MOCAFP: Meta-Optimisation Controller for Adaptive Fault Prediction in Real-Time Autonomous Vehicles Using Deep Learning

Abstract

Deep learning pipelines play a crucial role in the perception, planning, and control of autonomous vehicles (AVs). The optimization method, which gov- erns the effectiveness of models in learning and generalizing within dynamic en- vironments, represents a crucial yet underexplored component of these pipelines. Existing techniques, including SGD, Adam, and hybrid variations like BAAO and GASGD, remain static throughout the training process and are unable to adapt to the diverse conditions of the real world. By utilizing training inputs such as gradient variance and convergence rate, we present a Meta-Optimization Framework for Fault Prediction (MOCAFP) that functions as an adaptive con- troller, selecting and refining optimizers in a dynamic manner. The assessment of the system involved four datasets, which comprised both simulated and real- world driving environments: Udacity Jungle, Udacity Lake, KITTI, and a real- world Urban Roads dataset. Experimental results indicate that MOCAFP demon- strates superior performance compared to baseline optimizers regarding conver- gence speed and learning stability. Specifically, it achieves competitive training durations while enhancing the coefficient of determination (R2) by 12%, decreas- ing prediction variance by 15%, and minimising mean absolute error (MAE) by as much as 18%. The enhancements indicate that MOCAFP serves as a depend- able and scalable method for advancing fault prediction in AV pipelines.