Foundation CAN LM: A Pretrained Language Model For Automotive CAN Data
The Foundation CAN LM treats decoded automotive CAN data like a language to enable multi-objective generalization using a single pretrained model. It uses a unified tokenization scheme to handle mixed continuous and discrete signals, pretraining on large unlabeled datasets via Masked Language Modeling (MLM). The approach shows the foundation model can transfer effectively across different downstream tasks, though limitations exist in modeling subtle temporal cues.