Adding adaptive intelligence to sensor systems with MASS

In sensor systems, tracking gradual drift in a non-stationary environment is a challenging problem. The problem, a phenomenon also known as concept drift, is made even more difficult if the streaming data only consists of unlabeled data after initialization. This scenario is typically referred to as extreme verification latency (EVL), and is common in many sensor applications. In our previous work, we introduced a framework called COMPOSE (COMPacted Object Sample Extraction), which can handle the extreme verification latency problem, provided that the drift is limited. In this paper, we introduce a derivative of COMPOSE called MASS (Modular Adaptive Sensor System) as a solution to extreme verification latency in streaming sensor data, regardless of the particular application. To analyze the performance of MASS, the classification accuracy and execution time were compared to several variations of COMPOSE on synthetic benchmark datasets. The algorithm was then implemented on an Arduino sumo robot, where the objective was to keep the robot within a specific zone based on drifting data returned by the reflectance sensor.