Sensing the Unknowns: A Study on Data-Driven Sensor Fault Modeling and Assessing its Impact on Fault Detection for Enhanced IoT Reliability

Abstract

In the context of the Internet of Things (IoT), the effective operation of IoT applications heavily relies on the functionality of sensors. These sensors are prone to failures or malfunctions due to various factors, including adverse environmental conditions and aging components within sensors. To mitigate the impact of faulty sensors on system performance, notable research has focused on employing machine-learning techniques to detect faulty sensor data. In this context, due to the scarcity of real faulty data records and challenges in generating them even in controlled environments, researchers often model faulty data to create synthetic datasets containing normal and abnormal data for evaluating fault detection models. Our empirical investigation reveals that the current modeling approach to simulate faulty sensor scenarios does not adequately mirror the complexity of real-world faulty sensor behaviors. Therefore, to improve the efficacy of fault detection algorithms in practical applications, it is imperative to investigate sensor fault models further. To address this gap, we conducted a comparative analysis of existing fault models and proposed a novel composite approach for modeling faulty sensor behaviors that can more effectively capture real-world sensor behaviors. Our focus was to evaluate how different fault models impact the effectiveness of anomaly detection algorithms when tested in real-world scenarios. The evaluation included algorithms trained on synthetic datasets derived from various fault models, assessing their performance in identifying real-world faulty data. We also provide diverse labeled datasets, including normal and abnormal data collected from real-world applications.