THE USE OF DIGITAL SIGNAL PROCESSING AND A NEURAL NETWORK WHEN GENERATING A FORECAST OF TIME SERIES OF DATA FOR THE PURPOSE OF DETECTING ANOMALIES IN THE IN THE AUTOMATED CONTROL OF TECHNOLOGICAL PROCESSES

Abstract

n order to detect anomalies and improve the quality of forecasting dynamic data flows observed from sensors in Industrial Control System (ACS)., it is proposed to use a predictive mod-ule consisting of a series-connected digital signal processing unit (DSP) and a predictive unit using a neural network (predictive autoencoder ( Auto Encoder), predictive Autoencoder (PAE)). The study showed that the preliminary DSP block of the predicted input signal, consisting of a parallel set (comb) of digital low-pass filters with finite impulse responses (FIR-LPF), leads to a non-equilibrium account of the correlation relationships of the time samples of the input signal and to increase the accuracy of the final prediction result. The predicted autoencoder (PAE) pro-posed and considered in the work, in addition to restoring the input signal or part of the input signal at the PAE output, also generates the predicted samples of the input signal for the speci-fied number of «forward» time steps at the output, which increases the accuracy of the predic-tion result. The reduction of the forecast error occurs due to the imposition of restrictions in the formation of the forecast, that is, an additional requirement to restore the input samples of the samples – «stabilizers» at the NS output. The introduction of «stabilizers» increases the accuracy of the prediction result.