Aircraft Engine Ageing Evaluation Using Flight Data Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for monitoring aircraft engine state fail to explicitly account for engine ageing over time, lacking monotonicity in calculated indicators and requiring a priori knowledge of engine models.
Innovation Solution
A method that constructs time-domain series from flight data, trains an embedding function and classifier to learn monotonicity in engine ageing, using a cost function to optimize the representation of flight data in an embedding space, and calculates an engine state indicator based on distance from a reference flight.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional indicator tracking methods are used, then engine state monitoring is performed, but the indicators lack monotonicity and may predict improvement without maintenance
Solution Approach 1:
The patent applies dynamics by making the indicator system adaptive through machine learning. The embedding function and classifier are trained dynamically on historical flight data to learn the actual degradation patterns of the specific engine, replacing static conventional indicators with dynamic, data-driven indicators that adapt to the engine's unique aging trajectory.
Solution Approach 2:
The patent changes parameters by transforming raw flight data into embedded representations that capture temporal evolution. The cost function optimizes parameter transformations to ensure monotonicity, and the system extracts degradation indicators that inherently reflect cumulative damage rather than allowing parameter reversals that suggest improvement without maintenance.
2Measurement precision
If physical or statistical models are used, then engine state evaluation is achieved, but a priori knowledge of engine models is required
Solution Approach 1:
The system applies self-service by enabling the engine data to train its own evaluation model. Instead of requiring external experts to construct physical models based on a priori knowledge, the system automatically learns degradation patterns directly from the engine's operational data, with the data serving as both input and training material for the embedding function and classifier.
Solution Approach 2:
The patent substitutes mechanical model construction with data-driven learning. Rather than manually building physical or statistical models based on engine mechanics knowledge, the system uses machine learning algorithms to automatically discover patterns in flight data, replacing the need for explicit mechanical understanding with empirical pattern recognition.
3Productivity
If traditional indicator methods are used, then flight data is tracked, but time-related information such as engine ageing is not explicitly utilized
Solution Approach 1:
The system applies preliminary action by training the embedding function and classifier on historical flight data before actual prediction. This pre-training phase allows the model to learn temporal patterns and degradation trajectories in advance, so that when evaluating current engine state, the system can immediately leverage accumulated time-related information without losing critical aging signals.
Solution Approach 2:
The patent applies dimensionality change by transforming time-series flight data into embedded representations that explicitly capture temporal evolution. The embedding function maps sequential data into a space where time-related information is preserved and enhanced, allowing the classifier to effectively utilize aging patterns that would be lost in traditional flat indicator tracking.
Data Source
AI summary
A method for evaluating the state of an aircraft engine for a given flight, each flight of the aircraft being associated with one time-domain series, includes, each time maintenance of the aircraft is performed, creating a dataset including the time-domain series associated with each flight carried out between the maintenance and the preceding maintenance; creating a set of datasets including each created dataset and dividing it into a training set and a validating set; conjointly training an embedding function and a classifier, selecting a reference time-domain series from a time-domain series of the validating set; and computing a distance between the embedment associated with the time-domain series of the given flight and the embedment of the reference time-domain series, with a view to computing an indicator of engine state.