Fault Grading Techniques of Software Test Libraries for Safety-Critical Applications

Abstract

The adoption of complex and technologically advanced integrated circuits (ICs) in safetycritical applications (e.g., in automotive) forced the introduction of new solutions to guarantee the achievement of the required reliability targets. One of these solutions lies in performing in-field test (i.e., the test performed when the device is already deployed in the mission environment) to detect faults that may arise in this phase of electronic circuit life. In this scenario, one increasingly adopted approach is based on the software test libraries (STLs), i.e., suitable code which is run by the CPU included in the system and is able to detect the existence of possible permanent faults both in the CPU itself and in the rest of the system. In order to assess the effectiveness of the STLs, fault simulation is performed, so that the achieved fault coverage (e.g., in terms of stuck-at faults) can be computed. This paper explains why the fault simulation of the STLs represents a different problem with respect to the classical fault simulation of test stimuli (for which very effective algorithms and tools are available), shows why it can be highly computationally expensive, and overviews some solutions to reduce the computational cost and possibly trade-off between results accuracy and cost.

Hardware-in-the-Loop-Based Real-Time Fault Injection Framework for Dynamic Behavior Analysis of Automotive Software Systems

“…Fault injection is a system reliability assessment technique in which faults are intentionally introduced into a system through controlled experiments and the behavior of the system in the presence of faults is observed [ 13 ]. This approach enables an effective evaluation of various dependability attributes, i.e., safety, reliability, availability and security against random faults [ 14 , 15 ]. Most of the faults considered in this approach include nonlinear and unpredictable faults, called random hardware (HW) faults, which are mainly caused by environmental factors and occur as degradation of component functionality.…”

Pros and Cons of Fault Injection Approaches for the Reliability Assessment of Deep Neural Networks

“…For example, a small CNN with only seven layers can take about 25 minutes to run a single inference [7]. Furthermore, existing commercial fault simulation tools are not tuned and neither optimized to face the complexity of the state-ofthe-art DNN applications [29] (with billions of neuronal computations). This means that a FI at the hardware level is accurate but very costly in terms of simulation time.…”

Special Session: AutoSoC - A Suite of Open-Source Automotive SoC Benchmarks

“…Computing these figures for a given STL also requires a new generation of tools called Functional Fault Simulators. Several recent works introduced guidelines on how to correctly generate STLs for CPUs [27], [28] and peripherals [29], how to speed up the FI experiments [30], how to maximize their fault coverage in the different scenarios (possibly minimizing the test time [31]), and how to re-use existing STLs.…”