Data-driven framework for detecting anomalies in field failure data

Abstract

This paper discusses the design of a data-driven framework for detecting anomalies in the automotive field failure and repair data. The anomaly detection framework detects anomalies at two levels: 1) It detects anomalies in repair data using system-level fault model (or fault dependency-matrix) and diagnostic reasoner; 2) It detects anomalies in diagnostic trouble code (DTC) data using operating sensory parameter identifiers (PIDs) data mining. The system-level fault model provides a way to capture causal relationships between failures and symptoms of a given system. A repair is declared as anomalous if it does not match the repair recommended by the fault model and diagnostic reasoner. The PIDs data mining detects anomalies in DTC data by detecting patterns in the associated PIDs using various statistical techniques such as scatter plots, clustering and decision trees. The DTC anomalies could be either due to errors in the preconditions under which the DTCs are designed to set or errors while implementing them in the software. The PIDs data mining module provides a focused feedback to engineers for detecting the errors in DTC software algorithms and enhancing the diagnostic design of DTCs during the early stages of vehicle production. We demonstrate the data-driven framework on automobile fuel vapor pressure sensor problem.