Developing a prediction model for estimating hardware requirements based on standardized embedded system characteristics in the automotive domain

Abstract

Introduction: Modern day automotive vehicles often contain complex embedded systems that handle behavior of vehicles. These systems often comprise more than a hundred electronic control units (ECUs) that are responsible for different vehicle functionality, each requiring various degrees of computational power, and thus hardware requirements. Objectives: This thesis presents the development and application of prediction models that are able to predict baseline hardware requirements for ECUs in terms of CPU, RAM and ROM metrics for vehicular embedded systems. These models are based on what common high-level features that ECUs have with the features largely being realized by standardized base software modules from the AUTOSAR Classic standard. The created models aim to introduce another way of validating hardware characteristics on ECUs or propose suitable hardware baselines for new ECU projects. Methods: From the base software modules standardized by AUTOSAR, a grouping of said modules was presented to be able to characterize common ECU high-level features in Volvo Cars Corporation embedded systems. This characterization was then used as input to create regression neural networks that, given a set of features, are able to predict an estimation on what CPU, RAM and ROM size is required for a proposed ECU. Results: The developed models are currently able to predict an estimation on resource utilization for 32-bit CPUs, RAM, and ROM at runtime for an ECU. This estimation is rounded up to a suitable baseline (total available measure on an ECU) based on identified Volvo Cars Corporation ECU baselines. The models are able to predict suitable hardware requirements for ECUs in 60-78% of cases.