Anomaly detection by neural network models and statistical time series analysis

The problem of detecting weak anomalies in temporal signals is addressed. The performance of statistical methods utilizing the evaluation of the intensity of time-dependent fluctuations is compared with the results obtained by a layered artificial neural network model. The desired accuracy of the approximation by the neural network at the end of the learning phase has been estimated by analyzing the statistics of the learning data. The application of the obtained results to the analysis of actual anomaly data from a nuclear reactor showed that neural networks can identify the onset of anomalies with a reasonable success, while usual statistical methods were unable to make distinction between normal and abnormal patterns.<>