Synergistic and Efficient Edge-Host Communication for Energy Harvesting Wireless Sensor Networks

TL;DR

Seeker addresses the challenge of running DNN inferences in energy-harvesting wireless sensor networks by tightly integrating edge computation with host-assisted processing. It introduces a store-and-execute edge paradigm plus application-aware coreset constructions to offload unfinished inferences to a mobile host, significantly reducing communication by up to 8.9x while maintaining high accuracy. The approach combines two lightweight DNN quantizations at the edge, memoization to skip redundant work, and recoverable coresets (including GAN-based reconstruction) to preserve inference quality on the host, achieving up to 86.8% top-1 accuracy on HAR tasks and substantial improvements over state-of-the-art EH-WSN methods. The results demonstrate strong practical benefits for energy-constrained IoT deployments, enabling reliable edge analytics in realistic EH scenarios with scalable hardware acceleration.

Abstract

There is an increasing demand for intelligent processing on ultra-low-power internet of things (IoT) device. Recent works have shown substantial efficiency boosts by executing inferences directly on the IoT device (node) rather than transmitting data. However, the computation and power demands of Deep Neural Network (DNN)-based inference pose significant challenges in an energy-harvesting wireless sensor network (EH-WSN). Moreover, these tasks often require responses from multiple physically distributed EH sensor nodes, which impose crucial system optimization challenges in addition to per-node constraints. To address these challenges, we propose Seeker, a hardware-software co-design approach for increasing on-sensor computation, reducing communication volume, and maximizing inference completion, without violating the quality of service, in EH-WSNs coordinated by a mobile device. Seeker uses a store-and-execute approach to complete a subset of inferences on the EH sensor node, reducing communication with the mobile host. Further, for those inferences unfinished because of the harvested energy constraints, it leverages task-aware coreset construction to efficiently communicate compact features to the host device. We evaluate Seeker for human activity recognition, as well as predictive maintenance and show ~8.9x reduction in communication data volume with 86.8% accuracy, surpassing the 81.2% accuracy of the state-of-the-art.

Figures (17)

• Figure 1: A primer on energy harvesting systems: Figure \ref{['Fig:EHPrimer']} shows the basic building blocks of an EH node equipped with sensing and computation. Some of the units change according to the harvested energy source. Figure \ref{['Fig:EHsota']} shows the capabilities of the current SOTA. The size of the circle representing the solutions depicts the compute capabilities of the sensor nodes, the shade shows the available power, and their position on the axes approximates the amount of compute done on the node and the amount of reliability on external communication. The power source is denoted in (Red) (notations used in Figure \ref{['Fig:EHsota']}: COTS: Commercial-off-the-shelf, Bat.: Battery, Bonito batteryfree, Chinchilla chinchilla, ResiRCA ResiRCA, Origin Origin)
• Figure 2: Accuracy comparison of various classical node-level optimization techniques. The Extended-Round-Robin policy (ERR) Origin takes a store-and-execute approach, and the number associated represents the ratio of store cycles vs execute cycles (e.g. RR3 is 3 store cycles followed by 1 execute cycle). The 'Baseline' model is a fully powered system with no energy restrictions, and the quantized model runs on harvested energy using a RR12 policy.
• Figure 3: An example of EH Sensor-Host ecosystem - the sensor transitions between multiple states and executes the compute as store and execute fashion Origin. The host receives the data in compressed form for the unfinished portion, decompresses it, runs inference and finally ensembles the results from multiple sensors to improve accuracy and robustness.
• Figure 4: A toy example of the coreset construction techniques in Seeker. Imp-sampling uses a probability based importance sampling; clustering preserves the geometric shape of the original data. In each case, the points/values in red are communicated to the host.
• Figure 5: Overall system design of Seeker