Sensing, Scheduling, and Learning for Resource-Constrained Edge Systems

Abstract

Recent advances in Internet of Things (IoT) technologies have sparked significant interest in developing learning-based sensing applications on embedded edge devices. These efforts, however, are challenged by adapting to unforeseen conditions in open-world environments and by the practical limitations of low-cost sensors in the field. This dissertation presents the design, implementation, and evaluation of resource-constrained edge systems that address these challenges through time-series sensing, scheduling, and classification.

First, we present OpenSense, an open-world time-series sensing framework for performing inference and incremental classification on an embedded edge device, eliminating reliance on powerful cloud servers. To create time for on-device updates without missing events and to reduce sensing and communication overhead, we introduce two dynamic sensor-scheduling techniques: (i) a class-level period assignment scheduler that selects an appropriate sensing period for each inferred class and (ii) a Q-learning–based scheduler that learns event patterns to choose the sensing interval at each classification moment. Experimental results show that OpenSense incrementally adapts to unforeseen conditions and schedules effectively on a resource-constrained device.

Second, to bridge the gap between theoretical potential and field practice for low-cost sensors, we present a comprehensive evaluation of a sensing and classification system for early stress and disease detection in avocado plants. The greenhouse deployment spans 72 plants in four treatment categories over six months. For leaves, spectral reflectance coupled with multivariate analysis and permutation testing yields statistically significant results and reliable inference. For soils, we develop a two-level hierarchical classification approach tailored to treatment characteristics that achieves 75–86\% accuracy across avocado genotypes and outperforms conventional approaches by over 20\%. Embedded evaluations on Raspberry Pi and Jetson report end-to-end latency, computation, memory usage, and power consumption, demonstrating practical feasibility.

In summary, the contributions are a generalized framework for dynamic, open-world learning on edge devices and an application-specific system for robust classification in noisy field deployments. These real-world deployments collectively outline a practical framework for designing intelligent, cloud-independent edge systems from sensing to inference.