CAN Anomaly Detection

STEP / 01

Brief

ADAS-equipped vehicles emit dense streams over the CAN bus. Latent failures hide inside that noise. The mission was to surface anomalous behaviour automatically, without labels, at signal-rate.

STEP / 02

Pipeline

From raw frames to a model that flags drift in real time.

• 01Extract raw CAN frames, decode each byte to a decimal cell✦
• 02Project frames into a multivariate matrix indexed by message ID✦
• 03Focus on critical IDs 0x2B6, 0x32D, 0x2D8✦
• 04Window into multivariate time series with overlap✦

STEP / 03

Model

A Variational Auto-Encoder built around a Transformer encoder/decoder. The model trains exclusively on PASSED runs — anything off the learned manifold scores a high reconstruction error and trips the detector.

STEP / 04

Friction points

What made it hard.

• 01Heavy noise floor on long acquisition runs✦
• 02High dimensionality (one channel per byte per ID)✦
• 03Long-range temporal dependencies across IDs✦

STEP / 05

Outcome

A reproducible pipeline that flags previously-unseen behaviours, generalises across vehicles, and is fully automated end-to-end.