A machine learning approach to diesel engine health prognostics using engine controller data

Abstract

Many military assets such as surface ships and ground vehicles use diesel engines as their prime movers, and accurately estimating remaining useful life has a high value for enabling predictive maintenance and improving fleet logistics. Most of these diesel engines are already equipped with an array of sensors and digital data busses to support the function of the integrated electronic control module (ECM). There are cost advantages to developing predictive analytics and prognostics using existing embedded sensors. This paper describes a hybrid approach to predictive capabilities that utilizes multiple techniques for the implementation of embedded prognostics using existing sensors. One of the challenges is the fidelity of the data. This paper describes an automated approach to feature and classifier selection for hybrid prognostics. Maintenance records with associated diesel engine sensor data for several different engine classes were acquired, which enabled the training data sets to be organized by failure modes. To help prevent false positives, some filtering of the maintenance logs was required to only include those records likely to be associated with the selected failure mode sensor data sets. The classifier-based, data-driven approach essentially maps multiple channels of the sensor data into subspaces trained to classify multiple distinct failure modes. The intent of this step is to enable fault isolation by quantitatively determining which failure mode class the data best fits statistically. The remaining useful life estimate is provided by tracking the temporal path of the data from the healthy engine classification to one of the known failure mode classes using engine load-hours as the metric for the prognostics.