Modeling the process of failure prediction in radio electronic equipment using a recurrent neural network

This study examines the problem of data-distribution shift in predicting failures of radio-electronic equipment with recurrent neural networks. A typology of distribution-shift classes, their taxonomy, and classical detection methods are presented. Results of statistical tests for normality of the input data and an evaluation of the model’s hidden layer are reported; the analysis demonstrates that structural variations in the hidden layer indicate a shift in the data distribution even when the model’s output error remains unchanged. Finally, the forecasting process and distribution-shift detection are simulated by tracking state dynamics of the output hidden layer of a Long Short-Term Memory recurrent neural network.