AI-based CAN bus analyzer/simulator system for in-vehicle networks

Analyzing and simulating Controller Area Network (CAN) traffic remain challenging because decoding often depends on proprietary formats and manual effort. This thesis propose an automated machine-learning framework for SAE~J1939 traffic that, after decoding, standardizes preprocessing, classifies anchor signals, reconstructs multiple unknown signals, assigns semantics via a bounded score fusion, and re-encodes protocol-valid traces. The system also supports optional adversarial injection for intrusion detection (IDS) research. Evaluated on public J1939 datasets from heavy-duty vehicles, tree-based classifiers reach 98.4–99.6 % accuracy, and multi-signal reconstruction attains (R^2>0.90). In a representative run, signal matching 23 signals matching. The framework provides a strong, computationally efficient baseline for reverse engineering and a reproducible facility for generating realistic and adversarial CAN traces to advance IDS benchmarking. During training, This work uses J1939 DBCs only to derive ground-truth field boundaries for supervision; at inference, the system consumes unlabeled, tokenized payloads and does not require a DBC.