Anomaly Detection of Environmental Sensor Data using Recurrent Neural Network at the Edge Device
JaeMyoung KIM, Young‐Wook Cho, Dohyun Kim · 2020
Advance in sensor technology brings numerous challenges with it in the context of data collection, storage and processing. Edge-enabled AI processing of sensor data is a large part of the sensor data processing. Sensor data about environments have natural errors and incompleteness in the collection process and need to be processed in real-time. Due to large volumes of data, monitoring and reporting of analyzed results need to be processed at the edge side of generated data. This paper proposes a time series data anomaly detection method that is based on neural network. Our models are evaluated using synthesis data generated from time series with trend, seasonal and noise component. In the case of zero-based dataset, GRU model with hidden cells (240 cells) and using only input values without additional features, applied with 0.3 dropout, produced 100% recall and 99.7% accuracy. In the case of non-zero-based dataset, LTSM model with hidden cells (240 cells) and using only input values without additional features, produced 86.7% recall and 99.5% accuracy. We also suggest the edge monitoring system with anomaly detection function of each environmental sensor using our pretrained detection model. Users can recognize an environmental status of the workplace using the prediction method with previous sensor outputs in real-time.