Datenbestand vom 13. Juni 2019
Tel: 089 / 66060798
Mo - Fr, 9 - 12 Uhr
Fax: 089 / 66060799
DER VERLAG IST IN DER ZEIT VOM 12.06.2019 BIS 23.06.2019 AUSCHLIESSLICH PER EMAIL ERREICHBAR.
aktualisiert am 13. Juni 2019
978-3-8439-3705-4, Reihe Informatik
Bordersearch: Efficient Characterization of Automotive Electronic Systems Through Machine Learning
142 Seiten, Dissertation Eberhard-Karls-Universität Tübingen (2016), Softcover, A5
The complexity of modern electrical and electronic systems increases steadily. In the automotive industry, this trend is amplified by current developments such as the growing number of advanced driver assistance systems and the emergence of electric drive vehicles. To guarantee their safety and reliability, chips and devices have to be verified thoroughly and comprehensively. This requirement is reinforced by new regulations such as ISO 26262.
As a consequence, for characterization and verification, chips and devices are now often tested within their application context. However, this drastically increases the number of parameters that have to be considered when investigating the reason of possible failure. Established characterization and verification methodologies only work well with a limited number of parameters, as they otherwise would have to perform an infeasible amount of tests to provide reliable results. Therefore, new methodologies are required that can efficiently handle systems which have dozens of parameters.
To overcome this problem, this thesis presents a new methodology, termed Bordersearch, which analyzes systems based on functional tests that only provide binary responses, i.e. “pass” or “fail”. Internally, it combines several machine learning techniques, which allow it to operate efficiently and reliably on systems with dozens of parameters: A classifier robustly predicts the border between pass and fail region; additionally, it estimates how certain that prediction is in different regions of the parameter space. Based on that information, an adaptive sampling strategy concentrates most tests close to the border. Thus, each test provides much information about the border, and even with comparatively few samples, Bordersearch can accurately resolve the border. Additionally, Bordersearch performs an automated detection of relevant parameters and provides novel ways to visualize the multi-dimensional interaction between them. With this assistance, the results of an experiment can easily be understood and interpreted without a deep knowledge of machine learning and statistics.
Bordersearch has been applied successfully for the characterization of several automotive systems. In all these experiments, it was consistently faster and more accurate than state-of-the-art methods. It is thus a vital contribution to the efforts that enhance the verification of e/e systems to keep pace with other advancements in the automotive industry.