Neural networks for early prediction of machine failure

It is shown that both neural networks and the more usual parameter trending are useful in condition monitoring. For the data analysed, it appears that network training is best done using several different data sets, although it is noted that other work has yielded a different conclusion. Parameter trending was considered to be worthwhile only with two of the summary statistics discussed. Despite the obvious ease of use and effectiveness of parameter trending, neural networks are viewed as being more useful because they consider the data as a whole rather than as a series of individual plots. This has the advantage that, although some statistics may not be useful on their own, their combined information could be significant. It is not possible to detect this visually, but a neural network could identify it, and therefore has an additional source of information on which to base its output. It is planned that, after further experimentation on the default training technique, the results will be extended to form an artificially intelligent supplement to a condition monitoring program