The paper "Method of Natural Data Analysis and Anomaly Detection Based on a Collective of NARX Neural Networks" has been published in a Special Issue of the "Pattern Recognition and Image Analysis" Journal by Springer

The authors are the Researchers of the Laboratory of System Analysis of Atmospheric Processes of IKIR FEB RAS, Dr. Sci. (Phys.-Math.) Mandrikova O.V., Cand. Sci. (Tech.) Polozov Yu.A., Cand. Sci. (Tech.) Mandrikova B.S.

Neural networks expand the possibilities of classical methods of data analysis and are effective in the case of a high uncertainty of knowledge about the information component and interference.

However, the elements of heuristics in the neural network approach complicate the process of finding the optimal solution for a particular practical problem and require the synthesis of new approaches on the neural network basis. The article proposes a method of building a time series model of ionospheric parameters by creating a collective of NARX neural networks (NNC).

When building a NNC for data preprocessing, a threshold wavelet filtering is used. It makes it possible to solve the well-known NARX problem related to the vanishing gradient effect. The proposed approach is considered for the problem of analysis of ionospheric parameters and detection of anomalies.

Ionospheric anomalies lead to malfunctioning of modern technical means, navigation systems, etc. Therefore, the analysis of ionospheric anomalies is relevant and carried out by a large number of authors and scientific groups.

The developed anomaly detection algorithm includes NNC-based data modeling and anomaly detection rule. Comparison with the results of previous studies and the traditional way of application of NARX networks has shown the effectiveness of the method.

The NARX network architecture is shown in the Figure.

Full text of the paper is available at link.

For reference: Mandrikova, O.V., Polozov, Y.A. & Mandrikova, B.S. Method of Natural Data Analysis and Anomaly Detection Based on a Collective of NARX Neural Networks. Pattern Recognit. Image Anal. 34, 1223–1232 (2024). https://doi.org/10.1134/S1054661824701293