APPLICATION OF NEURAL NETWORKS TO CONDITION MONITORING AND PREDICTION OF ACOUSTIC FATIGUE FAILURES PACS REFERENCE:

Under consideration is condition monitoring of thin-walled structures, which suffer fatigue failures resulted from high-level wide-band loads in acoustic frequency range. Destructiveness is estimated with the aid of spectral characteristics of structure parameters measured in checkpoints. It was shown that Kohonen networks made it possible to diagnose a structure when neither all possible types of destruction nor the corresponding changes induced in the spectral characteristics were unpredictable in advance. The ways of reduction of input data dimension to improve the learning and operation network quality are described. The system of statistical goodness-of-fit tests for estimating the effectiveness of network functioning is proposed. Prediction of damage probabilities is carried out on the base of accumulated observations. Probability dynamics is described by continuous time, discrete state Markov processes. The method of minimum chi-square is used to identify unknown model characteristics by means of comparison of observed and expected histograms presenting damage distributions at the given time points