Intelligent Fault Diagnosis in Gasoline Engines Using Convolutional Neural Networks

This research focuses on the application of convolutional neural networks (CNNs) for fault detection in ignition coils and fuel injectors of a YESA 3140 gasoline engine. The objective is to design a CNN capable of identifying when the spark ignition engine (SIE) is operating under optimal conditions and when it presents specific power supply disconnection faults in the four injectors and four coils. Signals from the knock sensor (KS) and camshaft position sensor (CMP) of the SIE were acquired using a MyDAQ data acquisition card and LabVIEW software version 2024. A strict sampling protocol was followed: each replicate had a duration of 5 s while the engine was running at normal operating temperature and idle speed. Prior to each sampling, the SIE was operated with the corresponding fault induced for 5 min. The signals obtained from the KS sensor were transformed into spectrograms, which were then used to train various CNN models. The resulting CNN achieved a classification error of 3.21%. The algorithm was validated by inducing supervised faults in various Otto cycle engines.