An aircraft engine residual life prediction method based on mixed distribution learning
By employing a hybrid distribution learning algorithm and random sampling techniques, the problems of lack of confidence and uncertainty aliasing in the prediction of the remaining life of aero-engines have been solved. This has enabled the quantification of the confidence of the prediction results and risk assessment, thereby improving the robustness of the prediction and the scientific nature of the decision-making.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-05
AI Technical Summary
Existing methods for predicting the remaining life of aero-engines lack confidence metrics, and sparse samples lead to large prediction biases. The overlapping of heterogeneous uncertainties makes it difficult to accurately assess the risk level.
A hybrid distribution learning approach is adopted. By constructing a hybrid distribution learning layer, the uncertainty of lifetime evolution is simulated by weighted combination of multiple sets of probability parameters. Multiple random sampling mechanisms are introduced, and the uncertainty sources are decoupled by combining the principle of full variance decomposition, and the complete probability distribution and confidence interval are output.
It enables the quantification of confidence in prediction results, improves the risk assessment capability in complex environments, ensures the robustness and scientific nature of predictions, avoids errors caused by sample sparsity and noise interference, and improves the accuracy of maintenance decisions.
Smart Images
Figure CN122153301A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for predicting the remaining life of an aircraft engine, belonging to the field of engine life prediction technology. Background Technology
[0002] With the evolution of aircraft engine health management technology, Remaining useful life (RUL) prediction has become a key basis for condition-based maintenance. Accurate RUL prediction can provide airlines with maintenance decision support and avoid unplanned downtime for maintenance. However, current methods for predicting aircraft engine RUL have the following shortcomings:
[0003] The prediction results lack confidence metrics: Most existing prediction methods are point predictions, only outputting a single lifetime estimate. When faced with complex on-wing environmental interference or sensor failures, they cannot provide a degree of certainty in the prediction results, making it difficult for maintenance personnel to assess the risk level of the prediction conclusions.
[0004] Sparse samples lead to large prediction bias: The high cost of acquiring full-lifecycle data for aircraft engines results in extremely sparse samples at the end of the degradation period. Traditional models, when processing out-of-distribution (OOD) data, are prone to producing prediction biases with inflated confidence levels due to a lack of awareness of unknown feature spaces. That is, even when faced with cognitive limitations caused by insufficient training data, the model still outputs erroneous predictions with extremely small variance, failing to accurately reflect the risk of prediction failure, thus leading to model performance degradation.
[0005] Heterogeneous uncertainty aliasing: Existing probabilistic prediction schemes struggle to effectively decouple the sources of uncertainty. Specifically, they cannot distinguish between aleatoric uncertainty caused by sensor random noise and epistemic uncertainty caused by limitations in model architecture and training samples. This makes it impossible to specifically assess the impact of data quality on model performance, thereby hindering accurate maintenance decisions.
[0006] Therefore, there is an urgent need for a method that can accurately predict the remaining lifespan of aircraft engines to support airlines' maintenance and protection decisions. Summary of the Invention
[0007] To address the problems of existing aircraft engine remaining life prediction methods, such as the lack of confidence metrics in prediction results, large prediction bias due to sample coefficients, and aliasing of source uncertainties, this invention proposes an aircraft engine remaining life prediction method based on mixed distribution learning.
[0008] The technical solution adopted by the present invention to solve the above problems is as follows: The steps of the present invention include: Step 1: Access the timing signals from multiple sensors of the aircraft engine, and perform cleaning, normalization, and sliding window slicing processing. Step 2: Construct a fusion-hybrid distributed learning layer at the end of the network, and use a weighted combination of multiple sets of probability parameters to simulate complex lifetime evolution uncertainties; Step 3: Introduce multiple random sampling mechanisms during the inference phase; Step 4: Using the principle of full variance decomposition, extract two types of indicators from the output composite probability model.
[0009] Furthermore, step 1 specifically includes: The processed time-series data is input into a deep neural network encoder, which uses its multi-layer nonlinear transformation capability to extract deep feature vectors that reflect the evolution of engine performance, and uses them as the input benchmark for subsequent probability distribution mapping.
[0010] Furthermore, step 2 specifically includes: Step 201: Map the input features to Each group consists of three independent Gaussian distribution parameter clusters; each cluster includes three core parameters: mixture weights, predicted mean, and variance term. Step 202, The weighted summation of the Gaussian distributions yields the final lifetime prediction probability distribution. Step 203: Perform Softmax normalization on the mixed weights to ensure that the sum of probabilities is 1; perform positive value constraint processing on the adversarial variance to ensure that the random uncertainty of quantization has physical meaning.
[0011] Furthermore, step 3 specifically includes: By observing the drift of the fusion distribution under multiple samplings, cognitive uncertainty is calculated. This indicator can reflect the cognitive blind spot of the model caused by the sparse sample at a specific degradation stage, and provide risk warning for the prediction conclusion.
[0012] Furthermore, the two types of indicators in step 4 include: Predicted mean: as the best estimate of remaining lifespan; Decoupled uncertainty: Output the random uncertainty representing random noise and the cognitive uncertainty representing lack of knowledge respectively, and form the final dynamic confidence interval.
[0013] The beneficial effects of this invention are: 1. This invention addresses the problem of lacking confidence metrics in prediction results by outputting a complete probability distribution through a fusion-hybrid distribution learning algorithm. To overcome the shortcomings of existing methods that only output a single lifetime value and cannot perceive risk levels, this invention constructs a fusion-hybrid distribution learning layer in step ②. This algorithm, by fusing multiple probability distribution components, not only provides the mean of the lifetime prediction but also outputs the confidence interval of the prediction result in real time. This provides maintenance personnel with a quantifiable confidence index, enabling them to accurately assess the risk level of the prediction conclusion based on the dispersion of the distribution when facing complex on-wing environmental interference.
[0014] 2. This invention introduces cognitive uncertainty, solving the problems of inflated confidence levels and prediction deviations caused by sample sparsity. Addressing the shortcomings of traditional models that easily produce inflated confidence levels and fail to reflect the true risk of failure in the late stages of degradation (out-of-distribution data), this invention utilizes a fusion hybrid distribution learning framework combined with random sampling techniques in step ③. This mechanism can identify the model's perception of unknown feature spaces. When entering regions with extremely sparse training samples, the algorithm spontaneously increases the width of the prediction distribution to reflect cognitive uncertainty. This effectively avoids prediction failures caused by cognitive limitations, ensuring robustness of predictions throughout the entire lifecycle (especially in the late stages of degradation).
[0015] 3. This invention achieves effective decoupling of heterogeneous uncertainties, solving the problem of overlapping uncertainty sources. Addressing the deficiency of existing probabilistic prediction schemes in distinguishing between noise interference and model limitations, this invention, in step ④, decomposes the total prediction risk into accidental uncertainty caused by sensor random noise and cognitive uncertainty caused by insufficient samples, based on the principle of total variance decomposition. This decoupling capability enables the system to specifically assess data quality and model generalization performance, thereby guiding maintenance personnel to distinguish between errors caused by data fluctuations and risks caused by insufficient model cognition, significantly improving the scientific nature and accuracy of maintenance decisions. Attached Figure Description
[0016] Figure 1 This is a diagram of the algorithm structure of the present invention; Figure 2 This is an application effect diagram of the present invention. Detailed Implementation
[0017] Specific implementation method one: as follows Figure 1 and Figure 2 As shown, the steps of the aircraft engine remaining life prediction method based on hybrid distribution learning described in this embodiment include: Step 1: Access the time-series signals from multiple sensors of the aircraft engine, and perform cleaning, normalization, and sliding window slicing processing; input the processed time-series data into a deep neural network encoder, and use its multi-layer nonlinear transformation capability to extract deep feature vectors that reflect the evolution of engine performance, as the input benchmark for subsequent probability distribution mapping; Step 2: Construct a fusion-hybrid distributed learning layer at the end of the network. The core logic is to use a weighted combination of multiple sets of probability parameters to simulate complex lifetime evolution uncertainties. The specific steps are as follows: Step 201: Map the input features to Each group consists of an independent cluster of Gaussian distributed parameters. Each cluster contains three core parameters: the mixture weights. (Characterizing the importance of this distribution in global prediction), prediction mean (Characterizing the lifetime prediction conclusions under this path) and the variance term (Characterizing random uncertainties affected by sensor noise); Step 202, put this The final lifetime prediction probability distribution is obtained by weighted summation of the Gaussian distributions. : (1), In formula (1), The remaining lifetime value to be predicted; Representing a normal distribution, this fusion algorithm enables the model to overcome the symmetry limitation of a single Gaussian distribution and flexibly capture the asymmetric, multimodal, and heavy-tailed distribution characteristics that may occur during the degradation process. Step 203: Adjust the mixed weights Perform Softmax normalization to ensure the sum of probabilities is 1; adversarial variance Implement positive value constraint processing to ensure that the random uncertainty of quantification has physical meaning; Step 3: Introduce a multiple random sampling mechanism (MC-dropout) during the inference phase. By observing the drift of the fusion distribution under multiple sampling, cognitive uncertainty is calculated. This indicator can reflect the cognitive blind spots of the model caused by the sparsity of samples at specific degradation stages, providing a risk warning for the prediction conclusion.
[0018] Step 4: Using the principle of full variance decomposition, extract two types of indicators from the output composite probability model: Predicted mean: as the best estimate of remaining lifespan.
[0019] Decoupled uncertainty: Output the random uncertainty representing random noise and the cognitive uncertainty representing lack of knowledge respectively, and form the final dynamic confidence interval.
[0020] Example Step 1: Data Access and Standardization Processing (1) Obtain sensor data streams throughout the engine's life cycle and select 14 key performance parameters, such as compressor temperature, low-pressure turbine temperature, fan speed, and engine pressure ratio.
[0021] (2) Convert the preprocessed temporal features into tensor form and input them into the prediction model.
[0022] Step 2: Network parameter configuration The number of mixed distribution components is set to n=4. Among them, component 1 captures the dominant degradation trend, and components 2, 3 and 4 are used to capture the skewed distribution caused by sudden changes in operating conditions, failure modes and sensor noise.
[0023] Step 3: Model training based on negative log-likelihood (NLL) The goal of model training is to ensure that the true RUL falls within the "peak" region of the network's predicted probability distribution. Loss function The calculation is as follows:
[0024] In the formula, The true value of the engine's RUL represents the actual lifespan of the engine in the simulation dataset. The Adan optimizer is used, with an initial learning rate of 0.001. Through gradient backpropagation using NLL loss, the model automatically adjusts the weights of each component. This allows it to automatically increase the prediction variance when there is a lot of data noise.
[0025] Step 4: Decoupling and Extraction Process of Heterogeneous Uncertainties During the reasoning (testing) phase, the risk is quantified through the following steps: Step 401: Directly calculate the mixture variance of the network output. If this value increases, it indicates that the signal-to-noise ratio of the current sensor signal has decreased.
[0026] Step 402: Enable MC-Dropout, perform T=80 random sampling forward computations on the same input, and record the fused mean of the network output for each iteration. Calculate the variance of the mean of these 80 predictions. .
[0027] Step 403, when Exceeding the preset threshold When this happens, the system determines that it is currently in a sparse sample region (end of degradation) and sends a risk warning to the maintenance end.
[0028] Performance Verification: This dataset contains 6 complex operating conditions and 2 failure modes. The performance after feature extraction is as follows: Figure 2As shown, the results clearly demonstrate the predicted RUL curve and its envelope range. In the early stages of the lifetime, the envelope is extremely narrow, indicating high certainty; towards the end of the lifetime (e.g., when the number of cycles > 250), due to… and At the same time, as the size increases, the envelope widens significantly.
[0029] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.
Claims
1. A method for predicting the remaining life of an aircraft engine based on hybrid distribution learning, characterized in that, The specific steps include: Step 1: Access the timing signals from multiple sensors of the aircraft engine, and perform cleaning, normalization, and sliding window slicing processing. Step 2: Construct a fusion-hybrid distributed learning layer at the end of the network, and use a weighted combination of multiple sets of probability parameters to simulate complex lifetime evolution uncertainties; Step 3: Introduce multiple random sampling mechanisms during the inference phase; Step 4: Using the principle of full variance decomposition, extract two types of indicators from the output composite probability model.
2. The method for predicting the remaining life of an aircraft engine based on hybrid distribution learning according to claim 1, characterized in that, Step 1 specifically includes: The processed time-series data is input into a deep neural network encoder, which uses its multi-layer nonlinear transformation capability to extract deep feature vectors that reflect the evolution of engine performance, and uses them as the input benchmark for subsequent probability distribution mapping.
3. The method for predicting the remaining life of an aircraft engine based on hybrid distribution learning according to claim 1, characterized in that, Step 2 specifically includes: Step 201: Map the input features to Each group consists of three independent Gaussian distribution parameter clusters; each cluster includes three core parameters: mixture weights, predicted mean, and variance term. Step 202, The weighted summation of the Gaussian distributions yields the final lifetime prediction probability distribution. Step 203: Perform Softmax normalization on the mixed weights to ensure that the sum of probabilities is 1; perform positive value constraint processing on the adversarial variance to ensure that the random uncertainty of quantization has physical meaning.
4. The method for predicting the remaining life of an aircraft engine based on hybrid distribution learning according to claim 1, characterized in that, Step 3 specifically includes: By observing the drift of the fusion distribution under multiple samplings, cognitive uncertainty is calculated. This indicator can reflect the cognitive blind spot of the model caused by the sparse sample at a specific degradation stage, and provide risk warning for the prediction conclusion.
5. The method for predicting the remaining life of an aircraft engine based on hybrid distribution learning according to claim 1, characterized in that, The two types of indicators in step 4 include: Predicted mean: as the best estimate of remaining lifespan; Decoupled uncertainty: Output the random uncertainty representing random noise and the cognitive uncertainty representing lack of knowledge respectively, and form the final dynamic confidence interval.