Unleashing the Power of Anomaly Data for Soft Failure Predictive Analytics

Testing challenges of soft failures, including transient and intermittent failures, are being compounded by the fact that these failures are often induced by interference from time-varying stress factors, especially under a harsh environment in safety-critical applications. This paper presents a predictive analytics methodology using a continuing stream of anomaly data to tackle soft failure testing challenges. It is within a proposed silicon health prognosis framework. Multi-State Models (MSM) are applied to model soft failure progression with prognostic factors (e.g. interference) as time-varying covariates. The unique power of anomaly data can be unleashed with statistical machine learning techniques to infer potential interference effects on failure evolution and recovery rates. Failure prediction results can further be used for safety mitigation decision making. Several examples in context of 3-D mixed-signal SOC are analyzed to illustrate the proposed method. These predictive analytics methodology and prognosis framework are expected to pave an alternative way to improve dependability of safety-critical systems.