Software Implemented Detection and Recovery of Soft Errors in a Brake-by-Wire System

This paper presents an experimental study of the impact of soft errors in a prototype brake-by-wire system. To emulate the effects of soft errors, we injected single bit-flips into "live" data in the architected state of a MPC565 microcontroller. We first describe the results of an error injection campaign with a brake-by-wire controller in which hardware exceptions are the only means for error detection. In this campaign, 30% of the injected errors passed undetected and caused the controller to produce erroneous outputs to the brake actuator. Of these, 15% resulted in critical failures. An analysis showed that a majority of the critical failures were caused by errors affecting either the stack pointer or the controller's integrator. Hence, we designed two software implemented error handling mechanisms that protect the stack pointer and the integrator state, inducing an overhead of 4% in data and 8% in speed. A second error injection campaign showed that these mechanisms reduced the proportion of critical failures one order of magnitude, from 4.6% to 0.4% of the injected soft errors.