ai_can_anomaly_detection_data

The rows the detectors in asana17/ai_can_anomaly_detection are trained, calibrated and tested on, built from CAN logs by assemble.dataset there. A run reads them at one revision and records that revision, so the models in asana17/ai_can_anomaly_detection_runs each name the data they were fitted on.

python3 -m evaluate.pc.run asana17/ai_can_anomaly_detection_data <revision> out runs_clone

main holds the dataset built from every log. A smaller one built for a quick try goes to a branch of its own.

Source

Built from the University of Turku J1939 truck dataset, a Renault Euro VI truck on the road, normal traffic only. It is CC BY 4.0, and so is this. https://etsin.fairdata.fi/dataset/7586f24f-c91b-41df-92af-283524de8b3e/data

The raw logs are not here. Fetch them from the page above.

Files

file	holds
seconds.json	MIN_SPEED, and each log's seconds above it, which splits the logs into train and test
grid.json	the train logs and the grid settings the grid was built with
grid_raw.npy	the train logs on a 0.1 s grid, float32, 17 signals a row. The train and calibration rows are cut from it
grid_t.npy, grid_seg.npy	each grid row's time and segment
scale.npy	the mean, then the std, every row is z-scored by, fitted to the train rows
built.json	the logs and the settings the attack set and the scale were built with
attacked.json	where each attack sits, its PGN, time span, rows and how far it moved them
attacked_raw.npy	the test logs with one attack each, on the grid
attacked_rows.npy	attacked_raw.npy z-scored by scale.npy
attacked_t.npy, attacked_seg.npy, attacked_label.npy, attacked_wheel.npy	each test row's time, segment, attack label and wheel speed

The signals and the settings are in grid.json and built.json.

Frames

frames/ holds the same attacked test logs frame by frame, for sending the attacks over a real CAN bus. They go with the rows above, the attack set the runs 20260916-001002, 20260916-064753, 20260916-234726 and 20260919-025203 scored.

file	holds
frames/frames.parquet	every test log's frames with its attack in, one row group per log
frames/attacked.json	attacked.json with each attack's log added

column	holds
log	the source log, such as part_3/20210204093406279004.csv
timestamp	epoch seconds
can_id	the 29-bit identifier
data	the payload bytes
attacked	True where the attack changed the payload

A frame can change while no row moves, so score a detector on attacked_label.npy, not on attacked. A row cut short in the source log was dropped on reading.