This paper proposes a new methodology for detecting and diagnosing faults found in heavy-duty diesel engines based upon spectrometric analysis of lubrication samples and is compared against a conventional method, the redline limits, which is utilized in a number of major laboratories in the U.K. and across Europe. The proposed method applies computational power to a well-known maintenance technique and consists of an improved method of preprocessing to form a derivative tuple, which extracts further information from the measured elemental concentrations. To identify incipient faults, the distance in vector space is calculated using a Gaussian contour, generated from prior data, as the zero crossing, which enables novel samples to be classified as normal or abnormal. This information is utilized as the input to a probabilistic directed acyclic graph in the form of a belief network. This network provides a prognosis for the mechanism as well as suggesting possible actions that could be taken to rectify the diagnosed problem, supported with confidence probabilities. The proposed method is evaluated for both accuracy in detecting a fault as well as the duration of time that is provided before the event occurs, with significant improvements in both metrics demonstrated over the conventional method.