A framework for sensor fault detection and management in low-power IoT edge devices

Abstract

With the rapid development of the Internet of Things (IoT), reliable and energy-efficient provision of IoT applications is of utmost importance. In this context, the effective operation of IoT applications largely depends on sensor functionality, which can be compromised by various factors such as environmental conditions, vandalism, or sensor degradation. Ensuring the reliability and efficiency of \ac{IoT} applications requires robust, flexible, and effective tools for fault detection and management. However, achieving robustness, accuracy, and efficiency in current anomaly detection and management techniques is challenging. Faulty sensor measurements might not always manifest as obvious deviations from the norm and can resemble normal behavior. Conversely, legitimate fluctuations may be misinterpreted as anomalies. Moreover, the scarcity of labeled real-world faulty sensor data complicates the accurate evaluation of fault detection models. Additionally, the limited operational flexibility and resources of low-power edge devices introduce difficulties in timely fault management and reconfiguration. This thesis presents an advanced framework for sensor fault detection and management, to address the above challenges. The proposed fault detection solution, AssureSense, integrates a robust feature extraction method called TsAssure, which effectively identifies subtle and hidden faults in sensor data. TsAssure captures essential temporal, local, and spatial features from sensor measurements and their correlations, enhancing the understanding of sensor behavior. This capability enables AssureSense to detect anomalies with a shorter training phase, making it suitable for real-world applications. Once faults are identified, the framework also includes two strategies for remotely managing and mitigating the impact of faulty sensors. These strategies leverage the Over-the-Air (OTA) update paradigm to reconfigure the system. To evaluate the proposed methods, collections of labeled datasets were obtained from experimental situations containing both normal and faulty sensors. Testing with these datasets demonstrates that AssureSense outperforms existing methods in accurately detecting anomalies. Furthermore, experiments reveal that TsAssure surpasses other established feature extraction techniques in accurately capturing sensor behaviors and helps the fault detection techniques to more accurately detect abnormal measurements. In parallel, this thesis also addresses the challenge of collecting real-world faulty data by proposing a novel approach to model faulty sensor measurements, which generates more realistic synthetic data that can be used to train fault detection algorithms. Evaluating the proposed fault model in comparison with current models demonstrates that this approach more effectively represents sensor faults, leading to improved identification of real-world faulty data compared to traditional fault models.