Context
A test aircraft carries many sensors (several hundred) and consequently measures large quantities of signals (several TB). This data is used to analyze the proper functioning of the aircraft, and to help the R&D team refine test reports. It is therefore important that the data generated is qualitative, so that any deterioration in sensor quality can be quickly detected.
However, it seems tedious for engineering departments to have to examine each sensor by hand and determine whether all the resulting time series have integrity.
Need
The customer wanted to be able to perform automatic fault detection on sensor data. Thanks to this detection, he could highly reactive to interrupt the test, trigger intervention on the device to carry out the necessary repairs, and quickly restart the measurements.
Solution
DiagFit analyzed the sensor data and generated an automatic fault detection model, enabling the teams to launch rapid diagnostics on the sensors.
To train the model, 3 different contexts were used to check its resilience:
- Training on data from 2 flights of the same aircraft, and testing the model on data from 2 other flights of the same aircraft.
- Perform training on data from one aircraft, and test on data from another aircraft
- Test outlier detection, i.e. perform anomaly detection on a flight with no healthy data
Results
DiagFit was able to demonstrate a performance reaching 100% detection for the anomalies listed, on 3 different use cases, even though 2 of these cases were outside the software’s specifications.
DiagFit also detected other anomalies that had not been identified upstream by the customer, but which were nevertheless collaborated on by the customer’s teams.
