FAILURE PREDICTION METHOD FOR INTERNET OF THINGS NETWORKS USING MACHINE LEARNING

Abstract

The article addresses the problem of ensuring the reliability and uninterrupted operation of Internet of Things networks consisting of a large number of sensor nodes, gateways, and distributed computing elements. Due to the high heterogeneity of devices, rapid topology changes, and heterogeneity of data flows, such networks are vulnerable to various types of failures—hardware, network, and software. In view of this, fault prediction methods capable of detecting risks of system destabilization in advance are becoming increasingly relevant. The advantages of using hybrid machine learning approaches that combine time series analysis and the assessment of spatial interaction between network nodes are substantiated. A fault prediction method LGFP is proposed, built on the combination of graph neural networks (GNN) and LSTM architecture, which provides comprehensive data interpretation. The method allows estimating the probability of failure occurrence based on current and previous telemetry parameters while considering the mutual influence of network elements. An analysis of existing approaches is carried out, a comparison of machine learning models is performed, and the process of dataset formation for the study is described. Special attention is paid to the problem of class balancing and filtering of noise structures in the data, which are critically important stages for improving prediction accuracy. The obtained experimental results demonstrate the advantage of the proposed method compared to traditional models such as Random Forest, SVM, and isolated LSTM architectures, which is confirmed by increased classification accuracy and F1-score. The practical application of the proposed approach can provide a significant improvement in the robustness of IoT systems in industrial, energy, and household environments. Prospects for further research include the expansion of the feature space, the integration of attention mechanisms to improve model interpretability, and the testing of the method in real industrial conditions.