A multi-domain diagnostic framework for CRD fuel injection systems under water-in-diesel emulsion fuel conditions

Abstract

From a global perspective, combustion emission from diesel engines has negatively impacted our environment as observed from the increasing global warming and other environmental pollution. In the quest for mitigating these consequences, diesel emulsification offers a cost-efficient and eco-friendly alternative but the corrosion, wear, and power loss effects on the engine pose a strong concern for global adoption. In the right proportions and mixing conditions, water-in-diesel (WiD) emulsion fuels can offer long-term solutions to reducing the number of combustion residuals-nitrogen oxides (Nox), particulate matter (PM), carbon oxides (COx) and with appropriate condition monitoring techniques in place, engine efficiency can be maintained. Interestingly, the rail pressure sensor (RPS) of a common rail (CR) diesel engine reflects the underlying transient and spectral dynamics in the CR system at varying engine speeds and WiD emulsion compositions and when properly harnessed can be used for condition monitoring/diagnosis. This paper proposes a machine learning (ML)-based diagnostic framework for CR diesel engines that functions on the use of continuous wavelet coefficients (CWCs) and short-time Fourier transform (STFT) outputs from the first-order derivatives of the RPS signals. Following an investigative experiment on a KIA Sorento 2004 four-cylinder line engine at various WiD emulsion compositions at various engine RPMs, diagnostic results show that exploring a multi-domain approach of feature extraction offers more discriminative diagnostic parameters thereby providing a wider perspective to condition monitoring and with the robustness of ML classifiers, accurate diagnostics can be achieved. A comparative assessment using global and local evaluation tools reveal the diagnostic performances of the ML algorithms and the trustworthiness of the most accurate model- random forest (RF) with a test accuracy of 96.65 % following a grid search against the other widely-used ML-based diagnostic tools.