An image-classifying-based diagnostic method applied to electronic fuel injection system for gasoline engine

Abstract The electronic fuel injection (EFI) system significantly governs modern gasoline engine performance. Dirty sensors, weak actuators, broken electric wires, shorted circuits, and mechanical issues are most causes of abnormal system operations. Technicians can use the integrated onboard diagnostic (OBD) function to monitor system behaviour and then apply a standard Symptom-System-Component-Cause (SSCC) diagnostic process to determine what are causing rough idling, hesitate acceleration, excess exhaust gas emissions, and reduced fuel economy. This is a complicated and time-consuming procedure that is prone to human errors, particularly when mechanical causes occur, requiring skilled and experienced technicians. This study aims to demonstrate the feasibility of an automatic diagnostic solution for identifying causes of bad symptoms in EFI systems. Engine live data are collected from the OBD system while the engine is running through a diagnostic engine speed cycle. Live data features are then extracted and visualized as a radar-chart-type diagnostic image, which contains the signatures of normal or abnormal engine operations and can represent a specific cause of failure. By classifying the obtained diagnostic image, the cause of failure is identified. Experiments were conducted with a 4A91 engine, which calculates intake mass air flow from engine speed and intake air density instead of using an air flow meter. Mechanical issues such as cracked intake manifold, evaporative purge valve leak, and weak fuel pump were generated with different degrees of failures. Various pruned versions of the VGG-16 deep learning model were applied to classify diagnostic images thereby can indentify the causes of failure. The results showed that in the case VGG-16 model reduces 2 convolutional layers and 1 dense layer, it is most suitable for this automatic diagnostic application. The proposed diagnostic method can be further improved to detect multi-cause failures as well as predict system degradation trends for early warning and preventive maintenance.