Plausibility Assessment and Validation of Deep Learning Algorithms in Automotive Software Development

Abstract

The implementation of artificial intelligence (AI) systems in automotive software development still is an obstacle. Despite of accelerating scientific research and big wins in this field, the practical application is only possible in restricted environments or non safety critical components. There is a need to develop methods to verify the robustness and safety of AI software modules. The data based generation of deep learning (DL) algorithms creates black box models, which properties inhibit a validation as it is done for deterministic algorithms following ISO 26262. This paper introduces methods to assess the plausibility of AI model outputs. A description of the training data domains for a robust training is accomplished by means of one-class support vector machines (OCSVMs). This anomaly detection process encloses valid data within a DB, to be able to verify model outputs during operation. A further categorization of the training data domain into 20, equally spaced sub-domains led to best results in detecting implausible model calculations.