ML Detection and Isolation of Functional Failures using Syndrome Diagnostics

Abstract

Failure identification in complex engineering systems often needs to be in form of multilevel analysis, the first of which involves detection and isolation of functional disruptions across the system down to a discrete item. Considering that there is a functional flow(s) of operation in the machine, the loss or deviation of that functional flow(s) will imply a Functional Failure of the system. Most often, the disruption may propagate from where it may be first found and thus, the root issue can be isolated via the causal relationship between components and their flows. This paper references the idea of Causation-based AI: intelligence that brings together the fast estimations of machine learning and the domain-based physics of failure via qualitative models in the form of Syndrome Diagnostics (SD). There are three main routines in SD. Firstly, the current operating mode of the system is determined. Secondly, functional failure detection techniques are used to detect the existence of an anomaly. Thirdly, a functional failure isolation routine is executed to isolate the failing component, which is composed of two steps: using classification methods to generate a predicted syndrome and matching this generated syndrome with the failure syndrome pattern extracted from the Digital Risk Twin of the system (where the causation aspect is taken advantage of). A series of experiments are conducted and 90% of the failures in validation data have been correctly identified, which verifies the effectiveness of SD in terms of functional failure detection and isolation.