Engine diagnosis system for automative industry

Abstract

The condition monitoring based on sound and vibration detection has benefited the machinery industry. Endless efforts have been put into the research of fault diagnosis based on sound. It offers concrete economic benefits, which can lead to high system reliability and save maintenance cost. Artificial Neural Network is a very demanding application and popularly implemented in many industries including condition monitoring via fault diagnosis. Artificial Neural Network has been implemented successfully in many aspects, but the implementation needs detailed knowledge and studies. In this work, the noise from the vehicle engine is recorded using a suitable experiments setting up and the noise signature of the vehicle engine noise is obtained and identified. Tests have been carried out to record the noise signal from different vehicles of the same model to identify the noise signature. The acoustical signals are collected by a set of directional microphones, which transducer the sound pressure signals affected by mechanical vibrations emitted by the transmission. The set of signals collected will contain fault signatures as well as signals from other interfering sources. The captured signal, which is in analog forms, is digitized. In digitizing process; the signals are digitized using different frequency octave. Here we are using full octave band. Digitized signals are used to generate frequency power spectrum. This permit the sound intensity level with its corresponding frequency obtained from the frequency power spectrum. These data will go through a pre-processing stage. The noise signature is obtained by applying two different methods of pre-processing, frequency spectrum analysis method and the Principal Component Analysis method. Both methods extract the noise signature of the recorded noise. The result of the pre-process will used as the input in the neural network model. Based on the noise signature, neural networks models diagnose the vehicle faults. Two different neural network architectures are proposed in this research, back propagation network and Learning Vector Quantization network. The performances of the two different pre-processing techniques are evaluated and the neural network architectures are compared in terms of accuracy, efficiency and speed.