An LSTM-Based Outlier Detection Approach for IoT Sensor Data in Hierarchical Edge Computing

Outlier detection in sensor data has recently gained significant recognition, particularly with the proliferation of wireless sensor networks (WSNs) and the Internet of Things (IoT). Several challenges face outlier detection in WSNs and IoTs, including sensor nodes’ limited energy and processing capabilities and high communication costs. This paper presents a novel deep learning-based outlier detection method for IoT sensor data in hierarchical edge computing. First, we proposed a hierarchical edge computing framework to save energy, provide load balance, and low latency data processing at sensor nodes. Then, we designed an outlier detection algorithm that resides on each edge server. The proposed algorithm consists of two modules: a predictor model and an outlier detector. The predictor module uses Long Short-Term Memory networks (LSTM) to forecast the subsequent data measurements of sensor nodes. The predicted values are then passed to an outlier detector module, which decides whether a data measurement is an outlier. The proposed method is evaluated on twelve different datasets. The simulation results provided by our proposal are promising in terms of a set of metrics.