Artificial Neural Network Predictive Autoencoder with Pre-Digital Signal Processing Unit

Abstract

In order to improve the quality of forecasting and detect anomalies in signals recorded from the outputs of sensors of automated process control systems (APCS), it is proposed to use an artificial neural network - a predictive auto-encoder with a preliminary digital signal processing (DSP) unit. It is shown that the preliminary DSP of the input predicted signal, consisting of a parallel set (comb) of digital low-pass filters with finite impulse responses (FIR-LPF), leads to non-equilibrium accounting for the correlations of time samples of the input signal and increases the accuracy of the prediction result. It is also shown that the predictive autoencoder (PAE) considered in the paper, in addition to restoring the PAE output of the input signal, additionally generates predicted samples of the input signal at the output, which also increases the accuracy of the prediction result. If anomalies occur in the signals (for example, as a result of the impact of cyberattacks), during the operation of the APCS, structural changes will occur in the error signal of the generated forecast, as a result of the analysis of these structural changes in the forecast error, anomalies are detected in the observed APCS processes.