Pre-processing of acoustic signals by neural networks for fault detection and diagnosis of rolling mill

Incipient faults and changes in the structure of any industrial process may be detected and known by their effects: vibration and/or acoustic signals. We consider some methods for pre-processing the acoustic signal, generated by a rolling mill process, for fault detection and structural classification. The pre-processing methods are based on artificial neural networks. The methods refer to: signal decomposition algorithms, a distances measure for spectral amplitude classification and neural network structures for spectrum compression. For the signal decomposition problem an adaptive neural network algorithm is proposed in which the number of inputs is adapted to the imposed error. When the training error for two successive steps is very little, then the number of inputs in network is increased. If the spectral components are zero for sufficient time, then the number of inputs is decreased. The Hausdorff distance is proposed for spectrum classification as the distance measure for the frequency domain in a pattern recognition context. It shown that the Hausdorff distance has a monotone relationship with the signal-to-noise-ratio. Finally, the possibility of decreasing the number of spectrum components as patterns is presented, by compression with neural networks. Spectral representations of the acoustic source show that signatures collected at rolling mill sensor locations can be successfully used to identify process and structural changes in the rolling mill monitoring system. The results obtained by simulation is encouraging for real-time implementation