Average precision peaks with six detectors combined.

Flag Scania trucks whose engine component is about to fail, learning only from 318 normal trips of noisy, irregular sensor data, with no deep learning.

I engineered temporal features from the sensor series and averaged six classical outlier detectors into one trip-level score.

Temporal features lifted average precision from 0.31 to 0.53, and the ensemble pushed it to 0.59, double the 0.29 of random guessing.