Comparing hidden Markov models with artificial neural network architectures for condition monitoring applications

The use of hidden Markov models (HMM's), multilayer perceptrons (MLP's) and Kohonen self-organising maps (SOM's) has been proposed previously as efficient analysis and detection tools for condition monitoring of industrial plants and processes (Yin, 1993). Work on such applications with these techniques, has identified a need for a reassessment of these alternative recognition systems with a view to establishing their relative merits. In this paper the three systems are compared for two test data sets, one identifying the response of the systems to varying fault severity, the other showing recognition of faults which are independent of load. It is shown that for the MLP and SOM, implementing multiple networks improved the recognition of faults of varying severity. Possibly of more importance, this technique provided a means of diagnosing combinations of faults. For faults produced under differing load conditions, it is shown that the data cannot be classified by the SOM, and the supervised training regimes of the HMM and MLP provide the only means of classifying the data. Improved recognition is obtained with the HMM, although it is believed that the spectral pre-emphasis which was carried out on the input data could have contributed to this fact, suggesting that implementation of such pre-processing for the artificial neural network architectures, may be beneficial.