Automated Diagnosis of Engine Misfire Faults Using Combination Classifiers

Existing on-board diagnostics vehicle systems can detect the existence of faults, but their diagnostic (fault isolation) capabilities are rather low. Extensions to on-board diagnostics are needed in order to provide a high degree of automated diagnostic support. In this context, we study in this article the problem of internal combustion engine misfires, which constitute a class of automotive faults known to be difficult to diagnose, and present a combination classifier that has excellent performance in classifying the various root causes of misfire faults. We first obtained real-life data and built a database consisting of 2,299 time instances of actual misfire and misfire-free cases. Fault data were captured on several different vehicle makes and models, with each misfire fault belonging to one of three different categories (air-intake, coil-ignition, and fuel-injection), further subdivided into a total of seven subcategories. We then developed a combination classifier (referred to as TMF, “Trees for MisFires”) and obtained its performance on the real-life misfire fault dataset. Extensive simulation results show that TMF outperforms a variety of other standard classifiers (whether single or ensemble) as well as other combination classifiers, both for “low-resolution” diagnosis (classification consisting of three misfire fault categories plus misfire-free) and “high-resolution” diagnosis (classification consisting of seven misfire fault subcategories plus misfire-free). Moreover, these results hold even when the training set is restricted to be a very small portion of the available dataset, which is a valuable asset of a realistic classifier.