Two-dimensional Dataset Reduction in Data-Driven Fault Detection for IoT-based Cyber Physical Systems

Internet of Things (IoT)-based Cyber-Physical Systems (CPS) are the integration of computational, networking and physical processes. IoT devices are used to enhance monitoring and control of such systems. CPS normal operation might be altered by physical component faults, which can be exploited to generate abnormal behaviour. Data-driven fault detection can be used in CPS to take advantage of the large amount of data being generated by IoT devices. However, the data can be noisy, corrupted or redundant, which can lead to misclassification and/or degradation in the performance of the detectors. A proper selection of data in terms of features and instances can improve the quality of the data and enhance the performance of such detectors. Moreover, it can allow for the implementation of lightweight but still accurate fault detectors, which are required in resource constrained edge-based IoT environments. This work studies the unified problem of instance and feature selection, for the purpose of two-dimensional reduction, in class-based fault detection datasets used in IoT-based CPS. To avoid the high complexity required by an optimal two-dimensional reduction approach, we examine order-based feature and instance dataset reduction and evaluate the new fault detectors which are trained based on the reduced datasets, in terms of accuracy gain/loss. A new weighted metric which combines accuracy and dataset reduction rates is introduced that enables the selection of an appropriate reduction scheme per application. Experimental results suggest that a proper selection of instance and feature selection algorithms can significantly reduce dataset size by up to 90%.