Bordersearch: An adaptive identification of failure regions

The reliability and safety of modern analog devices, e.g. in automotives, aircraft or consumer electronics, is influenced by many input parameters like supply voltage, ambient temperature or load resistances. In certain regions of this large parameter space, the device exhibits degraded performance or it fails completely. The validation of such a device has to find the regions of the input parameter space in which the device misbehaves. However, with several parameters, it is a complex task to determine these regions, especially if parameters interact. In this paper, we present the Bordersearch algorithm, which combines adaptive testing with a machine learning classifier to efficiently determine the border between passing and failing regions in the parameter space. Furthermore, this method enables sophisticated post-processing analysis, e.g. better visualizations and automatic ranking of the parameters according to their influence. This algorithm scales well to a high-dimensional parameter space and is robust against outliers and fuzzy borders. We show the effectiveness of this method on an automotive electromechanical system with eleven input parameters.