ThingsDND: IoT Device Failure Detection and Diagnosis for Multi-User Smart Homes

IoT systems are composed of devices with limited resources, and thus they are always prone to failure. In this paper, we propose ThingsDND, an automatic context-aware method for detecting and diagnosing sensor failures in smart home IoT verticals. The method involves three phases: a preprocessing phase that extracts the context; a training phase that provides the extracted context to an LSTM neural network to learn sensor data sequences; and finally, a prediction phase that predicts sensor data sequences to detect and diagnose failures. We evaluated ThingsDND against five smart home datasets with up to nine users. The results show that the method achieves an average F-score of 88% to detect the failure and an average F-score of 84% to diagnose faulty sensors. Moreover, ThingsDND imposes negligible performance overhead. It is worthwhile to mention that obtained results outperform other existing well-known methods.