Automated Fault Detection Framework for Reliable Provision of IoT Applications in Agriculture

With the growth of Internet-of-Things (IoT), smart agriculture has become one of the most compelling IoT applications that supports crop management and better resource utilization. In this context, the quality of data gathered by widely distributed IoT edge devices has become critical to guarantee the accuracy of decisions in data-driven applications and cost-effectiveness. The data may be inaccurate and contain errors due to adverse environmental conditions or device faults. Supporting knowledge-based systems for monitoring and analyzing collected data to ensure the reliability of IoT services is vital. However, several challenges are encountered in fault detection for IoT applications, such as mimicking normal sensor behaviour by a faulty sensor, limited time and workforce. Also, the lack of labelled datasets containing both normal and real abnormal data points has affected the set of satisfactory data analysis methods. This work aims to propose a novel fault detection framework by utilizing a systematic feature engineering technique which is able to automatically identify abnormal data points, even nontrivial ones. The feature engineering technique helps to build a more reliable anomaly detection model, shortens the training phase, and preserves the sensor against unseen anomalies. Moreover, we provide collections of labelled datasets obtained from experimental situations from various sensors with and without sensor faults to evaluate our approach. The experimental results indicate that the proposed anomaly detection approach combined with the feature engineering technique outperforms established approaches, which are applied to the raw data without any features. It can be seen that extracting meaningful features is a pivotal step for having more precise anomaly detection.