A Multi Model Approach: A Data Engineering Driven Pipeline Model for Detecting Anomaly in Sensor Data using Stacked LSTM

Abstract

Abstract: This study uses a self-supervised learning technique based on auto encoders to find anomalous nodes. Only temporal variables have been taken into account and researched so far for use in identifying anomalies in wireless sensor networks (WSNs). This method fully utilises the geographic and temporal information of the WSN for anomaly identification by incorporating the extraction of geographic location features, intermodal WSN correlation features, and temporal WSN data flow characteristics into the design of the autoencoder. First, by focusing on a single mode from a local spatial perspective, a fully connected network is used to temporal nodes. Second, the spatial and temporal characteristics of the data flows of the nodes and their neighbours are retrieved by concentrating on a specific mode and seeing the WSN topology from a global spatial perspective for anomaly identification. The adaptive fusion method's weighted summation step is then used to extract the relevant features from the various models. An LSTM is used in this study to solve the problem of long-term dependence in the temporal dimension. The decoder's reconstructed output and the hidden layer representation of the are used to calculate the anomaly probability of the current system utilising a fully linked network.