Machinery Diagnostics Using Polynomial Neural Networks

The vibration signatures (mechanical, acoustic, and electromagnetic) produced by machine components may be used for machinery diagnostics. By regularly measuring vibration levels, defects can be detected and diagnosed before causing extensive damage or failure, a process known as predictive maintenance. The main advantage of predictive maintenance is that problems can be identified without disassembling a machine, or even removing it from service. Conventional machinery diagnostics generally require significant human involvement and expertise. Essential requirements include data pre-processing for feature extraction, detection of signals of interest, and classification of these signals. Automated systems often utilize deductive approaches that fail to capitalize on the benefits offered by inductive techniques such as neural networks; these benefits include performance advantages and reduced development and maintenance. Machine diagnostics is essentially a pattern- recognition task, for which neural networks are ideally suited due to their speed and ability to recognize complex high-dimensional relationships. Classification polynomial neural networks emphasize discrimination among fault classes and thereby offer advantages over estimation neural networks for diagnostics applications. Signal-processing and pattern-recognition algorithms, and the hardware on which they run, are sufficiently advanced for rapid on-line detection and classification of changes in the condition of electro-mechanical systems. -