On-board neural network-based sensor fault diagnosis system in automotive engine

In the new millennium, automobile designs emphasize improved performance, safety and comfort with reduced exhaust emissions which must comply with increasingly stringent federal and states laws. These require an increasing number of actuators and sensors for the sophisticated control techniques applied in the modern automobiles. In order to promote the reliability and maintainability of automotive systems, an on-board fault detection and isolation (FDI) system capable of detecting and identifying actuator and sensor faults is necessary.; In this work, an on-board two-stage neural network-based fault detection and isolation system in a highly nonlinear internal combustion (IC) engine is proposed. The first stage generates the residuals, computed as the difference between the actual and predicted engine system response. The task of the second stage network is to classify the residual vector pattern corresponding to the various preset faults that are to be detected and isolated in the engine system. In order to implement an on-board detection system for an IC engine, a load torque estimator using a neural network is also developed. Furthermore, in order to obtain input-output pairs training data for Lervenberg-Marquardt backpropagation network algorithm, a nonlinear identified engine model is validated and used.; The performance evaluation of the proposed on-board FDI system is investigated by addressing issues related to the detection and isolation of engine sensor faults of varying severity in a variety of different noise level. The simulation results demonstrated that the two-stage neural network-based FDI system is a very promising method for detecting and isolating sensor faults in a significantly nonlinear IC engine system.