic scheduling factor ,in, The current training cycle number; determined by the first dynamic scheduling factor. and the second dynamic scheduling factor The first basic adjustment amount and the second basic adjustment amount are scaled respectively to obtain the first semantic coefficient. Second semantic coefficient ,in, , .

[0036] In specific implementation, the first mapping function is constructed through the following steps. Second mapping function :

[0037] Based on the safety standards for aircraft engines, a first semantic score is set. Total number of discrete levels Second semantic score Total number of discrete levels Second mapping function The theoretical upper limit of the output value Based on the total number of discrete levels Construct the first mapping function Based on the total number of discrete levels and theoretical upper limit Construct the second mapping function .

[0038] In practice, the following steps are used to set the first dynamic scheduling factor related to the training progress. Second dynamic scheduling factor :

[0039] The training schedule is divided into a warm-up period, a stabilization period, and an annealing period; if the training schedule... During the preheating period, the scheduling factor ,in, As a dynamic scheduling factor, The preset maximum strength, This represents the maximum training progress during the warm-up period; if the training progress... During the stable period, the scheduling factor If training progress During the annealing period, the scheduling factor ,in, scheduling factor The theoretical lower limit throughout the entire training process. This represents the maximum training progress during the annealing period, and .

[0040] In specific implementation, the following steps are used to construct a training set based on the health representation vector, the first semantic coefficient, and the second semantic coefficient, and to add the risk constraint term to the objective function of the risk constraint reinforcement learning model for training:

[0041] Based on each sample time t Obtain the corresponding health representation vector. First semantic coefficient Second semantic coefficient and actual remaining lifespan and construct a training set; then, use the health representation vectors... The first semantic coefficient and the second semantic coefficient The process involves fusing the data to construct an engine state representation for the risk-constrained reinforcement learning model. ,in, Vector concatenation; representing the engine state The input is fed into the policy network of the risk-constrained reinforcement learning model to obtain the initial remaining lifetime prediction distribution parameters. ,in, For the engine state representation The mean of the Gaussian distribution, For the engine state representation The logarithmic standard deviation; through the first semantic coefficient Predicted distribution parameters of the initial remaining lifetime Modulation is performed to generate modulated distributed parameters. ,in, , Element-wise multiplication; based on actual remaining lifetime value and modulated distributed parameters Calculate instant rewards ,in, For probability quantiles, In order to be in The absolute error at that time Based on the loss weight, To overestimate the penalty coefficient, () is an indicator function. For the modulated distribution parameters The median estimate of the remaining lifespan predicted by calculation; through the second semantic coefficient For instant rewards Risk-weighted calculation to generate weighted returns The optimization problem of constructing the risk-constrained reinforcement learning model is based on maximizing expected return and satisfying the conditional value at risk (CVaR) constraint. The parameters of the risk-constrained reinforcement learning model are iteratively updated until the risk-constrained reinforcement learning model converges, thus obtaining the trained risk-constrained reinforcement learning model.

[0042] In specific implementation, the optimization problem of the risk-constrained reinforcement learning model is constructed by taking the maximization of expected return and the conditional value at risk (CVaR) constraint as the optimization objective, based on the weighted return. The parameters of the risk-constrained reinforcement learning model are iteratively updated until the model converges, resulting in the trained risk-constrained reinforcement learning model.

[0043] Through the weighted returns Generate a trajectory reward dataset for policy evaluation. ,in, , The training trajectory is defined by H, which is the preset trajectory field of view length. For time step index, The discount factor is preset; based on the trajectory reward dataset. Constructing constrained optimization problems ,in, For the parameters of the policy network, For expectation operator, For parameters Defined strategy function, For the condition of risk value, The risk performance threshold is used to transform the constrained optimization problem into a Lagrangian function maximization / minimization problem. ,in, As the dual variable; repeatedly perform the Lagrange function minimization problem and update the parameters of the policy network. and the dual variable Continue until the convergence condition is met.

[0044] In specific implementation, the following steps are used to input the health representation vector, the first semantic coefficient, and the second semantic coefficient as engine state inputs into the trained risk-constrained reinforcement learning model, output remaining life prediction distribution parameters, and modulate the remaining life prediction distribution parameters using the first semantic coefficient to output modulated distribution parameters:

[0045] Through the health representation vector The first semantic coefficient and the second semantic coefficient Constructing an engine state representation The engine state is represented. The input is fed into the trained risk-constrained reinforcement learning model, and the initial remaining life prediction distribution parameters are output. ; through the first semantic coefficient Predicted distribution parameters of the initial remaining lifetime Modulation is performed to generate modulated distributed parameters. .

[0046] Before providing a further detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0047] (1) Remaining Useful Life (RUL): refers to the expected time that a device or system can continue to operate from the current moment until its function fails or it can no longer meet the predetermined performance requirements. In scenarios such as discrete cyclic operation of aircraft engines, the remaining useful life is usually represented by the "remaining cycles" and is a core prediction target in the field of prediction and health management (PHM).

[0048] (2) Large Language Model (LM): refers to a type of deep learning model trained on large-scale text data, which has powerful natural language understanding, generation and reasoning capabilities. In the embodiments of this application, the large language model is used as a semantic reasoning engine to convert numerical sensor trend information into interpretable semantic scores with engineering significance, thereby injecting domain knowledge and risk judgment into the subsequent prediction model.

[0049] (3) Conditional Value-at-Risk (CVaR): This is a metric used to quantify and manage risk in the financial or engineering fields. It measures the conditional expected value of loss under extremely adverse conditions (e.g., the tail of the return distribution) exceeding a certain Value-at-Risk (VaR) threshold. In embodiments of this application, CVaR is used as an optimization constraint to explicitly control and reduce the tail risk of catastrophic overestimation in the prediction of the remaining life of aero-engines.

[0050] (4) Trend Summary: This refers to a concise natural language description generated by symbolizing and textualizing the dynamic trends in multivariate time-series data from key sensors. It typically includes the direction and intensity of changes in key parameters, as well as information on the operating conditions of the equipment, aiming to transform complex numerical signals into a format that is easy for humans and machines (especially large language models) to understand.

[0051] (5) Semantic Action Coefficient and Semantic Risk Coefficient: These are two numerical coefficients obtained by mapping the semantic score output by the large language model through a specific mathematical function. The semantic action coefficient is used to directly and finely adjust or modulate the predicted value output by the model during the inference stage, reflecting the semantic guidance for the prediction strategy (such as conservative or aggressive). The semantic risk coefficient is used to weight the model's learning process during the training stage, for example, amplifying the influence of high-risk samples in the loss or reward function, thereby guiding the model to pay more attention to and avoid specific risk scenarios.

[0052] (6) Overestimation: In the context of remaining life prediction, this refers to the model predicting a remaining life value that is significantly higher than the actual remaining life value. In safety-critical systems such as aero-engines, overestimation is a more dangerous error than underestimation (predicted value is lower than the actual value) because it may lead to delays in maintenance decisions, thereby increasing the risk of unexpected failures during operation.

[0053] This invention provides a method for predicting the remaining life of aero-engines based on semantic injection and risk constraints using a large language model. By introducing the semantic understanding capabilities of a large language model and a reinforcement learning framework for risk constraints, it achieves high-precision and high-reliability prediction of the remaining life of aero-engines.

[0054] In the embodiments of the present invention, the data acquisition and preliminary processing steps are performed first.

[0055] The system acquires multi-dimensional time-series sensor data from the engine's health monitoring unit or historical database. This data may include dozens of physical quantities reflecting the engine's internal state, such as turbine outlet temperature, high-pressure compressor outlet pressure, and fan speed. Simultaneously, operating condition data synchronized with this time-series data is also acquired, such as flight altitude, Mach number, and throttle lever angle. This operating condition data is crucial for correctly interpreting the sensor signals. After acquiring the raw data, a series of preprocessing operations are required, including data cleaning to remove outliers and fill in missing values, data standardization to eliminate the influence of different sensor dimensions, and optional filtering and smoothing operations to reduce signal noise. Subsequently, a sliding window technique is used to truncate the preprocessed time-series data, generating a series of fixed-length window sample sequences. Each window sample represents the engine's recent operational history, serving as the basic unit for subsequent analysis.

[0056] After generating the window sample sequence, the method enters the feature extraction and information synthesis stage. For each window sample sequence, the system performs a temporal feature extraction process, aiming to extract a low-dimensional feature vector from the high-dimensional, complex time-series data that can effectively characterize the overall health degradation state of the engine. This vector is called the health characterization vector. This vector captures the complex correlations between multi-sensor data and the degradation patterns that evolve over time. Simultaneously, to provide more targeted information for subsequent semantic analysis, the system also extracts trend features from some key sensors from the window samples. These trend features are typically quantitative indicators describing the direction and rate of change of sensor readings within the window, such as the slope of a linear fit.

[0057] Next, leveraging the features extracted in the previous step, semantic input is generated for interaction with the large language model. Specifically, based on the extracted key sensor trend features and corresponding operating condition data, the system generates a natural language text, i.e., a trend summary text, according to preset rules and templates. This text transforms abstract numerical changes (such as "the slope of sensor A is -0.05") into descriptions that both humans and machines can understand (such as "under cruising conditions, sensor A shows a slight downward trend"). This transformation is a crucial bridge connecting the purely data-driven model with the world of semantic knowledge.

[0058] After obtaining the trend summary text, the method utilizes a large language model for semantic reasoning. The system inputs the same trend summary text into two functionally different cue word templates. The first cue word template (first cue word template) is designed to guide the large language model to assess the current engine degradation trend from an engineering conservatism perspective and output a corresponding score, called the first semantic score. The second cue word template (second cue word template) guides the large language model to assess from a safety risk perspective, particularly from the perspective of overestimating risk, and outputs the second semantic score. These two scores represent the qualitative judgments that domain experts might make based on the current trend. Subsequently, the system uses a pre-defined mapping function to convert these two discrete semantic scores into continuous numerical coefficients, namely the first semantic coefficient and the second semantic coefficient, for integration and use in subsequent mathematical models.

[0059] During the training phase, this method constructs a special training set and employs a risk-constrained reinforcement learning paradigm to optimize the prediction model. Specifically, each sample in the training set includes a health representation vector extracted from the data and two semantic coefficients obtained through inference from a large language model. The prediction model itself is constructed as a risk-constrained reinforcement learning model. A key innovation in defining the model's optimization objective is the inclusion of an explicit risk constraint term in its objective function. This risk constraint term aims to penalize extreme adverse behaviors that the model may exhibit during the prediction process, such as producing excessively high remaining life expectancy predictions. By training under the guidance of this risk-constrained objective function, the model not only learns how to minimize the average prediction error but also learns how to proactively avoid predictions that could lead to safety hazards.

[0060] Once the model training is complete, it can enter the online inference or prediction phase. At this point, for a new engine state, the system extracts its health representation vector and obtains the corresponding first and second semantic coefficients through the large language model. These three together constitute a comprehensive description of the current engine state and are provided as input to the trained risk-constrained reinforcement learning model. After receiving the input, the model outputs a set of remaining life prediction distribution parameters. These parameters describe a probabilistic prediction of the future remaining life, rather than just a single point estimate. To apply the semantic judgment of the large language model to the prediction results in real time, the method modulates this initial set of distribution parameters using the first semantic coefficients, generating a set of semantically adjusted distribution parameters. This modulation process can be understood as using the qualitative judgment of experts (reflected in the first semantic coefficients) to fine-tune the quantitative output of the model, making it more consistent with the conservative or aggressive requirements of engineering practice.

[0061] Finally, based on this set of semantically modulated distribution parameters, the system calculates and outputs the final remaining lifetime indicators. This includes calculating point estimates of the remaining lifetime, such as the median or mean of the probability distribution, which provides the user with the most probable remaining lifetime prediction. Simultaneously, the system calculates prediction intervals, such as 90% confidence intervals, which give the range of possible remaining lifetime, reflecting the uncertainty of the prediction. In addition, it can output a series of risk indicators, such as the probability that the predicted value exceeds a certain danger threshold, or the risk level associated with the second semantic coefficient. These diverse outputs collectively constitute a comprehensive health assessment report, providing rich, reliable, and interpretable evidence for aircraft engine maintenance decisions.

[0062] The core technical idea of ​​this invention lies in its approach: instead of simply applying deep learning models to time-series prediction tasks, it creates a hybrid intelligent framework that combines data-driven and knowledge-driven approaches. By transforming difficult-to-quantify domain knowledge and risk preferences (e.g., aversion to overestimation) into computable semantic coefficients through a large language model and seamlessly integrating them into the risk-constrained reinforcement learning training and inference process, this method significantly improves the credibility and security of prediction results. It allows the prediction model to pursue high accuracy while its behavior is constrained by clear risk boundaries and semantic logic, making it more suitable for safety-critical fields such as aviation, aerospace, and nuclear energy.

[0063] The benefits of this method are multifaceted. First, by explicitly incorporating a risk constraint term into the objective function, the model can effectively suppress catastrophic overestimation that is prone to occur at the end of the lifespan or during sudden changes in operating conditions, thus improving the safety of the prediction. Second, the introduced trend summarization and semantic scoring mechanisms provide a clear, traceable, and auditable explanation path for the previously "black box" prediction process, enhancing user trust in the prediction results. Finally, the dynamic modulation of the model output by semantic coefficients makes the prediction more flexible and adaptable to different scenarios, enabling the prediction strategy to be dynamically adjusted in terms of conservatism based on real-time degradation trends and risk assessments.

[0064] Furthermore, in a preferred embodiment, the specific process of extracting temporal features from the window sample sequence to generate a health representation vector, and simultaneously extracting trend features from key sensors, is refined. This temporal coding can employ advanced deep learning structures such as Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Gated Recurrent Units (GRUs), Temporal Convolutional Networks (TCNs), or Transformers. These structures excel at capturing long-term dependencies and complex dynamic patterns in time-series data. The role of temporal coding is to learn and extract a compact and information-rich health representation vector from multivariate sensor time-series data. This vector is designed to characterize the global, comprehensive degradation state of the engine.

[0065] At the same time, in order to generate a more physically meaningful trend summary, it is necessary to identify which sensors are "critical." Key Sensors The determination of key sensors can be based on two or a combination of criteria. The first is based on domain knowledge and physical relevance. For example, aero-engine experts, based on their understanding of engine failure modes (such as compressor blade wear and turbine blade creep), directly designate sensors closely related to these failure modes (such as compressor outlet pressure and turbine outlet temperature) as key sensors. The second is a data-driven approach, such as calculating the mutual information (MIA) index between each sensor signal and its real remaining lifetime (RUL) on the training set, and selecting the sensors with the highest MIA values ​​as key sensors. Mutual information measures the nonlinear correlation between two variables, thus making it an effective feature selection tool.

[0066] Key sensors identified Then, trend features can be extracted from the window sample sequence. A specific and effective trend feature is the normalized slope. The calculation formula is as follows: .in, For the current moment, The length of the sliding window. Indicates key sensors The data sequence within the current time window. The function is used to calculate the slope of the sequence. It can be estimated using simple linear regression (least squares), or, to enhance robustness to outliers, the Theil-Sen estimator or Huber regression can be used. Denominator It is this key sensor Normalizing using the standard deviation across the entire training set eliminates differences in the dimensions and fluctuation ranges of different sensors, making the trend features of different sensors comparable. This strategy of separating and extracting global health representations from local key trend features ensures that the model has a comprehensive grasp of the overall state and provides clear and focused input for subsequent semantic analysis. The technical effect is to improve the representational ability and interpretability of features.

[0067] In another preferred embodiment, the process of generating trend summary text based on trend features and operating condition data is specified. This process aims to transform numerical features into structured natural language, which is the first step in achieving semantic injection. First, the continuous trend features need to be discretized and symbolized. The normalized slope calculated above... For example, the system presets a series of discretization thresholds, such as those based on the percentiles of all slope values ​​in the training set (e.g., 20%, 40%, 60%, 80%). Then, the calculated slope values ​​are compared with these thresholds and mapped to a set of predefined symbolic representations. For instance, a very large positive slope is mapped to "sharp increase (↑↑)", a moderate positive slope to "increase (↑)", a slope close to zero to "flat (→)", and a negative slope to "decrease (↓)" or "sharp decrease (↓↓)". This symbolic representation intuitively conveys the direction and strength of the trend.

[0068] Secondly, similar processing is required for the operational data acquired synchronously with the sensor data. Operational data typically consists of multiple continuous variables (such as altitude and Mach number) or discrete variables. The system will map these combinations of operational variables to corresponding operational status labels, such as "takeoff," "climb," "cruise," and "descent," based on preset rules or cluster analysis. This helps to consider the impact of different operating conditions on sensor signals in subsequent semantic analysis.

[0069] Finally, the symbolic representations of the key sensors obtained from the above processing are combined with the operating condition labels according to a preset text template to generate the final trend summary text in natural language form. For example, a template could be: "Under the [operating condition label] condition, the key sensors behave as: [sensor 1 symbol][sensor 2 symbol] …". After filling in the blanks, the generated text might be: "Under the cruise condition, the key sensors behave as: T50↑↑ P30↓ Nf→". The technical advantage of this method is that it creates a standardized, information-rich, and easily readable intermediate representation for both humans and LLM processing, successfully transforming the raw, unstructured numerical data stream into structured text with preliminary semantics, laying a solid foundation for subsequent high-level semantic reasoning.

[0070] Furthermore, the complete process of inputting the trend summary text into the prompt word template, performing semantic reasoning through the Large Language Model (LLM), and finally generating semantic coefficients is also described in detail. This process is the core link in the knowledge injection of this invention. First, the system inputs the generated trend summary text into two carefully designed first and second prompt word templates simultaneously. The role of the first prompt word template is to guide the Large Language Model (LLM) to evaluate from the perspective of engineering conservatism. For example, the template may contain an instruction like this: "You are an experienced aero-engine health management expert. Based on the following trend summary, determine what strategy should be adopted for predicting the remaining life of this engine. Please output an integer score from 1 to N, where 1 represents 'extremely conservative' and N represents 'optimistic'." After understanding the trend summary and this instruction, the LLM will output a first discrete semantic score. The score is directly proportional to the aggressiveness of the prediction strategy.

[0071] Meanwhile, the second prompt template guides the LLM to assess from a security risk perspective, particularly focusing on the risk of overestimation. For example, the template might instruct: "You are a security assessment engineer. Based on the following trend summary, assess the risk level of overestimation of remaining lifetime in the current state. Output an integer score from 1 to M, where 1 represents 'very low risk' and M represents 'very high risk'." The LLM then outputs a second discrete semantic score accordingly. The higher the score, the greater the risk of overestimating remaining life expectancy.

[0072] After obtaining these two discrete scores, they need to be converted into continuous coefficients that can be used in the numerical model. To this end, the system constructs a first mapping function. Second mapping function Both of these functions are pre-defined monotonic functions, ensuring that the higher the score, the higher or lower the corresponding coefficient value (monotonically). First mapping function Used to represent the aggressiveness of the strategy. This is converted into mathematical coefficients used to adjust the model output, while the second mapping function... This is used to assign a score representing the risk level. This is converted into mathematical strength for the weighted training process. The first and second base adjustments are obtained through these two functions, respectively.

[0073] To make the semantic knowledge injection process smoother and more controllable, and to avoid excessive interference to the model in the early or late stages of training, this embodiment also introduces a dynamic scheduling mechanism. Specifically, a scheduling mechanism is set relative to the training progress (e.g., the current number of training epochs). The two related dynamic scheduling factors, namely the first dynamic scheduling factor Second dynamic scheduling factor These two factors are usually The function's value changes as training progresses. Finally, the first and second base adjustment values ​​are scaled using these two dynamic scheduling factors to obtain the final first semantic coefficients. Second semantic coefficient Their calculation formulas are: and This design allows the strength of semantic injection to be dynamically adjusted as the training process progresses, providing strong guidance in the early stages of training and gradually weakening it in later stages, allowing the model to fine-tune from the data. The technical effect of this approach is that it establishes a complete and controllable link from qualitative semantic judgment to quantitative numerical modulation, enabling domain knowledge to be deeply integrated into the model's learning and reasoning process in a flexible, dynamic, and mathematically rigorous manner.

[0074] In a more specific implementation, the first mapping function Second mapping function The construction provides a clear mathematical form. The specific form of these functions can be set according to engineering requirements and security standards. First, the first semantic score needs to be set according to the refined management requirements of the application scenario. Total number of discrete levels (For example, N=5, representing a five-tier strategy), and the second semantic score. Total number of discrete levels (For example, M=5 represents level five risk). Furthermore, a second mapping function needs to be set according to safety regulations. Theoretical upper limit of output value ,this The value determines the maximum strength of the risk weighting.

[0075] Based on these preset parameters, specific mapping functions can be constructed. For the first mapping function, a simple and effective choice is a linear mapping, normalizing its output range to [-1, 1]. Its formula can be: .when (At the most conservative) time, ;when (At its most radical) time, ;when When taking the median value, The transition is linear between -1 and 1. Thus, the final semantic action coefficient... The scale will be symmetrical with 1 as the center.

[0076] The purpose of the second mapping function is to map the risk score to a non-negative weighted intensity; therefore, its output range is typically [range missing]. A similar linear mapping function can be constructed as: .when (At the lowest risk) This means that no additional risk weights are applied; when (At the highest risk) Apply the highest risk weight. Thus, the final semantic risk coefficient... This approach involves non-negative enhancements based on the first approach. This explicitly functional implementation offers exceptional transparency and configurability. Engineers can precisely control the scope and intensity of semantic injection by adjusting the three hyperparameters N, M, and K, based on specific security requirements and operational experience, making the entire system easier to deploy and verify.

[0077] In addition, for dynamic scheduling factors related to training progress and The present invention also provides a preferred setting strategy. This strategy uses the entire training progress (which can be expressed as a percentage of the total number of training steps or the total number of cycles) as a percentage. The learning process is divided into three distinct phases: warm-up, stable, and annealing. This phased scheduling strategy aims to balance the relationship between prior knowledge guidance and data-driven learning.

[0078] Specifically, if the current training progress During the preheating period (e.g., ,in This is the end of the warm-up period, such as... Then the dynamic scheduling factor Adopt a linear growth strategy: .in, This is the preset maximum strength of semantic injection. This means that in the early stages of training, the strength of semantic guidance gradually increases from 0, which helps to guide the model's exploration direction and accelerate convergence by utilizing prior knowledge when the model parameters are not yet stable.

[0079] If training progress It has entered a stable period (e.g., ,in It is the end of the stable period, such as Then the dynamic scheduling factor Maintain at maximum value: At this stage, the model has already established a certain foundation. Maintaining the strongest semantic guidance can help the model fully learn and absorb domain knowledge, and form correct decision-making patterns.

[0080] Finally, when the training progress Entering the annealing period (e.g., Then the dynamic scheduling factor Using a cosine annealing strategy, the temperature smoothly decays from the maximum value to a preset minimum value. (Usually 0). Its formula is: In the later stages of training, when the model's performance is already high, gradually reducing the intensity of semantic guidance allows the model to rely more on data for fine-tuning and optimization, avoiding overfitting to prior knowledge and thus achieving better generalization performance. This three-stage scheduling strategy achieves an intelligent, adaptive knowledge injection process that mimics the pattern of human learning—the transition from strong guidance to autonomous learning—significantly improving the stability of model training and the final prediction accuracy.

[0081] In a core implementation, the training process of the risk-constrained reinforcement learning model is described in detail. This process is crucial for achieving risk control and performance optimization in this method. First, a dataset for training needs to be constructed. For each time step in the historical data... All of these can obtain their corresponding health representation vectors. The first semantic coefficient obtained through LLM inference Second semantic coefficient And the actual remaining lifetime value at that moment. These elements together constitute the basic unit of the training set.

[0082] Within the framework of reinforcement learning, it is necessary to define a "state". This embodiment uses the health representation vector... First semantic coefficient Second semantic coefficient By fusing the data, an engine state representation for the risk-constrained reinforcement learning model is constructed. A simple way to merge them is through vector concatenation, i.e. ,in This indicates a concatenation operation. This state... A comprehensive description of the engine in Risk / strategy assessment at the physical state and semantic level at any given moment.

[0083] Next, the state will be represented. The input is fed into the policy network (ActorNetwork) of the risk-constrained reinforcement learning model. The policy network outputs an initial distribution of remaining lifetime prediction parameters. For example, under the Gaussian distribution assumption, the output It can be a value that includes the mean. and logarithmic standard deviation The vector, i.e. .this This represents the model's initial prediction based on the current state. Then, the first semantic coefficients are used... For this initial distribution parameter Modulation is performed to generate modulated distributed parameters. A simple modulation method is element-wise multiplication, i.e. This step enables semantic knowledge to directly intervene in the prediction results. This is element-wise multiplication.

[0084] To guide model updates, a "reward" needs to be defined. This embodiment designs an immediate reward that is particularly sensitive to overestimation behavior. The calculation formula is as follows: .in, It is based on the modulated distribution parameters The calculated median of the RUL prediction. At quantile Time is equivalent to absolute error , It's about its weight. The key lies in the second item. It is an indicator function, when the predicted value Higher than the true value When an overestimation occurs, the function value is 1; otherwise, it is 0. It is a relatively large overestimation penalty coefficient. Therefore, once an overestimation occurs, the reward value will be subject to an additional, substantial penalty. The design of this reward function directly translates the engineering preference of "aversion to overestimation" into the optimization signal of the model.

[0085] To further enhance risk control, this embodiment also utilizes a second semantic coefficient. For instant rewards Risk-weighted calculation to generate weighted returns .For example, This means that when LLM determines that the current state has a high risk ( If the reward is relatively large, the corresponding reward (or penalty) will be amplified, thus taking on a larger weight in model parameter updates.

[0086] Finally, the optimization objective of the entire model is set as: maximizing the expected cumulative return while satisfying the Conditional Value at Risk (CVaR) constraint. This is a typical constrained optimization problem. This embodiment uses the Lagrange dual method to solve it. By introducing Lagrange multipliers, the constrained problem is transformed into an unconstrained minimax problem. Then, gradient descent (updating model parameters) and gradient ascent (updating Lagrange multipliers) are alternately iterated until the model converges. The technical effect of this complete training process is that it constructs an end-to-end learning loop that deeply integrates semantic information and risk preferences, so that while optimizing prediction accuracy, the model's behavior is always guided by both domain knowledge and strict risk constraints.

[0087] In a deeper implementation of the above optimization problem, the method of iteratively updating model parameters using Lagrange duality is specifically explained. First, the cumulative reward needs to be calculated along the reinforcement learning episode. This is based on the aforementioned weighted instantaneous reward... The trajectory reward used for strategy evaluation can be calculated. The calculation formula is as follows: .in, Represents a time from The initial training trajectory, where H is the preset trajectory horizon length. It is a preset discount factor ( This is used to balance the importance of immediate rewards and future rewards. It is a time step index.

[0088] Based on calculated trajectory rewards This allows us to construct a formalized constrained optimization problem. The goal of this problem is to find a set of optimal policy network parameters. This makes the strategy Expected return on the generated trajectory To maximize returns, a constraint regarding risk must be satisfied, namely, the condition for return lies in the value at risk. It must be no less than a preset risk performance threshold. Mathematically, this is represented as:

[0089] .

[0090] Here Indicates the worst The average return along the trajectory of the proportion. This constraint directly imposes a hard requirement on the model's performance in extreme cases.

[0091] To solve this constrained optimization problem, this embodiment employs the Lagrangian duality method, transforming it into an equivalent and more easily tractable maxima-minima problem of the Lagrangian function. The constructed Lagrangian function is as follows:

[0092] .in, It is a non-negative dual variable (or Lagrange multiplier). Intuitively, this function multiplies the original constraint term by... This is then added to the objective function, forming a new, unconstrained objective. The solution process becomes a two-layer optimization: the inner layer optimization is for a fixed... By maximizing To update strategy parameters Outer layer optimization is achieved by minimizing... To update the dual variable (Equivalent to gradient ascent on the dual problem).

[0093] In the actual training cycle, the system repeatedly performs the following steps: First, using the current strategy... The system interacts with the environment to collect a batch of trajectory data. Then, based on this data, the parameters of the policy network are updated using the gradient ascent method. To maximize the Lagrangian function (e.g., using policy gradient algorithms such as PPO). Then, based on the same set of data, the dual variable is updated using gradient ascent. Its update rules are roughly as follows: ,in It is the learning rate. If the CVaR constraint is violated (i.e. ),but This will increase, thus increasing the focus on the CVaR term and forcing the strategy to become more conservative; if the constraints are satisfied, The value will then decrease or remain at 0. This iterative process continues until the policy parameter... and dual variables All convergence conditions are met. The technical advantage of this method is that it provides a systematic, theoretically convergent algorithm for solving reinforcement learning problems with risk constraints, ensuring that the final policy not only has superior performance but also meets the preset safety standards in terms of risk level.

[0094] Accordingly, the process of using the model for prediction during the inference phase, i.e., after training, is also clarified. This process is similar to the forward propagation step in training, but uses the fixed, optimal model parameters. First, for the current time step... The aircraft engine system acquires real-time data through sensors, and after preprocessing and feature engineering, constructs a data structure containing health representation vectors. First semantic coefficient Second semantic coefficient Engine status indication It is worth noting that in the pure inference phase, since there is no training cycle... The concept of a dynamic scheduling factor used to calculate semantic coefficients. and They are usually given a fixed value, such as their value at the end of training, or simply set to 1.

[0095] Then, this real-time engine status is represented. The input is fed into the already trained risk-constrained reinforcement learning model. The model's policy network performs a forward computation, outputting an initial remaining lifetime prediction distribution parameter. .this It is the most direct quantitative judgment made by the model on the current state based on its experience learned from massive historical data and semantic knowledge.

[0096] Finally, in order to incorporate specific semantic guidance within the current context, the system will utilize state... The first semantic coefficients calculated together For the initial distribution parameters Modulation is performed to generate the modulated distributed parameters used in the final output. For example, if an LLM analyzes the trend summary to determine that the current degradation trend is gradual, a more optimistic forecasting strategy can be adopted. It will be higher, leading to This may cause the modulated distribution mean to be lower. Slightly increase. Conversely, if the trend is judged to be perilous, The value may be less than 1, thus lowering the mean of the prediction. This step ensures that each prediction is not only a reproduction of historical data, but also a deliberate result combined with real-time, interpretable semantic judgment. The technical effect of this implementation is that it clearly defines how to apply the model trained by this invention to real-world scenarios, demonstrating how semantic injection plays a role in the inference stage, making the final prediction results more dynamic, intelligent, and context-aware.

[0097] Finally, based on the modulated distribution parameters, the specific steps for calculating and outputting the point estimate, prediction interval, and risk index of the remaining life of the aero-engine are as follows:

[0098] The remaining lifetime is estimated using the modulated distribution parameters; the remaining lifetime prediction interval is calculated based on multiple quantiles in the modulated distribution parameters at a preset confidence level; the prediction result, including the remaining lifetime estimate and the remaining lifetime prediction interval, is output, along with a risk level indication associated with the second semantic score and / or the second semantic coefficient; and a complete audit chain, including trend summary text, first semantic score, second semantic score, first semantic coefficient, second semantic coefficient, modulated distribution parameters, and prediction result, is output.

[0099] Based on the same inventive concept, this invention also provides an aero-engine remaining life prediction device based on large language model semantic injection and risk constraints, as described in the following embodiments. Since the principle of solving the problem in the aero-engine remaining life prediction device based on large language model semantic injection and risk constraints is similar to that of the aero-engine remaining life prediction method based on large language model semantic injection and risk constraints, the implementation of the aero-engine remaining life prediction device based on large language model semantic injection and risk constraints can refer to the implementation of the aero-engine remaining life prediction method based on large language model semantic injection and risk constraints; repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0100] Figure 2 This is a structural block diagram of an aero-engine remaining life prediction device based on large language model semantic injection and risk constraints according to an embodiment of the present invention, as shown below. Figure 2 As shown, the system includes: a health representation vector construction module 201, a semantic coefficient inference module 202, a model training module 203, an output distribution parameter module 204, and a remaining lifespan estimation module 205. The structure is described below.

[0101] A health representation vector module 201 is constructed to acquire multivariable sensor time-series data and operating condition data of aero-engines, perform preprocessing and sliding window truncation to generate window sample sequences, extract time-series features from the window sample sequences to generate health representation vectors, and simultaneously extract trend features of key sensors.

[0102] The semantic coefficient reasoning module 202 is used to generate trend summary text based on the trend features and the working condition data, input the trend summary text into the first prompt word template and the second prompt word template respectively, perform semantic reasoning through the large language model, output the first semantic score and the second semantic score, and map the first semantic score to the first semantic coefficient and the second semantic score to the second semantic coefficient.

[0103] The model training module 203 is used to construct a training set based on the health representation vector, the first semantic coefficient and the second semantic coefficient, and to add the risk constraint term to the objective function of the risk constraint reinforcement learning model to train the risk constraint reinforcement learning model.

[0104] The output distribution parameter module 204 is used to input the health representation vector, the first semantic coefficient and the second semantic coefficient as engine state inputs to the trained risk constraint reinforcement learning model, output the remaining life prediction distribution parameter, and modulate the remaining life prediction distribution parameter through the first semantic coefficient to output the modulated distribution parameter.

[0105] The remaining life estimation module 205 is used to calculate and output the point estimate, prediction interval and risk index of the remaining life of the aero-engine based on the modulated distribution parameters.

[0106] In one embodiment, constructing a health representation vector module includes:

[0107] The extraction vector unit is used to input the window sample sequence into the time encoder to extract the health characterization vector, wherein the time encoder is used to extract features characterizing the global degradation state of the engine from the multivariable sensor time series data;

[0108] A key sensor unit is set up to select key sensors from all sensors installed on the aero-engine based on the physical correlation between the sensor and the engine failure mode and / or the mutual information index calculated based on the training set. ;

[0109] The trend extraction unit is used to extract the key sensor from the window sample sequence. Trend characteristics ,in, , This is the slope calculation function. The standard deviation of the key sensor on the training set is given by [reference to a specific sensor]. For the current moment, The length of the sliding window.

[0110] In one embodiment, constructing the health representation vector module further includes:

[0111] A symbolic representation unit is used to map the direction and intensity of change in the trend features into symbolic representations according to a preset discretization threshold, wherein the symbolic representations include symbols representing rising, falling, and remaining unchanged.

[0112] A mapping label unit is used to map the working condition data to corresponding working condition status labels;

[0113] The summary text generation unit is used to combine the symbolic representation with the working condition label according to a preset text template to generate the trend summary text in natural language form.

[0114] In one embodiment, the semantic coefficient inference module includes:

[0115] Output a first semantic scoring unit, used to input the trend summary text into the first prompt word template and output a first discrete semantic score. The first prompt word template is configured to guide the large language model to assess the current degradation trend from an engineering conservatism perspective, and the first discrete semantic score is positively correlated with the aggressiveness of the prediction strategy. The number of rating levels;

[0116] Output a second semantic scoring unit, used to input the trend summary text into the second prompt word template and output a second discrete semantic score. The second prompt word template is configured to guide the large language model to assess the current degradation trend from a security risk perspective, and the second discrete semantic score is positively correlated with the risk of overestimating remaining lifetime. The number of risk levels;

[0117] Mapping function building unit, used to construct the first mapping function Second mapping function , where the first mapping function and the second mapping function All are preset monotonic functions, the first mapping function The second mapping function is used to convert discrete semantic scores into action modulation mathematical coefficients. Used to convert discrete semantic scores into risk-weighted mathematical strengths;

[0118] Mapped to a second basic adjustment value, used through the first mapping function The first discrete semantic score Mapped to a first basic adjustment value, and through the second mapping function The second discrete semantic score Mapped to the second basic adjustment amount;

[0119] The parameter setting unit is used to set the first dynamic scheduling factor related to the training progress. Second dynamic scheduling factor ,in, This represents the current training cycle number.

[0120] The scheduling factor calculation unit is used to calculate the first dynamic scheduling factor. and the second dynamic scheduling factor The first basic adjustment amount and the second basic adjustment amount are scaled respectively to obtain the first semantic coefficient. Second semantic coefficient ,in, , .

[0121] In one embodiment, the model training unit is also used to set a first semantic score based on the safety standards of the aero-engine. Total number of discrete levels Second semantic score Total number of discrete levels Second mapping function The theoretical upper limit of the output value Based on the total number of discrete levels Construct the first mapping function Based on the total number of discrete levels and theoretical upper limit Construct the second mapping function .

[0122] In one embodiment, the scheduling factor calculation unit is further used to divide the training progress into a warm-up period, a stabilization period, and an annealing period; if the training progress... During the preheating period, the scheduling factor ,in, As a dynamic scheduling factor, The preset maximum strength, This represents the maximum training progress during the warm-up period; if the training progress... During the stable period, the scheduling factor If training progress During the annealing period, the scheduling factor ,in, scheduling factor The theoretical lower limit throughout the entire training process. This represents the maximum training progress during the annealing period, and .

[0123] In one embodiment, the model training module includes:

[0124] Construct training set units for each sample time step. t Obtain the corresponding health representation vector. First semantic coefficient Second semantic coefficient and actual remaining lifespan And construct a training set;

[0125] Feature fusion unit, used to fuse the health representation vector The first semantic coefficient and the second semantic coefficient The process involves fusing the data to construct an engine state representation for the risk-constrained reinforcement learning model. ,in, This involves concatenating vectors.

[0126] The remaining life prediction distribution parameter calculation unit is used to represent the engine state. The input is fed into the policy network of the risk-constrained reinforcement learning model to obtain the initial remaining lifetime prediction distribution parameters. ,in, For the engine state representation The mean of the Gaussian distribution, For the engine state representation The logarithmic standard deviation;

[0127] Distributed parameter modulation unit, used to transmit the first semantic coefficient Predicted distribution parameters of the initial remaining lifetime Modulation is performed to generate modulated distributed parameters. ,in, , This is element-wise multiplication;

[0128] Instant reward calculation unit, used to calculate based on actual remaining lifespan value and modulated distributed parameters Calculate instant rewards ,in, For probability quantiles, In order to be in The absolute error at that time Based on the loss weight, To overestimate the penalty coefficient, () is an indicator function. For the modulated distribution parameters The median estimate of the remaining lifespan predicted by calculation.

[0129] Weighting unit, used to pass the second semantic coefficient For instant rewards Risk-weighted calculation to generate weighted returns ;

[0130] The model training unit is used to construct the risk-constrained reinforcement learning model to solve the optimization problem of maximizing expected return and satisfying the conditional value at risk (CVaR) constraint, based on the weighted return. The parameters of the risk-constrained reinforcement learning model are iteratively updated until the risk-constrained reinforcement learning model converges, thus obtaining the trained risk-constrained reinforcement learning model.

[0131] In one embodiment, the model training unit is also used to pass the weighted reward. Generate a trajectory reward dataset for policy evaluation. ,in, , The training trajectory is defined by H, which is the preset trajectory field of view length. For time step index, The discount factor is preset; based on the trajectory reward dataset. Constructing constrained optimization problems ,in, For the parameters of the policy network, For expectation operator, For parameters Defined strategy function, For the condition of risk value, The risk performance threshold is used to transform the constrained optimization problem into a Lagrangian function maximization / minimization problem. ,in, As the dual variable; repeatedly perform the Lagrange function minimization problem and update the parameters of the policy network. and the dual variable Continue until the convergence condition is met.

[0132] In one embodiment, the output distribution parameter module includes:

[0133] Construct an engine state representation unit for use with the health representation vector The first semantic coefficient and the second semantic coefficient Constructing an engine state representation ;

[0134] Output distributed parameter unit, used to represent the engine state The input is fed into the trained risk-constrained reinforcement learning model, and the initial remaining life prediction distribution parameters are output. ;

[0135] Modulation distribution parameter unit, used to pass the first semantic coefficient Predicted distribution parameters of the initial remaining lifetime Modulation is performed to generate modulated distributed parameters. .

[0136] The embodiments of the present invention achieve the following technical effects:

[0137] By explicitly incorporating risk constraints into the objective function, the model effectively suppresses catastrophic overestimation that can easily occur at the end of the lifespan or during sudden changes in operating conditions, thus improving the safety of predictions. Secondly, the introduced trend summarization and semantic scoring mechanisms provide a clear, traceable, and auditable explanation path for the previously "black box" prediction process, enhancing user trust in the prediction results. Finally, the dynamic modulation of the model output by semantic coefficients makes predictions more flexible and adaptable to different scenarios, dynamically adjusting the conservatism of its prediction strategy based on real-time degradation trends and risk assessments.

[0138] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

[0139] 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 aero-engines based on semantic injection and risk constraints using a large language model, characterized in that, include: Acquire multivariable sensor time-series data and operating condition data of aero-engines, perform preprocessing and sliding window truncation to generate window sample sequences, extract time-series features from the window sample sequences to generate health characterization vectors, and simultaneously extract trend features of key sensors. Based on the trend features and the working condition data, a trend summary text is generated. The trend summary text is input into the first prompt word template and the second prompt word template respectively. Semantic reasoning is performed through a large language model to output a first semantic score and a second semantic score. The first semantic score is mapped to a first semantic coefficient and the second semantic score is mapped to a second semantic coefficient. A training set is constructed based on the health representation vector, the first semantic coefficient, and the second semantic coefficient, and a risk constraint term is added to the objective function of the risk constraint reinforcement learning model to train the risk constraint reinforcement learning model. Based on each sample time t Obtain the corresponding health representation vector. First semantic coefficient Second semantic coefficient and actual remaining lifespan And construct a training set; The health representation vector The first semantic coefficient and the second semantic coefficient The process involves fusing the data to construct an engine state representation for the risk-constrained reinforcement learning model. ,in, This involves concatenating vectors. Representing the engine state The input is fed into the policy network of the risk-constrained reinforcement learning model to obtain the initial remaining lifetime prediction distribution parameters. ,in, For the engine state representation The mean of the Gaussian distribution, For the engine state representation The logarithmic standard deviation; Through the first semantic coefficient Predicted distribution parameters of the initial remaining lifetime Modulation is performed to generate modulated distributed parameters. ,in, , This is element-wise multiplication; Based on actual remaining lifespan and modulated distributed parameters Calculate instant rewards ,in, For probability quantiles, In order to be in The absolute error at that time Based on the loss weight, To overestimate the penalty coefficient, () is an indicator function. For the modulated distribution parameters The median estimate of the remaining lifespan predicted by calculation. Through the second semantic coefficient For instant rewards Risk-weighted calculation to generate weighted returns ; To maximize expected return and satisfy the conditional value at risk (CVaR) constraint as the optimization objective, an optimization problem is constructed using the risk-constrained reinforcement learning model, based on the weighted return. The parameters of the risk-constrained reinforcement learning model are iteratively updated until the risk-constrained reinforcement learning model converges, thus obtaining the trained risk-constrained reinforcement learning model. The health representation vector, the first semantic coefficient, and the second semantic coefficient are used as engine state inputs to the trained risk-constrained reinforcement learning model, outputting the remaining life prediction distribution parameters, and modulating the remaining life prediction distribution parameters through the first semantic coefficient, outputting the modulated distribution parameters. Based on the modulated distribution parameters, the point estimate, prediction range, and risk index of the remaining life of the aero-engine are calculated and output.

2. The method for predicting the remaining life of aero-engines based on semantic injection and risk constraints of a large language model as described in claim 1, characterized in that, Temporal feature extraction is performed on the window sample sequence to generate a health representation vector, and trend features of key sensors are extracted simultaneously, including: The window sample sequence is input into the time encoder to extract the health representation vector, wherein the time encoder is used to extract features representing the global degradation state of the engine from the multivariable sensor time series data; Based on the physical correlation between the sensor and the engine failure mode and / or the mutual information index calculated based on the training set, key sensors are selected from all sensors installed on the aero-engine. ; Extract the key sensor from the window sample sequence. Trend characteristics ,in, , This is the slope calculation function. The standard deviation of the key sensor on the training set is given by [reference to a specific sensor]. For the current moment, The length of the sliding window.

3. The method for predicting the remaining life of aero-engines based on semantic injection and risk constraints of a large language model as described in claim 1, characterized in that, Based on the trend characteristics and the operating condition data, a trend summary text is generated, including: The direction and intensity of change in the trend features are mapped to symbolic representations according to a preset discretization threshold, wherein the symbolic representations include symbols for rising, falling, and remaining unchanged. Map the operating condition data to corresponding operating condition status labels; The symbolic representation and the operating condition label are combined according to a preset text template to generate the trend summary text in natural language form.

4. The method for predicting the remaining life of aero-engines based on semantic injection and risk constraints of a large language model as described in claim 1, characterized in that, The trend summary text is input into the first prompt word template and the second prompt word template, respectively. Semantic reasoning is performed through a large language model to output a first semantic score and a second semantic score. The first semantic score is mapped to a first semantic coefficient, and the second semantic score is mapped to a second semantic coefficient, including: Input the trend summary text into the first prompt word template and output the first discrete semantic score. The first prompt word template is configured to guide the large language model to assess the current degradation trend from an engineering conservatism perspective, and the first discrete semantic score is positively correlated with the aggressiveness of the prediction strategy. The number of rating levels; Input the trend summary text into the second prompt word template and output the second discrete semantic score. The second prompt word template is configured to guide the large language model to assess the current degradation trend from a security risk perspective, and the second discrete semantic score is positively correlated with the risk of overestimating remaining lifetime. The number of risk levels; Construct the first mapping function Second mapping function , where the first mapping function and the second mapping function All are preset monotonic functions, the first mapping function The second mapping function is used to convert discrete semantic scores into action modulation mathematical coefficients. Used to convert discrete semantic scores into risk-weighted mathematical strengths; Through the first mapping function The first discrete semantic score Mapped to a first basic adjustment value, and through the second mapping function The second discrete semantic score Mapped to the second basic adjustment amount; Set the first dynamic scheduling factor related to the training progress respectively. Second dynamic scheduling factor ,in, This represents the current training cycle number. Through the first dynamic scheduling factor and the second dynamic scheduling factor The first basic adjustment amount and the second basic adjustment amount are scaled respectively to obtain the first semantic coefficient. Second semantic coefficient ,in, , .

5. The method for predicting the remaining life of aero-engines based on semantic injection and risk constraints of a large language model as described in claim 4, characterized in that, Construct the first mapping function Second mapping function ,include: Based on the safety standards for aircraft engines, a first semantic score is set. Total number of discrete levels Second semantic score Total number of discrete levels Second mapping function The theoretical upper limit of the output value ; Based on the total number of discrete levels Construct the first mapping function ; Based on the total number of discrete levels and theoretical upper limit Construct the second mapping function .

6. The method for predicting the remaining life of aero-engines based on semantic injection and risk constraints of a large language model as described in claim 4, characterized in that, Set the first dynamic scheduling factor related to the training progress respectively. Second dynamic scheduling factor ,include: The training schedule is divided into a warm-up period, a stabilization period, and an annealing period. If training progress During the preheating period, the scheduling factor ,in, As a dynamic scheduling factor, The preset maximum strength, This represents the maximum training progress during the warm-up period; If training progress During the stable period, the scheduling factor ; If training progress During the annealing period, the scheduling factor ,in, scheduling factor The theoretical lower limit throughout the entire training process. This represents the maximum training progress during the annealing period, and .

7. The method for predicting the remaining life of aero-engines based on semantic injection and risk constraints of a large language model as described in claim 1, characterized in that, To maximize expected return and satisfy the conditional value at risk (CVaR) constraint as the optimization objective, an optimization problem is constructed using the risk-constrained reinforcement learning model, based on the weighted return. Iteratively update the parameters of the risk-constrained reinforcement learning model until the model converges, obtaining the trained risk-constrained reinforcement learning model, including: Through the weighted returns Generate a trajectory reward dataset for policy evaluation. ,in, , The training trajectory is defined by H, which is the preset trajectory field of view length. For time step index, This is a preset discount factor; Based on the trajectory reward dataset Constructing constrained optimization problems ,in, For the parameters of the policy network, For expectation operator, For parameters Defined strategy function, For the condition of risk value, This represents a risk performance threshold. The constrained optimization problem is transformed into a Lagrangian function maximization and minimization problem. ,in, As dual variables; Repeat the Lagrange function minima problem and update the parameters of the policy network. and the dual variable Continue until the convergence condition is met.

8. The method for predicting the remaining life of aero-engines based on semantic injection and risk constraints of a large language model as described in claim 1, characterized in that, The health representation vector, the first semantic coefficient, and the second semantic coefficient are used as engine state inputs to the trained risk-constrained reinforcement learning model to output remaining life prediction distribution parameters. The remaining life prediction distribution parameters are then modulated using the first semantic coefficient to output modulated distribution parameters, including: Through the health representation vector The first semantic coefficient and the second semantic coefficient Constructing an engine state representation ; Representing the engine state The input is fed into the trained risk-constrained reinforcement learning model, and the initial remaining life prediction distribution parameters are output. ; Through the first semantic coefficient Predicted distribution parameters of the initial remaining lifetime Modulation is performed to generate modulated distributed parameters. .

9. A device for predicting the remaining life of an aero-engine based on semantic injection and risk constraints using a large language model, characterized in that, include: A health representation vector module is constructed to acquire multivariable sensor time-series data and operating condition data of aero-engines, and to perform preprocessing and sliding window truncation to generate a window sample sequence. The time-series features of the window sample sequence are extracted to generate a health representation vector, and the trend features of key sensors are extracted simultaneously. The semantic coefficient reasoning module is used to generate trend summary text based on the trend features and the working condition data, input the trend summary text into the first prompt word template and the second prompt word template respectively, perform semantic reasoning through the large language model, output the first semantic score and the second semantic score, and map the first semantic score to the first semantic coefficient and the second semantic score to the second semantic coefficient. The model training module is used to construct a training set based on the health representation vector, the first semantic coefficient and the second semantic coefficient, and to add the risk constraint term to the objective function of the risk constraint reinforcement learning model to train the risk constraint reinforcement learning model. The model training module includes: Construct training set units for each sample time step. t Obtain the corresponding health representation vector. First semantic coefficient Second semantic coefficient and actual remaining lifespan And construct a training set; Feature fusion unit, used to fuse the health representation vector The first semantic coefficient and the second semantic coefficient The process involves fusing the data to construct an engine state representation for the risk-constrained reinforcement learning model. ,in, This involves concatenating vectors. The remaining life prediction distribution parameter calculation unit is used to represent the engine state. The input is fed into the policy network of the risk-constrained reinforcement learning model to obtain the initial remaining lifetime prediction distribution parameters. ,in, For the engine state representation The mean of the Gaussian distribution, For the engine state representation The logarithmic standard deviation; Distributed parameter modulation unit, used to transmit the first semantic coefficient Predicted distribution parameters of the initial remaining lifetime Modulation is performed to generate modulated distributed parameters. ,in, , This is element-wise multiplication; Instant reward calculation unit, used to calculate based on actual remaining lifespan value and modulated distributed parameters Calculate instant rewards ,in, For probability quantiles, In order to be in The absolute error at that time Based on the loss weight, To overestimate the penalty coefficient, () is an indicator function. For the modulated distribution parameters The median estimate of the remaining lifespan predicted by calculation. Weighting unit, used to pass the second semantic coefficient For instant rewards Risk-weighted calculation to generate weighted returns ; The model training unit is used to construct the risk-constrained reinforcement learning model to solve the optimization problem of maximizing expected return and satisfying the conditional value at risk (CVaR) constraint, based on the weighted return. The parameters of the risk-constrained reinforcement learning model are iteratively updated until the risk-constrained reinforcement learning model converges, thus obtaining the trained risk-constrained reinforcement learning model. The output distribution parameter module is used to input the health representation vector, the first semantic coefficient and the second semantic coefficient as engine state inputs to the trained risk constraint reinforcement learning model, output the remaining life prediction distribution parameter, and modulate the remaining life prediction distribution parameter through the first semantic coefficient to output the modulated distribution parameter. The remaining life estimation module is used to calculate and output the point estimate, prediction range, and risk index of the remaining life of the aero-engine based on the modulated distribution parameters.

Images