Neural networks for engine fault diagnostics

A dynamic neural network is developed to detect soft failures of sensors and actuators in automobile engines. The network, currently implemented offline in software, can process multi-dimensional input data in real time. The network is trained to predict one of the variables using others. It learns to use redundant information in the variables such as higher order statistics and temporal relations. The difference between the prediction and the measurement is used to distinguish a normal engine from a faulty one. Using the network, we are able to detect errors in the manifold air pressure sensor (V-s) and the exhaust gas recirculation valve (V-a) with a high degree of accuracy.