Abstract
As the automobile industry moves toward greater efficiency, safety, and autonomous operation, the demand for Electronic Control Units (ECUs) - the microprocessors controlling everything from engine functions to infotainment systems - has exploded globally. Currently, there are more than 200 million ECUs in the world, and that number is expected to rise to 700 million by 2030, necessitating enhanced ECU performance, reliability, and safety. However, the growing complexity of ECUs has ironically led to increased false alarms and failures, which in some cases have endangered user safety and privacy, causing heavy penalties for manufacturers. Although artificial intelligence (AI) and machine learning (ML) are showing promise in addressing ECU-related issues, existing methods remain insufficient. Manufacturers need to employ AI-driven, end-to-end, standardized solutions that help design, train, test, and deploy models without deep AI expertise and allow real-time runtime monitoring and retraining of the ECUs.Drawing on decades of experience in the electronics and automotive industries, as well as a track record of successfully deploying AI-based solutions in safety-critical systems like avionics and diesel engine control, a comprehensive method is proposed. It includes an array of novel functionalities that increase transparency, reliability, and safety while keeping development times low. Central to the method is a feature that creates an environment-sensitive digital twin of the ECU by assimilating data from ECUs and the vehicle, thus improving model fidelity and monitoring for unforeseen edge cases. The proposal is based on co-design and training of AI-based perception and prediction models, which can monitor the relevant environmental parameters both on-board and in the cloud. The on-board model is lightweight yet deterministic and can trigger warnings in case of model uncertainty and prediction errors, while the corrective action is taken by the re-licensed cloud-based model. A dataset of more than 33 million kilometers of driving from passenger vehicles in Northern Europe with SaaStronic and Focus models has been provided, using compute-efficient methods for interpretation and simplification of AI-based models.