Sequential design of experiments for effective model-based validation of electronic control units

Abstract

Model-based verification of automotive electronic control units (ECUs) must ensure compliance with the target requirements in a short time frame. On the other hand, an increasing number of sources of variability (e.g. operating conditions, block parameters) impact system performance. To reduce the overall verification effort, much focus has been put on performance in simulation, and little on how to plan simulation experiments to yield maximum information with minimum number of runs. This paper shows how the classical framework of statistical Design of Experiments can be extended to perform faster and more reliable multivariable sensitivity and worstcase studies. Simulation experiments of a state of the art ECU modelled in SystemC/SystemC-AMS show significant increase in efficiency as compared to traditional approaches. Ó Springer-Verlag 2010 IntroductionAutomotive ECUs must face ever increasing demands. Requirements relate to functionality, computing power, safety, energy consumption. These all translate into more complexity, both as density of integration and heterogeneity. Pre-silicon verification phases must provide more reliable outcome. On the other hand, overall verification time must be reduced, to shorten development cycles.Solutions based on specification languages such as SystemC manage to speed-up steps of the system-level model-based verification. Raising the abstraction level gains performance in simulation, at the expense of a reduced model accuracy. This is one way to address the issue of long verification times. However, not much work focuses on more efficient and faster methods for multivariable sensitivity and worst-case analyses. These take part in the verification to ensure system reliability when multiple sources of variability (factors) are present.Factors vary in specified windows and can be both internal and external to the verified system. For sensitivity analysis, factors must be identified that have high impact on system outputs/performance (response). Worst-case analysis ensures the target requirements are met, under any factor conditions. In addition, configurations are found that ensure optimum performance. A challenge is how to cover the highly dimensional and continuous verification space with fewer simulation runs, to realize reliable sensitivity and worst-case studies.Simulation practitioners without statistical background often rely on trial-and-error methods (best-guesses, random corner-cases). These can easily end-up in a large number of runs without getting closer to the worst-case, thus they are not reliable. Pessimistic guesses lead to overestimations (oversize the specification range; over-restrict admitted factor variations). Most methods assume little or no impact of design factors on the system response, and do not discover interaction effects, i.e. that can be tracked only by concom-

Design of experiments for reliable operation of electronics in automotive applications

“…Previous work found DoE able to detect important effects and response extremes, for static responses in automotive applications [7], [8]. For more details on metamodelling methods, [9] provides an extensive review.…”

On more dependable assertion-based verification

“…To increase the problem coverage while decreasing the number of simulation runs, advanced methods like Corner Case Analysis [5] and Design of Experiments [6] were proposed. In contrast to statistical approaches these methods are based on "worst case" system analysis.…”

Verification of Mixed-Signal Systems with Affine Arithmetic Assertions

“…However, even if all corner cases are considered, the dependability of the result cannot be guaranteed since corner cases are not necessarily worst cases. Design of Experiments [4] allows reducing the number of simulation runs significantly and finding worst case performances more accurately. Even with high number of simulation runs, there is no guarantee that worst case performances are found.…”