Observation framework of errors in microprocessors with machine learning location inference of radiation-induced faults

This paper presents a non-intrusive failure reason capturing approach for processor-based system-on-a-chip (SoC) integrated as an intellectual property (IP). It provides diagnostic information about the origin of single-event upsets (SEU) in the context of technology qualification to radiations. Thanks to the combination of trace events buffering and error detection with triggering mechanisms, the module is able to capture an execution trace containing the error propagation, using only 1 KB of memory. The execution trace is supplemented by a configurable set of pipeline registers. For single-event functional interrupt (SEFI), we also propose a technique based on a machine-learning algorithm to find from which register the SEU comes from. The captured CPU trace is processed by a classification algorithm, trained on a fault-injection campaign database and providing an accuracy score up to 87 %.