IMU road-anomaly simulator and cross-vehicle generalization benchmark (code, simulated data and results)

Abstract

Code, simulated data and per-run results accompanying the article "Cross-Vehicle Generalization of IMU-Based Road-Anomaly Detection: A Physics-Based Simulation Study with a Lightweight CNN-GRU for IoT Edge Nodes" (submitted to Results in Engineering). All data are simulated; no field measurements are included. Contents: (1) simulation/ - physics-based generator of vehicle-mounted 6-axis IMU signals: ISO 8608 random roads with potholes, IRC:99-2018 speed humps and worn segments; a driver model; a batched 7-DOF full-car model of five vehicle classes with randomised instances; and an MPU-6050-class IMU error model. (2) ml/ - windowing and labelling, CNN-GRU and parameter-matched baselines (CNN-LSTM, 1D-CNN, GRU, LSTM, TCN, Transformer, random forest, SVM, threshold detectors), random-window / trip-disjoint / leave-one-vehicle-class-out protocols, distribution-shift and training-diversity experiments, edge-cost profiling, and the analysis scripts that generate every table and figure. (3) results/ - per-run predictions and metrics used in the paper. (4) data/ - the generated datasets (main dataset and shifted test sets), which can also be regenerated deterministically from the stated seeds. See README.md for exact reproduction commands and software versions. A third archive contains the decoy-braking ablation dataset (data/sim_decoy), in which traffic-like braking episodes are added away from road events.