Fault diagnosis application in an automotive diesel engine using auto-associative neural networks

The application of a new method for fault diagnosis in an automotive diesel engine is presented. Two common types of fault are investigated: (i) sensor fault, caused by a bias in the inlet manifold pressure sensor and (ii) process fault, caused by small air leaks in the inlet manifold plenum chamber. Such faults may lead to increased emission levels which, if left undetected, can eventually degrade engine performance. A diagnostic model using a variant of auto-associative neural networks (AANN), which has a unique architecture which shows great potential for fault diagnosis, is investigated. This new variant uses a reduced set of original data for the input-target network and is denoted as a generalized serial T2T (GST2T) network. The proposed GST2T model is experimentally validated using data from a real engine embedded in a chassis dynamometer at a test-cell facility. It is demonstrated that using just one diagnostic model, the different types of common faults mentioned above can be accurately detected and diagnosed. This new GST2T network constitutes a nonlinear extension of linear principal component analysis