The application of digital signal processing technology to improve the accuracy of forecasting time series data in anomaly detection systems in the observed processes of automated process control systems

Abstract

Forecasting dynamic time series of data plays an important role in the construction of
anomaly detection systems for information protection in various automated process control
systems (APCS). To improve the accuracy of forecasting dynamic time series of data, preliminary
digital processing (digital filtering) of signals is used to decompose the observed time series
coming from the APCS sensors into separate components. With this approach, component decomposition and filtering of the original signal are performed using a digital filter comb, which
significantly improves the quality of the generated forecast. The error signal of the prediction
result was analyzed using digital spectral and bispectral analysis. It is shown that for the case of
«perfect prediction» the prediction error signal is an unpredictable residual, that is, it tends to
the state of white noise. The paper shows that the analysis of the forecast error using the methods of digital spectral and bispectral analysis makes it possible to form an assessment of the
quality of the forecast result. The comparison shows a significant increase in the efficiency of
using preliminary digital filtering to improve the accuracy of forecasting the observed dynamic
time series of APCS data. Work with neural networks was carried out in the MATLAB «Deep
Learning Toolbox» extension package. For spectral and bispectral analysis of signals, the «Higher Order Spectral Analysis Toolbox» package was used