Strategy for health monitoring and fault detection in heavy-duty diesel engines

Performance tuning, health monitoring, fault diagnosis, etc. are important aspects of testing a pre-production engine out in the field. To fulfill these tasks, a robust data-based strategy for detecting anomalies in diesel engines is proposed in this work. Such a strategy separates the healthy data from outlying, anomalous trends of faulty data. The data classifier used here is based on fundamental principles of statistical learning and derived from linear discriminant analysis. Further efforts to improve the classification performance led to the finding that steady state data makes for a more accurate classification of the working conditions of individual trucks. Hence an algorithm to extract steady data is suggested. After achieving nearly 100% accuracy for classifying data with one fault, this work is naturally extended to handle multiple trucks and the associated faults. Subsequently, a two-fault classifier using trucks belonging to distinct fleets and ratings, and a three-fault classifier using trucks belonging to the same family were devised. It was observed that data of trucks with similar power ratings and calibrations is more easily classified. The efforts towards clustering the healthy data of three different trucks together and separating it from the clusters of faulty data were successful. This was achieved at an acceptable accuracy of greater than 90%. The multi-fault, multi-class allocation scheme was validated using five completely new, healthy sets of data. These trucks contained engines with similar hardware and calibrations. It was observed that the data of all but one trucks were grouped in the healthy cluster. That one outlying truck, on further investigation, yielded unusually high values of crankcase pressure as compared to the other four. This established the validity of the technique in characterizing healthy data and identifying anomalous behavior of the on-field trucks