Implementation of a Neural Network-based In-Vehicle Engine Fault Detection System

Abstract

Arti cial neural networks (ANNs) are a powerful processing units inspired by the human brain. They can be used in many applications due to their pattern classi cation abilities, ability to model complex nonlinear input-output mappings, and their ability to adapt and learn. The relatively new Smooth Variable Structure Filter (SVSF) has recently been applied to the training of feedforward multilayered neural networks. It has shown to have good accuracy and a fast speed of convergence. In this thesis, an engine fault detection system using an ANN will be implemented. ANNs are used in engine fault detection due to the high-noise environment that engine operate in. Additionally the fault detection system must work while the engine is mounted in a vehicle, which provide additional sources of noise. The SVSF training method is evaluated and compared to other traditional training methods. Also di erent accelerometer types are compared to evaluate whether lower cost accelerometers can be used to keep the system cost down. The system is tested by inducing a missing spark fault, a fault that has a complex fault signature and is di cult to detect, especially in an engine with a high number of cylinders.