Context
Airbus Aircraft uses acoustic test bench to observe and reduce noise on their aircrafts engines.
The acoustic test bench embeds more than 170 microphones. Measurements from these microphones are made at high sampling frequency for a long period of time, which increases the likelihood of having a problem on a sensor. To check sensors individually is a laborious and sometimes infeasible task.
The goal was to determine if, thanks to Machine Learning (ML) and Artificial Intelligence (AI), it would be possible to automatise this task.
Acoustic test bench:
Need
Be able to quickly detect the failure in order to identify faulty measures on the test bench and restart a measure campaign.
To go further, also be able to identify the type of failure to assess its criticality.
Solution
In order to create a fault detection model, our DiagFit software was used. The software enables to create a model that detects anomalies on the test bench but also classified them. This classification allows to identify them as precisely as possible to facilitate the intervention of operators.
Our approach being based on an “unsupervised” model, healthy data was used to determine the normality of the test bench. Then when an anomaly was detected, a classification was carried out in order to be able to label them and recognize them subsequently.
Results
The model demonstrated excellent defect detection and classification prediction capabilities on labeled data compared to the reference.
The model also highlighted the detection of some defects not observed by the operators. And conversely, it demonstrated “tagging” errors: the model’s prediction was more reliable than the “ground truth” given by the operators.
Comparison of results between human identification (left) and model prediction (right). Colors indicate the type of detection defect over time (x-axis) for all microphones (y-axis) for a 5-second measurement.
