An Explainable DL-Based Condition Monitoring Framework for Water-Emulsified Diesel CR Systems

Abstract

Despite global patronage, diesel engines still contribute significantly to urban air pollution, and with the ongoing campaign for green automobiles, there is an increasing demand for controlling/monitoring the pollution severity of diesel engines especially in heavy-duty industries. Emulsified diesel fuels provide a readily available solution to engine pollution; however, the inherent reduction in engine power, component corrosion, and/or damage poses a major concern for global adoption. Notwithstanding, on-going investigations suggest the need for reliable condition monitoring frameworks to accurately monitor/control the water-diesel emulsion compositions for inevitable cases. This study proposes the use of common rail (CR) pressure differentials and a deep one-dimensional convolutional neural network (1D-CNN) with the local interpretable model-agnostic explanations (LIME) for empirical diagnostic evaluations (and validations) using a KIA Sorento 2004 four-cylinder line engine as a case study. CR pressure signals were digitally extracted at various water-in-diesel emulsion compositions at various engine RPMs, pre-processed, and used for necessary transient and spectral analysis, and empirical validations. Results reveal high model trustworthiness with an average validation accuracy of 95.9%.