Machine Learning Based Error Detection in Transient Susceptibility Tests

Compliance to electromagnetic compatibility (EMC) standards is a fundamental requirement for modern integrated circuits (ICs). In this framework, error detection in transient susceptibility tests is of crucial importance to assess the circuit robustness. However, performing such tests is expensive and requires an ad hoc hardware, whose configuration must be adapted for the different test setups, i.e., the transient waveforms, defined in the EMC standards. This paper describes a novel machine learning based approach for error detection, which only requires the raw output data from a susceptibility test: neither additional information about the architecture of the device under test nor the test configuration is needed. We applied and evaluated anomaly detection techniques (a branch of machine learning methods focused on error detection) for transient susceptibility tests with two pertinent examples (one simulation- and one measurement-based). The proposed techniques detected errors successfully in both unsupervised and supervised scenarios. Moreover, these can give insight on the output behaviors that are more likely to cause errors during the test. As shown by our results, an anomaly detection-based approach is an applicable and viable solution for automatic error detection in transient susceptibility tests.