Project case study · 2026

PX4 Flight Log Anomaly Analyzer

An interpretable anomaly-detection system for PX4 flight logs that combines unsupervised ML scoring with rule-based findings.

Role
Machine Learning Engineer
Year
2026
Technologies
PythonScikit-learnPyTorchStreamlitPX4
0.940Held-out AUC
55Features per timestep
View project on GitHub

Problem

PX4 logs contain dense, multivariate signals that make abnormal behavior difficult to spot manually.

Approach

The system extracts 32 raw and 23 engineered features, trains on 110 normal multirotor logs, compares five unsupervised approaches, and combines model scores with interpretable operational rules.

Architecture

.ulg upload → parser → 55-feature pipeline → OCSVM scoring + rule detectors → contribution analysis → Streamlit dashboard

Evaluation

OCSVM with an RBF kernel reached 0.940 AUC on a held-out set of 34 normal and 31 anomalous logs, ahead of the four reported alternatives.

Product

The Streamlit interface brings together summary metrics, a score timeline, anomaly contributors, rule overlays, and a signal explorer so an operator can move from a flagged interval to the telemetry behind it.

Limitations

Training is multirotor-only. The evaluation set contains 65 logs, GPS-only fault sensitivity is limited, and thresholds are relative within each log rather than globally calibrated.