The notebook of this folder presents a solution for an unsupervised anomaly detection task on accelerometer data that were acquired during helicopter flights.
This exercice was presented as a data challenge competition during my Post Master program in Big Data at Télécom Paris. I ranked 2nd out of 51 participants.
The notebook has been redacted in french but you can find the corresponding english article in my portfolio here : https://antonindurieux.github.io/portfolio/2_accelerometer_anomaly_detection/.
The whole code and explanation is in the Data_challenge-Detection_anomalies_non_supervisee.ipynb notebook.
The data is voluminous and thus hasn't been uploaded on this Github repo.
It consists 4188 time-series of accelerometer data, that were acquired during helicopter flights. Each observation was a 1 minute recording, sampled at 1024 Hz.
- Data import
- Statistical features extraction
- Extraction of frequency information
- Anomaly scores calculation
- Other approaches and kernel-PCA