Lightweight AI Deployment for Legacy Automotive ECUs: A Practical Optimization Approach

Automotive systems are increasing adopting data-driven and intelligent functionality in the areas of predictive maintenance, virtual sensors and diagnostics. This has led to a need for the AI models to be directly run on vehicle ECUs. However, most of these ECUs – especially those in cost-sensitive or legacy platforms lack the computational capacity and parallel processing support required for standard AI implementations. Given the stringent real-time and reliability requirements in automotive environments, deploying such models presents a unique challenge. This paper proposes a practical methodology to optimize both the training and deployment phases of AI models for low-computation ECUs that operate without parallelism. Designing lightweight model architectures, using pruning and quantization techniques to minimize resource utilization, and putting in place a strategy appropriate for single-threaded execution are the three main objectives of the developed approach. The goal is to guarantee that the final models satisfy real-time requirements without sacrificing prediction accuracy. To evaluate the effectiveness of the proposed method, a set of use cases relevant to automotive virtual sensors and condition monitoring were selected. Importantly, latency remains within the acceptable limits for in-vehicle ECUs. Analysis was performed for the memory and the CPU usage to ensure that the reliability of the operation is never compromised. This work provides a path for deploying AI in embedded automotive environments without the need for hardware upgrades. It supports OEMs and suppliers aiming to bring intelligent features to market efficiently while staying within the constraints of existing ECU platforms.