Aero-engine residual life parallel prediction method, system, equipment and medium

Abstract

The invention provides an aero-engine residual life parallel prediction method, system and device and a medium, and relates to the field of aero-engine health management.The aero-engine residual life parallel prediction method comprises the steps that sensor time sequence data of an aero-engine is obtained, and an input feature sequence is obtained; inputting the feature sequence into a continuous dynamic modeling path and a multi-resolution feature extraction path at the same time, and outputting a first residual useful life prediction value and a corresponding first uncertainty measure as well as a second residual useful life prediction value and a corresponding second uncertainty measure; and identifying the current flight condition of the engine, and based on an uncertainty perception fusion mechanism, according to the first uncertainty measure, the second uncertainty measure and the current flight condition, performing adaptive weighted fusion on the first prediction value of the remaining useful life and the second prediction value of the remaining useful life to obtain a prediction result of the remaining useful life. According to the method, continuous dynamic and multi-scale transient characteristics are captured at the same time, and high-precision and high-reliability prediction of the remaining useful life of the aero-engine is realized.

Description

Technical Field

[0001] This invention relates to the field of aero-engine health management technology, specifically to a method, system, device, and medium for parallel prediction of the remaining life of an aero-engine. Background Technology

[0002] Predicting the remaining useful life (RUL) of an aircraft engine is a core task of predictive and health management systems, and its accuracy is directly related to flight safety and maintenance costs. The engine degradation process is non-stationary, nonlinear, and multi-scale, including both slow, continuous degradation trends and high-frequency fluctuations caused by local anomalies or transient events.

[0003] Traditional methods for predicting the remaining life of aero-engines mainly fall into two categories: one is dynamic modeling based on continuous-time recurrent neural networks, which can capture continuous degradation processes but is insensitive to transient anomalies; the other is feature extraction methods based on wavelet multi-resolution analysis, which excels at capturing multi-scale local features but lacks the ability to model continuous dynamics. Therefore, existing methods struggle to simultaneously capture both continuous evolution and transient patterns, leading to inaccurate predictions. Furthermore, most methods fail to consider the uncertainties in the prediction results, resulting in large prediction biases and poor robustness under complex operating conditions. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a parallel prediction method, system, device and medium for the remaining useful life of aero-engines. The present invention can simultaneously capture continuous dynamic and multi-scale transient features to achieve high-precision and high-reliability prediction of the remaining useful life of aero-engines.

[0005] This application provides the following technical solution: a parallel prediction method for the remaining life of an aero-engine, comprising:

[0006] The sensor time-series data of the aero-engine is acquired and preprocessed to obtain the input feature sequence; wherein, the sensor time-series data includes at least: total fan inlet temperature, total high-pressure compressor outlet pressure, total high-pressure turbine outlet temperature, fuel flow rate, high-pressure rotor speed, low-pressure rotor speed, exhaust temperature, and vibration amplitude, with a sampling frequency of 1Hz, and the original time-series data is divided into time-series segments of fixed length using a sliding window; The input feature sequence is simultaneously input into the continuous dynamic modeling path and the multi-resolution feature extraction path for parallel processing; The continuous dynamic modeling path models the continuous degradation process of the engine based on a liquid neural network, and outputs a first predicted value of the remaining useful life and a corresponding first uncertainty measure; the multi-resolution feature extraction path performs multi-scale decomposition and feature extraction on the input feature sequence based on a wavelet multi-resolution mixer, and outputs a second predicted value of the remaining useful life and a corresponding second uncertainty measure. Based on the altitude, Mach number, throttle lever angle, fuel flow rate, and high-low pressure rotor speed difference in the input feature sequence, the current flight condition of the aero engine is identified in real time. The flight condition includes: ground idle, takeoff and climb, cruise, maneuvering flight, and approach and landing. Based on the uncertainty perception fusion mechanism, the first prediction value of the remaining useful life and the second prediction value of the remaining useful life are adaptively weighted and fused according to the first uncertainty metric, the second uncertainty metric and the current flight condition of the aero-engine, to generate the final prediction value of the remaining useful life and the corresponding comprehensive uncertainty metric. Based on the magnitude of the comprehensive uncertainty metric, the prediction reliability level is determined, and according to the preset reliability level-decision rule mapping, prediction and maintenance decision suggestions of the corresponding level are output.

[0007] According to one embodiment of the present invention, the continuous dynamic modeling path specifically includes: Using the input feature sequence as input, a dynamic model of ordinary differential equations for the hidden state is constructed using a continuous-time recurrent neural network; The fourth-order Runge-Kutta method is used to numerically integrate the ordinary differential equations to obtain the hidden state evolution trajectory in continuous time. Temporal pooling is performed on the hidden state evolution trajectory to generate path feature representation; Based on the path feature representation, the first predicted value of the remaining useful lifetime and the first uncertainty measure are obtained by predicting the output layer with Gaussian parameterization. The remaining useful life is defined as the time remaining from the current moment to the engine failure threshold, which includes the exhaust temperature exceeding the preset red line value or the high-pressure rotor speed dropping to the design lower limit.

[0008] According to one embodiment of the present invention, the multi-resolution feature extraction path specifically includes: The input feature sequence is subjected to multi-level discrete wavelet transform to decompose the sensor time-series signal into low-frequency approximation coefficients and high-frequency detail coefficients. The approximation coefficients correspond to the slow trend component in the engine degradation process, and the detail coefficients correspond to the transient event component. For the approximation coefficients and detail coefficients obtained from each level of decomposition, feature transformation is performed through independent neural network modules to obtain the resolution features at each level. All resolution features are aggregated, and the second predicted value of the remaining useful life and the second uncertainty measure are obtained by predicting through a Gaussian parameterized regression output layer.

[0009] According to one embodiment of the present invention, the uncertainty-aware fusion mechanism employs a precision-weighted mechanism, specifically including: Calculate the reciprocals of the first uncertainty measure and the second uncertainty measure, and use them as the first precision weight and the second precision weight, respectively; dynamically adjust the first precision weight and the second precision weight according to the current flight conditions of the aero-engine to obtain the condition-adaptive fusion weight; Based on the fusion weights, the first predicted value of remaining useful life and the second predicted value of remaining useful life are subjected to condition-aware adaptive weighted fusion to generate the final predicted value of remaining useful life and the corresponding comprehensive uncertainty measure.

[0010] According to one embodiment of the present invention, it further includes: Construct a prediction model that includes a continuous dynamic modeling subnetwork, a multi-resolution feature extraction subnetwork, a working condition recognition subnetwork, an uncertainty perception fusion module, and a decision rule mapping module; The prediction model is trained using a labeled training dataset. The loss function used in the training process includes a working condition identification loss term, a decision consistency loss term, a prediction error loss term, and a monotonicity regularization loss term. The monotonicity regularization loss term is used to constrain the monotonically decreasing trend of the remaining useful life prediction result with running time.

[0011] According to one embodiment of the present invention, the discrete wavelet transform used in the wavelet multiresolution mixer is one of Daubechies wavelet, Haar wavelet or Symlet wavelet.

[0012] According to one embodiment of the present invention, the liquid neural network uses continuous-time network units based on a gating mechanism.

[0013] The present invention also provides a parallel prediction system for the remaining life of an aero-engine, comprising: The data acquisition module is used to acquire and preprocess the time-series sensor data of the aero-engine to obtain the input feature sequence. The time-series sensor data includes at least: total fan inlet temperature, total high-pressure compressor outlet pressure, total high-pressure turbine outlet temperature, fuel flow rate, high-pressure rotor speed, low-pressure rotor speed, exhaust temperature, and vibration amplitude. The sampling frequency is 1Hz, and the original time-series data is divided into fixed-length time-series segments using a sliding window. The continuous dynamic modeling and multi-resolution feature extraction module is used to simultaneously input the input feature sequence into the continuous dynamic modeling path and the multi-resolution feature extraction path for parallel processing. The continuous dynamic modeling path models the continuous degradation process of the engine based on a liquid neural network, and outputs a first predicted value of the remaining useful life and a corresponding first uncertainty measure; the multi-resolution feature extraction path performs multi-scale decomposition and feature extraction on the input feature sequence based on a wavelet multi-resolution mixer, and outputs a second predicted value of the remaining useful life and a corresponding second uncertainty measure. The operating condition identification module is used to identify the current flight operating condition of the aero engine in real time based on the altitude, Mach number, throttle lever angle, fuel flow, and high-low pressure rotor speed difference in the input feature sequence. The flight operating conditions include: ground idle, takeoff and climb, cruise, maneuvering flight, and approach and landing. The fusion module is used to perform condition-aware adaptive weighted fusion of the first predicted value of the remaining useful life and the second predicted value of the remaining useful life based on the uncertainty-aware fusion mechanism, according to the first uncertainty metric, the second uncertainty metric and the current flight condition of the aero-engine, to generate the final predicted value of the remaining useful life and the corresponding comprehensive uncertainty metric. The hierarchical uncertainty decision module is used to determine the prediction reliability level based on the magnitude of the comprehensive uncertainty metric, and output prediction and maintenance decision suggestions corresponding to the level according to the preset reliability level-decision rule mapping.

[0014] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described parallel prediction method for the remaining life of an aero-engine.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that performs the above-described parallel prediction method for the remaining life of an aero-engine.

[0016] Compared with existing technologies, the beneficial effects achieved by at least one of the above-mentioned technical solutions adopted in the embodiments of this specification include at least the following: The embodiments of this invention propose a parallel dual-path prediction architecture, including a liquid neural network path and a wavelet multi-resolution mixer path, and intelligently fuse the outputs of the two paths through an uncertainty-aware fusion mechanism. The liquid neural network path models the continuous degradation process of the engine through a continuous-time dynamic system; the wavelet multi-resolution mixer path extracts degradation features at different time scales through discrete wavelet transform. The two paths output prediction results and their uncertainties respectively, and finally fuse them based on accuracy weighting to obtain an accurate and reliable remaining useful life prediction. The embodiments of this invention, through parallel dual-path design, simultaneously capture continuous dynamic and multi-scale transient features, significantly improving prediction accuracy. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a schematic diagram of the parallel prediction method for the remaining life of an aero-engine according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the parallel prediction system for the remaining life of an aero-engine according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the computer device of the present invention. Detailed Implementation

[0019] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0020] 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.

[0021] The key terms involved in the embodiments of this invention are defined as follows: LNN (Liquid Neural Network): A continuous-time dynamic system used to model non-stationary degrading trajectories.

[0022] WRM (Wavelet Multi-Resolution Mixer): A wavelet multi-resolution mixer that captures scale-specific degradation patterns through DWT decomposition.

[0023] CTRNN (Continuous-Time Recurrent Neural Network): The core of LNNs, used for hidden state evolution.

[0024] DWT (Discrete Wavelet Transform): This transforms a sequence into approximation and detail coefficients.

[0025] Uncertainty-Aware Fusion: This term refers to a fusion method that uses precision-weighted fusion to predict outputs in parallel.

[0026] Gaussian parameterization: Predicts outputs that follow a normal distribution in terms of mean and variance.

[0027] Temporal Pooling: Extracts the final representation from the hidden state, such as the final state or attention pooling.

[0028] Approximation Coefficients: These are the low-frequency components of wavelet decomposition, capturing slowly changing parameters.

[0029] Detail Coefficients: These are the high-frequency components of wavelet decomposition, used to capture transient anomalies.

[0030] Precision Weighting: A fusion mechanism based on the inverse of variance.

[0031] like Figure 1 As shown, this embodiment of the invention provides a parallel prediction method for the remaining life of an aero-engine, including: S101. Acquire and preprocess the time-series sensor data of the aero-engine to obtain the input feature sequence; The sensor time-series data includes, but is not limited to: total fan inlet temperature (T2), total fan inlet pressure (P2), total low-pressure compressor outlet temperature (T24), total high-pressure compressor outlet pressure (P30), total high-pressure turbine outlet temperature (T50), total high-pressure turbine outlet pressure (P50), fuel flow rate (Wf), high-pressure rotor speed (N2), low-pressure rotor speed (N1), exhaust temperature (EGT), combustion chamber pressure (Pcomb), vibration amplitude (Vib), lubricating oil temperature (OilTemp), and lubricating oil pressure (OilPress). The sensor data sampling frequency is uniformly 1 Hz. A sliding window technique is used to divide the original time-series data into fixed-length time series segments, with a window length of 50 time steps and a step size of 1, which serve as the input feature sequence for the model.

[0032] Preprocessing includes: missing value imputation (using linear interpolation or forward filling), outlier detection and removal based on the 3σ criterion, Z-score normalization, and alignment of multi-sensor data into a multi-dimensional time-series vector.

[0033] S102. The input feature sequence is simultaneously input into the continuous dynamic modeling path and the multi-resolution feature extraction path for parallel processing; The continuous dynamic modeling path models the continuous degradation process of the engine based on a liquid neural network, and outputs a first predicted value of the remaining useful life and a corresponding first uncertainty measure; the multi-resolution feature extraction path performs multi-scale decomposition and feature extraction on the input feature sequence based on a wavelet multi-resolution mixer, and outputs a second predicted value of the remaining useful life and a corresponding second uncertainty measure. S103. Based on the altitude, Mach number, throttle lever angle, fuel flow rate, and high-low pressure rotor speed difference in the input feature sequence, identify the current flight condition of the aero-engine in real time; The flight conditions include: ground idle, takeoff and climb, cruise, maneuvering flight, and approach and landing. S104. Based on the uncertainty perception fusion mechanism, according to the first uncertainty measure, the second uncertainty measure and the current flight condition of the aero-engine, the first predicted value of the remaining useful life and the second predicted value of the remaining useful life are subjected to adaptive weighted fusion based on the condition perception to generate the final predicted value of the remaining useful life and the corresponding comprehensive uncertainty measure. S105. Based on the magnitude of the comprehensive uncertainty metric, determine the prediction reliability level, and output prediction and maintenance decision suggestions corresponding to the level according to the preset reliability level-decision rule mapping.

[0034] This invention proposes a parallel dual-path fusion architecture that no longer relies on a single model. Instead, it designs two independent paths with complementary functions: a continuous dynamic modeling path and a multi-resolution feature extraction path, which respectively capture different aspects of the degradation process. Specifically, the system receives pre-processed sensor time-series data in parallel. One path constructs a continuous-time dynamic system using a liquid neural network (LNN) to simulate the macroscopic evolution trajectory of degradation. The other path performs multi-scale decomposition of the signal using a wavelet multi-resolution mixer (WRM) to capture local degradation patterns and anomalies in different frequency bands. Each path outputs not only point prediction values ​​but also a measure of the uncertainty of the prediction. Finally, the prediction results of the two paths are organically integrated through an uncertainty-aware intelligent fusion module.

[0035] In some embodiments of the present invention, the continuous dynamic modeling path specifically includes: Using the input feature sequence as input, a dynamic model of ordinary differential equations for the hidden state is constructed using a continuous-time recurrent neural network; The fourth-order Runge-Kutta method is used to numerically integrate the ordinary differential equations to obtain the hidden state evolution trajectory in continuous time. Temporal pooling is performed on the hidden state evolution trajectory to generate path feature representation; Based on the path feature representation, the first predicted value of the remaining useful lifetime and the first uncertainty measure are obtained by predicting the output layer with Gaussian parameterization. The remaining useful life is defined as the time remaining from the current moment to the engine failure threshold, which includes the exhaust temperature exceeding the preset red line value or the high-pressure rotor speed dropping to the design lower limit.

[0036] The embodiments of this invention employ a continuous-time recurrent neural network, whose hidden state changes are described by ordinary differential equations. The network uses an ordinary differential equation solver to process sequence data of variable length and unequal intervals, and has a strong ability to model long-term dependencies.

[0037] In practice, the continuous dynamic modeling path is based on a Liquid Neural Network (LNN) to model the continuous degradation process of the engine. The LNN uses continuous-time network units based on a gating mechanism, taking sensor feature sequences as input, to construct the continuous evolution process of the hidden state over time. The changes in the hidden state are described by the following ordinary differential equation:

[0038] Where h(t) is the hidden state vector, x(t) is the sensor input vector, f is the nonlinear function parameterized by the neural network, and θ is a learnable parameter; the temporal pooling uses the hidden state at the last time step as the path feature representation. The fourth-order Runge-Kutta method is used to numerically integrate the ordinary differential equation to obtain the evolution trajectory of the hidden state over continuous time. Temporal pooling is then performed on the hidden state evolution trajectory, using the hidden state at the last time step as the path feature representation. Based on the path feature representation, the first predicted value μ1 of the remaining useful life and the corresponding first uncertainty metric σ1 are predicted through a Gaussian-parameterized output layer, i.e., the predicted output follows a normal distribution N(μ1, σ1²). The remaining useful life is defined as the time remaining from the current moment to the engine failure threshold, which includes exhaust temperature exceeding a preset redline value, such as 950℃, or high-pressure rotor speed dropping to the design lower limit, such as 95% of the rated speed.

[0039] In some embodiments of the present invention, the multi-resolution feature extraction path specifically includes: The input feature sequence is subjected to multi-level discrete wavelet transform to decompose the sensor time-series signal into low-frequency approximation coefficients and high-frequency detail coefficients. The approximation coefficients correspond to the slow trend component in the engine degradation process, and the detail coefficients correspond to the transient event component. For the approximation coefficients and detail coefficients obtained from each level of decomposition, feature transformation is performed through independent neural network modules to obtain the resolution features at each level. All resolution features are aggregated, and the second predicted value of the remaining useful life and the second uncertainty measure are obtained by predicting through a Gaussian parameterized regression output layer.

[0040] This invention employs Discrete Wavelet Transform (DWT) as the front end to decompose the original signal into different resolutions without redundancy. Next, an independent lightweight neural network is designed for feature learning of the sub-band coefficients at each resolution, enabling the model to explicitly and scale-wise focus on slowly changing background trends (low-frequency approximation coefficients) and rapidly occurring transient events (high-frequency detail coefficients).

[0041] In specific implementation, the multi-resolution feature extraction path is based on a wavelet multi-resolution mixer (WRM) to perform multi-scale decomposition and feature extraction on the input feature sequence. A three-level discrete wavelet transform using Daubechies 4 wavelets is employed to decompose the sensor time-series signal into low-frequency approximation coefficients and high-frequency detail coefficients. The approximation coefficients correspond to the slow trend components in the engine degradation process, including performance degradation and thermomechanical fatigue accumulation, while the detail coefficients correspond to transient event components, including combustion chamber instability, blade vibration, and fuel system fluctuations. For each level of decomposition, the approximation coefficients and detail coefficients are transformed using a two-layer multilayer perceptron (MLP) with a hidden layer dimension of 32 and a ReLU activation function, yielding features at each resolution level. All resolution features are concatenated and aggregated, and the second predicted value μ2 of the remaining useful life and the corresponding second uncertainty metric σ2 are output through a fully connected layer.

[0042] In some embodiments of the present invention, the uncertainty-aware fusion mechanism employs a precision-weighted mechanism, specifically including: calculating the reciprocals of the first uncertainty metric and the second uncertainty metric, respectively serving as the first precision weight and the second precision weight; dynamically adjusting the first precision weight and the second precision weight according to the current flight conditions of the aero-engine to obtain condition-adaptive fusion weights; and performing condition-aware adaptive weighted fusion of the first predicted value of the remaining useful life and the second predicted value of the remaining useful life based on the fusion weights to generate the final predicted value of the remaining useful life and the corresponding comprehensive uncertainty metric. Specifically, the final predicted value of the remaining useful life is obtained by multiplying the first predicted value of the remaining useful life by the first precision fusion weight, adding the product of the second predicted value of the remaining useful life by the second precision fusion weight, and then dividing by the sum of the precision fusion weights.

[0043] The predictive reliability of two paths can dynamically change under different operating conditions or degradation stages due to different modeling focuses. This invention introduces an uncertainty-aware fusion mechanism. Specifically, the reciprocal of the uncertainty metric output by each path is used as the accuracy weight of that path's prediction. Then, based on the current flight conditions of the aero-engine, the accuracy weights of the two path predictions are dynamically adjusted to obtain condition-adaptive fusion weights. By analyzing features such as altitude, Mach number, throttle lever angle, fuel flow rate, and high / low pressure rotor speed difference in the input feature sequence in real time, a one-dimensional convolutional neural network or temporal attention network is used to identify the specific flight conditions or load states of the engine. These flight conditions include: ground idle, takeoff climb, cruise, maneuvering flight, and approach / landing. The weight allocation in the fusion module considers both prediction uncertainty and the current operating conditions, dynamically adjusting the contribution of the two paths. During fusion, the prediction result of the path with higher accuracy is given a larger weight. This mechanism allows the system to adaptively rely on the more reliable path in a specific situation, thereby improving overall robustness. This embodiment solves the problem of the engine degradation rate dynamically changing with operating conditions by dynamically adjusting the fusion weights based on the current flight conditions. For example, during stable cruise phases, the continuous dynamic path is more relied upon; during high-transient phases of intense maneuvering, the weight of the multi-resolution path is automatically increased. This upgrades the model from static fusion to context-aware dynamic fusion, significantly improving prediction accuracy and adaptability under complex flight profiles. The uncertainty-aware fusion mechanism in this embodiment enables the model to dynamically adjust its dependence on the two paths based on the characteristics of the current input data. For example, during stable degradation phases, the continuous dynamic path may be more reliable; when sudden anomalies occur, the role of the multi-resolution path is enhanced. This adaptive capability makes the model stable under varying operating conditions. This invention, through parallel dual-path design, simultaneously captures continuous evolution trends and multi-scale transient features, achieving information complementarity. Experiments show that on publicly available benchmark datasets, the RMSE index of the method of this invention is improved by more than 25% compared to the traditional single model.

[0044] In some embodiments of the present invention, the method further includes: constructing a prediction model comprising a continuous dynamic modeling subnetwork, a multi-resolution feature extraction subnetwork, a working condition identification subnetwork, an uncertainty perception fusion module, and a decision rule mapping module; training the prediction model using a labeled training dataset, wherein the loss function used during training includes a working condition identification loss term, a decision consistency loss term, a prediction error loss term, and a monotonicity regularization loss term, wherein the monotonicity regularization loss term is used to constrain the monotonically decreasing trend of the remaining useful life prediction result with running time.

[0045] In specific implementation, the discrete wavelet transform used in the wavelet multiresolution mixer is one of the Daubechies wavelet, Haar wavelet, or Symlet wavelet. The liquid neural network uses continuous-time network units based on a gating mechanism to improve computational efficiency.

[0046] In some embodiments of the present invention, the real-time identification of the current flight condition of the aero-engine includes: Extract multidimensional features, including at least altitude, airspeed, thrust, and vibration amplitude, from the input feature sequence; input the multidimensional features into a pre-trained operational condition recognition sub-network, and output the current discrete operational condition category, which includes at least ground idle, takeoff and climb, cruise, maneuvering flight, and approach and landing.

[0047] In some embodiments of the present invention, the working condition identification subnetwork is implemented by a convolutional neural network or a temporal attention mechanism, and is trained under the supervision of multiple working condition labels recorded from aviation flight data.

[0048] The magnitude of the comprehensive uncertainty metric determines the prediction reliability level, including: The comprehensive uncertainty measure is compared with three preset confidence thresholds: high, medium, and low. If the reliability level is below the low threshold, it is determined to be high reliability; if it is between the low and high thresholds, it is determined to be medium reliability; if it is above the high threshold, it is determined to be low reliability.

[0049] The reliability level-decision rule mapping includes: When the reliability level is high, the output is the estimated remaining useful life point and the narrowband confidence interval, and the maintenance is recommended according to the plan; When the reliability level is medium, the output provides an estimate of the remaining useful life range and suggestions to shorten the monitoring cycle and add additional inspection items. When the reliability level is low, a real-time warning signal is triggered, a conservative lifespan lower limit estimate is output, and it is recommended to immediately conduct ground inspections or replace components.

[0050] This embodiment not only outputs the final RUL prediction value but also constructs a hierarchical uncertainty decision-making framework. First, based on the first and second uncertainty measures, a meta-judgment of the current prediction reliability level is generated, such as: high confidence, medium confidence, and low confidence. Then, according to this reliability level, the system adaptively selects prediction output strategies with different levels of conservatism, such as: directly outputting point estimates, outputting pessimistic interval estimates, or triggering manual check alarms.

[0051] In one specific embodiment, the parallel prediction method for the remaining life of an aero-engine includes the following steps: 1. A subset of NASA's C-MAPSS dataset FD001 was used. This dataset contains time-series sensor data for 100 engine units, each running from its initial healthy state to failure, recording data from 21 sensor channels. Fourteen degradation-related sensor channels were selected: total fan inlet temperature (T2), total fan inlet pressure (P2), total low-pressure compressor outlet temperature (T24), total high-pressure compressor outlet pressure (P30), total high-pressure turbine outlet temperature (T50), total high-pressure turbine outlet pressure (P50), fuel flow rate (Wf), high-pressure rotor speed (N2), low-pressure rotor speed (N1), exhaust gas temperature (EGT), combustion chamber pressure (Pcomb), vibration amplitude (Vib), lubricating oil temperature (OilTemp), and lubricating oil pressure (OilPress). The raw sensor data was Z-score normalized, and a sliding window was used to generate sequence samples with a window length of 50 time steps and an input dimension of 14×50.

[0052] 2. Model Building: Continuous Dynamic Path: A single-layer CTRNN with a hidden layer dimension of 64 is used. The fourth-order Runge-Kutta method is employed for numerical integration of ordinary differential equations, with a step size of 0.1. Temporal pooling utilizes the hidden state of the last time step. The output layer is Gaussian parameterized, outputting the mean μ1 and variance σ1².

[0053] Multi-resolution path: Daubechies 4 (db4) wavelet is used for 3-level discrete wavelet transform. The approximation coefficients and detail coefficients of each level are processed by a two-layer MLP (32-dimensional, ReLU activation). Finally, the features of each level are concatenated and the mean μ2 and variance σ2² are output through a fully connected layer.

[0054] 3. Operating Condition Recognition: Based on the input feature sequence, the current flight operating condition of the aero-engine is identified in real time. Specifically, a one-dimensional convolutional neural network is used, containing two convolutional layers (with kernel sizes of 5 and 3, and channel numbers of 32 and 64, respectively) and a fully connected layer, outputting the Softmax probability distribution of 5 operating conditions.

[0055] 4. Fusion Module: The output layers of both paths are Gaussian parameterized and dynamically adjusted based on the current flight conditions, outputting the mean μ and variance σ. 2 During fusion, follow the formula. Calculate the final predicted value.

[0056] Training: The model is trained using the Adam optimizer.

[0057] On the test set, the method of this invention achieves a prediction accuracy of 8.5 RMSE, which is significantly better than using continuous dynamic path alone (RMSE: 16.2) or multi-resolution path alone (RMSE: 12.1).

[0058] like Figure 2 As shown, a parallel prediction system 200 for the remaining life of an aero-engine is characterized by comprising: The data acquisition module 201 is used to acquire and preprocess the sensor time-series data of the aero-engine to obtain the input feature sequence. The sensor time-series data includes at least: total fan inlet temperature, total high-pressure compressor outlet pressure, total high-pressure turbine outlet temperature, fuel flow rate, high-pressure rotor speed, low-pressure rotor speed, exhaust temperature, and vibration amplitude. The sampling frequency is 1Hz, and the original time-series data is divided into fixed-length time-series segments using a sliding window. The continuous dynamic modeling and multi-resolution feature extraction module 202 is used to simultaneously input the input feature sequence into the continuous dynamic modeling path and the multi-resolution feature extraction path for parallel processing. The continuous dynamic modeling path models the continuous degradation process of the engine based on a liquid neural network, and outputs a first predicted value of the remaining useful life and a corresponding first uncertainty measure; the multi-resolution feature extraction path performs multi-scale decomposition and feature extraction on the input feature sequence based on a wavelet multi-resolution mixer, and outputs a second predicted value of the remaining useful life and a corresponding second uncertainty measure. The operating condition identification module 203 is used to identify the current flight operating condition of the aero engine in real time based on the altitude, Mach number, throttle lever angle, fuel flow, and high-low pressure rotor speed difference in the input feature sequence. The flight operating conditions include: ground idle, takeoff climb, cruise, maneuvering flight, and approach and landing. The fusion module 204 is used to perform condition-aware adaptive weighted fusion of the first predicted value of the remaining useful life and the second predicted value of the remaining useful life based on the uncertainty-aware fusion mechanism, according to the first uncertainty metric, the second uncertainty metric and the current flight condition of the aero-engine, to generate the final predicted value of the remaining useful life and the corresponding comprehensive uncertainty metric. The hierarchical uncertainty decision module 205 is used to determine the prediction reliability level based on the magnitude of the comprehensive uncertainty metric, and output prediction and maintenance decision suggestions corresponding to the level according to the preset reliability level-decision rule mapping.

[0059] In specific implementation, this system also includes a model building and training module, which is used to build a prediction model that includes a continuous dynamic modeling subnetwork, a multi-resolution feature extraction subnetwork, a working condition recognition subnetwork, an uncertainty perception fusion module, and a decision rule mapping module. The prediction model is trained using a labeled training dataset. The loss function used in the training process includes a working condition recognition loss term, a decision consistency loss term, a prediction error loss term, and a monotonicity regularization loss term. The monotonicity regularization loss term is used to constrain the monotonically decreasing trend of the remaining useful life prediction result with running time.

[0060] The prediction system architecture of this invention combines a liquid neural network (continuous dynamic modeling) with a wavelet multi-resolution mixer (multi-scale feature extraction) in parallel. It utilizes the uncertainty metric of each path output and fuses the prediction results of the two paths in an adaptive weighted manner based on operating conditions. This solves the problem that existing single models cannot simultaneously take into account the continuous dynamics and multi-scale transients of the engine degradation process, and generally lack the quantification and utilization of the uncertainty of the prediction results, resulting in insufficient prediction accuracy under complex operating conditions.

[0061] In one embodiment, a computer device is provided, such as Figure 3 As shown, it includes a memory 301, a processor 302, and a computer program stored in the memory 301 and executable on the processor 302. When the processor 302 executes the computer program, it implements the above-mentioned parallel prediction method for the remaining life of aero-engines.

[0062] Specifically, the computer device can be a computer terminal, a server, or a similar computing device.

[0063] In this embodiment, a computer-readable storage medium is provided, which stores a computer program that performs the above-described parallel prediction method for the remaining life of an aero-engine.

[0064] Specifically, computer-readable storage media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable storage media does not include transient media, such as modulated data signals and carrier waves.

[0065] 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.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A parallel prediction method for the remaining life of an aero-engine, characterized in that, include: The sensor time-series data of the aero-engine is acquired and preprocessed to obtain the input feature sequence; wherein, the sensor time-series data includes at least: total fan inlet temperature, total high-pressure compressor outlet pressure, total high-pressure turbine outlet temperature, fuel flow rate, high-pressure rotor speed, low-pressure rotor speed, exhaust temperature, and vibration amplitude, with a sampling frequency of 1Hz, and the original time-series data is divided into time-series segments of fixed length using a sliding window; The input feature sequence is simultaneously input into the continuous dynamic modeling path and the multi-resolution feature extraction path for parallel processing; The continuous dynamic modeling path models the continuous degradation process of the engine based on a liquid neural network, and outputs a first predicted value of the remaining useful life and a corresponding first uncertainty measure; the multi-resolution feature extraction path performs multi-scale decomposition and feature extraction on the input feature sequence based on a wavelet multi-resolution mixer, and outputs a second predicted value of the remaining useful life and a corresponding second uncertainty measure. Based on the altitude, Mach number, throttle lever angle, fuel flow rate, and high-low pressure rotor speed difference in the input feature sequence, the current flight condition of the aero engine is identified in real time. The flight condition includes: ground idle, takeoff and climb, cruise, maneuvering flight, and approach and landing. Based on the uncertainty perception fusion mechanism, the first prediction value of the remaining useful life and the second prediction value of the remaining useful life are adaptively weighted and fused according to the first uncertainty metric, the second uncertainty metric and the current flight condition of the aero-engine, to generate the final prediction value of the remaining useful life and the corresponding comprehensive uncertainty metric. Based on the magnitude of the comprehensive uncertainty metric, the prediction reliability level is determined, and according to the preset reliability level-decision rule mapping, prediction and maintenance decision suggestions of the corresponding level are output.

2. The parallel prediction method for the remaining life of an aero-engine according to claim 1, characterized in that, The continuous dynamic modeling path specifically includes: Using the input feature sequence as input, a dynamic model of ordinary differential equations for the hidden state is constructed using a continuous-time recurrent neural network; The fourth-order Runge-Kutta method is used to numerically integrate the ordinary differential equations to obtain the hidden state evolution trajectory in continuous time. Temporal pooling is performed on the hidden state evolution trajectory to generate path feature representation; Based on the path feature representation, the first predicted value of the remaining useful lifetime and the first uncertainty measure are obtained by predicting the output layer with Gaussian parameterization. The remaining useful life is defined as the time remaining from the current moment to the engine failure threshold, which includes the exhaust temperature exceeding the preset red line value or the high-pressure rotor speed dropping to the design lower limit.

3. The parallel prediction method for the remaining life of an aero-engine according to claim 1, characterized in that, The multi-resolution feature extraction path specifically includes: The input feature sequence is subjected to multi-level discrete wavelet transform to decompose the sensor time-series signal into low-frequency approximation coefficients and high-frequency detail coefficients. The approximation coefficients correspond to the slow trend component in the engine degradation process, and the detail coefficients correspond to the transient event component. For the approximation coefficients and detail coefficients obtained from each level of decomposition, feature transformation is performed through independent neural network modules to obtain the resolution features at each level. All resolution features are aggregated, and the second predicted value of the remaining useful life and the second uncertainty measure are obtained by predicting through a Gaussian parameterized regression output layer.

4. The parallel prediction method for the remaining life of an aero-engine according to claim 1, characterized in that, The uncertainty-aware fusion mechanism employs a precision-weighted mechanism, specifically including: Calculate the reciprocals of the first uncertainty measure and the second uncertainty measure, and use them as the first precision weight and the second precision weight, respectively; dynamically adjust the first precision weight and the second precision weight according to the current flight conditions of the aero-engine to obtain the condition-adaptive fusion weight; Based on the fusion weights, the first predicted value of remaining useful life and the second predicted value of remaining useful life are subjected to condition-aware adaptive weighted fusion to generate the final predicted value of remaining useful life and the corresponding comprehensive uncertainty measure.

5. The parallel prediction method for the remaining life of an aero-engine according to claim 1, characterized in that, Also includes: Construct a prediction model that includes a continuous dynamic modeling subnetwork, a multi-resolution feature extraction subnetwork, a working condition recognition subnetwork, an uncertainty perception fusion module, and a decision rule mapping module; The prediction model is trained using a labeled training dataset. The loss function used in the training process includes a working condition identification loss term, a decision consistency loss term, a prediction error loss term, and a monotonicity regularization loss term. The monotonicity regularization loss term is used to constrain the monotonically decreasing trend of the remaining useful life prediction result with running time.

6. The parallel prediction method for the remaining life of an aero-engine according to claim 1, characterized in that, The discrete wavelet transform used in the wavelet multi-resolution mixer is one of the Daubechies wavelet, Haar wavelet, or Symlet wavelet.

7. The parallel prediction method for the remaining life of an aero-engine according to claim 1, characterized in that, The liquid neural network uses continuous-time network units based on a gating mechanism.

8. A parallel prediction system for the remaining life of an aero-engine, characterized in that, include: The data acquisition module is used to acquire and preprocess the time-series sensor data of the aero-engine to obtain the input feature sequence. The time-series sensor data includes at least: total fan inlet temperature, total high-pressure compressor outlet pressure, total high-pressure turbine outlet temperature, fuel flow rate, high-pressure rotor speed, low-pressure rotor speed, exhaust temperature, and vibration amplitude. The sampling frequency is 1Hz, and the original time-series data is divided into fixed-length time-series segments using a sliding window. The continuous dynamic modeling and multi-resolution feature extraction module is used to simultaneously input the input feature sequence into the continuous dynamic modeling path and the multi-resolution feature extraction path for parallel processing. The continuous dynamic modeling path models the continuous degradation process of the engine based on a liquid neural network, and outputs a first predicted value of the remaining useful life and a corresponding first uncertainty measure; the multi-resolution feature extraction path performs multi-scale decomposition and feature extraction on the input feature sequence based on a wavelet multi-resolution mixer, and outputs a second predicted value of the remaining useful life and a corresponding second uncertainty measure. The operating condition identification module is used to identify the current flight operating condition of the aero engine in real time based on the altitude, Mach number, throttle lever angle, fuel flow, and high-low pressure rotor speed difference in the input feature sequence. The flight operating conditions include: ground idle, takeoff and climb, cruise, maneuvering flight, and approach and landing. The fusion module is used to perform condition-aware adaptive weighted fusion of the first predicted value of the remaining useful life and the second predicted value of the remaining useful life based on the uncertainty-aware fusion mechanism, according to the first uncertainty metric, the second uncertainty metric and the current flight condition of the aero-engine, to generate the final predicted value of the remaining useful life and the corresponding comprehensive uncertainty metric. The hierarchical uncertainty decision module is used to determine the prediction reliability level based on the magnitude of the comprehensive uncertainty metric, and output prediction and maintenance decision suggestions corresponding to the level according to the preset reliability level-decision rule mapping.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the parallel prediction method for the remaining life of an aero-engine as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that executes the parallel prediction method for the remaining life of an aero-engine as described in any one of claims 1 to 7.

Images