[论文解读] Terracorder: Sense Long and Prosper

Josh Millar, Sarab S. Sethi+2 more|arXiv (Cornell University)|Aug 5, 2024

IoT and Edge/Fog Computing被引用 8

一句话总结

Terracorder 展示了一种在设备端实现的强化学习调度器，用于低功耗生物多样性传感器，在显著减少触发次数的同时实现更高的事件检测，并延长电池寿命，同时包括面向鲁棒性的协作网络调度。

ABSTRACT

In-situ sensing devices need to be deployed in remote environments for long periods of time; minimizing their power consumption is vital for maximising both their operational lifetime and coverage. We introduce Terracorder -- a versatile multi-sensor device -- and showcase its exceptionally low power consumption using an on-device reinforcement learning scheduler. We prototype a unique device setup for biodiversity monitoring and compare its battery life using our scheduler against a number of fixed schedules; the scheduler captures more than 80% of events at less than 50% of the number of activations of the best-performing fixed schedule. We then explore how a collaborative scheduler can maximise the useful operation of a network of devices, improving overall network power consumption and robustness.

研究动机与目标

• Develop a low-power, multi-sensor device for biodiversity monitoring with long deployment lifetimes.
• Introduce an on-device reinforcement learning scheduler to optimize event-driven sensor activations under power constraints.
• Evaluate isolated and collaborative scheduling to maximize useful active time and network robustness.

提出的方法

• Use ESP32s3-based PowerFeather hardware with deep-sleep and low-power I/O to build Terracorder.
• Implement an on-device Q-learning scheduler with a quantized action/state space to minimize activations while maximizing detections.
• Define a reward R_t = N_p_t - w1 * N_n_t to balance positive/negative activations for isolated scheduling.
• Extend the reward for network-wide collaboration to include overlaps and battery balance: R_t = N_p_t - w1*N_n_t - w2*sum_i(N_o_ti-1) - w3*B_sigma_t.
• Prototype with a 5MP camera, omnidirectional microphone, and PIR sensor; evaluate Goertzel filtering and a one-layer CNN on-device for event detection.
• Demonstrate practical on-device learning with a 24x7 Q-table (~3KB) and negligible overhead for Q-learning inference/update.

Figure 1 . Prototype configuration

实验结果

研究问题

• RQ1 Can an on-device reinforcement learning scheduler optimize event-driven sensing to extend device lifetime under power constraints?
• RQ2 How does a collaborative, network-aware scheduling strategy reduce redundant activations and improve overall network robustness and efficiency?
• RQ3 What are the practical performance and energy trade-offs of Goertzel filtering versus on-device CNN-based event detection in this architecture?

主要发现

• Isolated scheduling with Q-learning detects 85.3% of events, extending battery life by 55.1% over a fixed 3s baseline.
• Q-learning inference uses ~30 μA and updates ~70 μA with <0.1 s latency; Goertzel ~33.3 mA and TFLite ~33.1 mA with similar latency.
• Battery life extends from 0.69 to 1.07 years with Goertzel, and to ~1.05 years with TFLite-based event detection.
• The device achieves a deep-sleep current of ~100 μA while maintaining attached sensors and RTC; WiFi remains the largest power consumer for transmissions.
• Collaborative scheduling introduces a network-wide reward to minimize overlaps and balance battery life, enabling distributed, low-overhead coordination across nearby devices.
• A 24x7 Q-table (~3 KB) supports scheduling across 7 activation frequencies across 24 hours.

Figure 2 . Pareto curve of fixed vs. Q-learning schedules