Importance Weighting of Diagnostic Trouble Codes for Anomaly Detection

Abstract

Diagnostic Trouble Codes (DTCs) allow monitoring a wide range of fault conditions in heavy trucks. Ideally, a perfectly healthy vehicle should run without any active DTCs; in practice, vehicles often run with some active DTCs even though this does not pose a threat to their normal operation. When a DTC becomes active, it is therefore unclear whether it should be ignored or considered as a serious issue. Recent approaches in machine learning, such as training Variational Autoencoders (VAEs) for anomaly detection, do not help in this respect, for a number of reasons that we discuss based on actual experiments. In particular, a VAE tends to learn that a frequently active DTC is of no importance, when in fact it should not be dismissed completely; instead, such DTC should be assigned a relative weight that is smaller but still non-negligible when compared to other, more serious DTCs. In this work, we present an approach to measure the relative weights of active DTCs in a way that allows the analyst to prioritize them and focus on the most important ones. The approach is based on the concept of binary cross-entropy, and finds application not only in the analysis of DTCs from a single vehicle, but also in monitoring active DTCs across an entire fleet.