Skip to content

Preventive maintainence

Project Title: Preventive Maintenance of Aircraft Engines
Client/Collaboration: Rolls-Royce & Imperial College London
Timeline: January 2022 – December 2022
Technologies: Python, Machine Learning, Unsupervised Learning, Autoencoder-decoders
Team Size: Academic-industry collaboration

Publications:
1. Presented Scarecrow - Intelligent Annotation platform for Engine Health Management in AI ML Systems conference
2. Predictive maintainence of aircraft engines in 9th International Conference on Business Analytics and Intelligence (Best paper award)

Problem Statement

Aircraft engine failures can lead to significant safety risks and operational costs. Traditional maintenance schedules are often reactive or time-based, lacking the precision needed to detect early signs of failure. The objective of this project was to develop predictive models that could identify potential failures in aircraft engines before they occur, enabling proactive maintenance.

Solution

Sensor data from aircraft engines was analyzed to detect patterns indicative of component degradation. Machine learning models were developed to predict failures in specific engine parts, using historical flight and maintenance data. The models were trained to distinguish between normal and abnormal behavior, allowing for early warnings and targeted inspections. The project also involved collaboration with subject-matter experts to validate model outputs and refine feature engineering.

  1. Academic project: Explored various unsupervised failure identification methods on aircraft engine simulated data. Explored various ways of implementing said methods to predict failure in engines.
    • Different failure modes and degradation scenarios were observed, and three different unsupervised approaches were selected
    • Simulated data from CMAPSS was taken to test the different methods on real failure modes on aircraft engine data
  2. Implementation in industry: Implemented a novel autoencoder-decoder model to predict the ideal behavior of 250 plus parameters in an aircraft engine. This helped engineers identify anomalous behaviors of aircraft engines on test beds.
    • Implemented a novel autoencoder-decoder-based approach to predict the ideal behaviour of more than 250 parameters in steady and transient phases of flight
    • Detected anomalies on test bed experiments using z-scores and CUSUM

Impact

The predictive models demonstrated strong potential in reducing false positives and improving the accuracy of failure detection. By enabling earlier interventions, the solution aimed to reduce unplanned downtime, enhance safety, and optimize maintenance schedules. The project was recognized through academic publication and contributed to ongoing research in intelligent engine health monitoring.

Back to top