Multi-Resolution Inter-Level Refinement (MR-ILR) Architecture for Anomaly Prediction in IoT Data

Abstract

We develop a novel architecture called "Multi-Resolution Inter-Level Refinement (MR-ILR)"architecture for anomaly prediction in Internet of Things (IoT) data. Our architecture is comprised of three modules: First the Inter-Level OR module takes the logical OR of the vector that represents the past anomaly states in IoT data and represents the occurrence of an anomaly state at increasingly coarse resolutions. Second a Multi-Layer Perceptron (MLP) predicts the occurrence of anomaly states at any given resolution. Third the anomaly predictions at successive resolutions are combined in Refiner modules to produce more accurate anomaly predictions. Our architecture provides the flexibility to produce anomaly predictions at distinct temporal resolutions. We compare the performance of our MR-ILR architecture against MLP and Long Short-Term Memory (LSTM) benchmark models. The results show that our architecture significantly outperforms both of these benchmark models with respect to the F1-score. This work represents an important advance in solving the challenging problem of anomaly prediction in IoT data and has the potential to be applied to a much wider range of problems that target anomaly prediction. © 2022 Elsevier B.V. All rights reserved.