⚡ electrical engineering

Now or Never: Continuous Surveillance AIoT System for Ephemeral Events in Intermittent Sensor Networks

This paper proposes an Energy-aware Elastic Split Computing Algorithm for AIoT systems that overcomes the "Availability Gap" in intermittent sensor networks by dynamically offloading tasks to energy-rich neighbors, thereby enabling continuous surveillance and the capture of critical ephemeral events like poaching and forest fires.

Original authors: Joonhee Lee, Kichang Lee, Jeonggil Ko

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joonhee Lee, Kichang Lee, Jeonggil Ko

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to keep a 24/7 watch over a vast, wild forest to catch rare events like a poacher sneaking in or a tiny spark starting a fire. The problem? You can't plug these sensors into the wall, and you can't send a human to change their batteries every week. They have to run on solar power or wind, which means they are "intermittent"—they wake up when they have energy, do their job, and then fall asleep when they run out.

The authors of this paper, Joonhee Lee, Kichang Lee, and JeongGil Ko from Yonsei University, identified a major flaw in how these "sleepy" sensors usually work.

The Problem: The "Wait-and-See" Trap

Currently, most of these sensors operate like a sleepy baker. If the baker (the sensor) doesn't have enough flour (energy) to bake a whole cake (analyze a sound), they just sit there waiting until they have enough. They only wake up when their battery is full.

In a forest, this is a disaster. If a poacher steps on a twig or a fire starts, that event happens in a split second. By the time the sensor finishes "waiting" to gather enough energy to wake up, the event is already over. The sensor missed the "now or never" moment.

The Solution: The "Team Relay" Strategy

The researchers propose a new way to organize these sensors, which they call an Energy-Aware Elastic Split Computing Algorithm. Think of it not as a group of sleepy bakers, but as a relay race team where the runners help each other.

Here is how their system works, using simple analogies:

1. The "Energy Belief Map" (Reading the Room)
Instead of constantly shouting "How much energy do you have?" to their neighbors (which wastes precious battery), the sensors use a clever trick. They look at their own battery level and guess what their neighbor's battery looks like.

Analogy: If you and your neighbor both have solar panels on your roofs, and the sun just went behind a cloud for you, you can safely guess the sun went behind a cloud for your neighbor too. They don't need to call each other to confirm; they just "predict" the neighbor's status based on their own experience.

2. The "Coverage First" Schedule (No Gaps in the Blanket)
In the old system, if two sensors woke up at the same time, they would both look at the same spot, wasting energy, while a third spot nearby remained blind.

Analogy: Imagine a group of security guards. If they all wake up at 9:00 AM, they all stand in the same corner. The new algorithm acts like a smart dispatcher. It tells the guards to wake up at slightly different times so that someone is always watching every part of the forest, even if individual guards have to take naps.

3. The "Elastic Split" (Passing the Buck)
This is the core innovation. Deep learning (AI) is heavy work. If a sensor wakes up but only has a tiny bit of energy left, it shouldn't try to do the whole job.

Analogy: Imagine a heavy box (the AI task) needs to be moved across a room.
- Old Way: The person with the box waits until they have enough strength to carry the whole thing.
- New Way: The person picks up the box, carries it a few steps, and then hands it to a neighbor who has more energy. The neighbor carries it the rest of the way.
- In the paper's terms, the sensor processes the first few layers of the AI brain, then "offloads" the rest to a neighbor with a fuller battery. This ensures the event is captured even if the first sensor is almost dead.

The Results: Catching More "Now or Never" Moments

The researchers tested this idea using a computer simulation of 100 sensors in a forest, using real-world data on how much power their chips (ESP32-C3) actually use.

More Ground Covered: Their system managed to watch an extra 2,496 square meters (about the size of half a soccer field) compared to the old "wait until full" method.
More Events Caught: Because the sensors were awake more often and coordinated better, the system caught about 103 more critical events per day (like gunshots or animal calls) that would have been missed otherwise.
Better Accuracy: By splitting the work, they could use a much smarter, more complex AI model (98.4% accuracy) instead of a simple, dumb model (68.3% accuracy) that a single weak sensor could handle alone.

The Bottom Line

The paper argues that to protect nature effectively, we can't just rely on individual sensors waiting for a full battery. We need them to act as a cooperative team. By sharing the workload and predicting each other's energy levels, these tiny, battery-less devices can stay awake enough to catch the fleeting, critical moments that matter most in the wild.

Technical Summary: Now or Never: Continuous Surveillance AIoT System for Ephemeral Events in Intermittent Sensor Networks

Problem Statement

Wilderness monitoring tasks, such as poaching surveillance and forest fire detection, require pervasive, high-accuracy sensing to distinguish critical threats from background noise. While deploying Deep Neural Networks (DNNs) at the edge is essential for the precision required in these scenarios, maintaining stable power for large-scale deployments is logistically infeasible. Consequently, energy-harvesting AIoT is the only viable solution for pervasive deployment.

However, energy-harvesting systems introduce a fundamental conflict: frequent power interruptions clash with the computation-intensive nature of DNNs. Current intermittent computing strategies typically adopt a "store-and-compute" paradigm, where nodes accumulate energy and execute tasks only when sufficient power is available. This approach prioritizes eventual completion over continuous availability. The authors argue that this paradigm creates a critical "Availability Gap." In scenarios involving ephemeral, "now-or-never" events (e.g., a transient gunshot or the initial spark of a fire), the latency introduced by energy buffering can result in mission failure, as the event may cease before the node recharges and processes the data.

Methodology

To address the tension between individual device intermittency and the need for continuous spatiotemporal coverage, the authors propose a decentralized, energy-aware algorithm that leverages energy diversity across nodes. The system operates on heterogeneous harvesting conditions (e.g., sunny vs. shady areas) and utilizes the following core mechanisms:

1. Pre-Deployment Profiling and Static Mapping

Before deployment, the system undergoes an offline profiling phase using identical hardware specifications (ESP32-C3-SuperMini MCUs).

MCU Energy Usage Map: The system profiles power consumption across all operational states (sensing, computation, transmission, idle) to create a baseline for energy estimation.
DNN Layer Splitting Map: The target DNN model is pre-partitioned into logical layer groups based on profiled energy costs. This allows the system to adaptively determine the optimal partition point during runtime.

2. Energy-Adaptive Life-Cycle with Hysteresis

To prevent "zombie nodes" (nodes that repeatedly boot and fail due to insufficient energy), the system implements a hysteresis-based recovery mechanism.

Nodes enter deep sleep when energy drops below a critical lower threshold (e.g., 1%) and remain inactive until recharging to a recovery threshold (e.g., 20%).
Once active, nodes calculate a maximum sensing duration ( $T_{sense}$ ) based on residual energy, reserving a safety margin for essential computation and transmission.

3. Neighbor Voltage Prediction (NVP)

To facilitate cooperation without energy-intensive handshakes, the system uses a piggybacking mechanism where communication packets include the sender's current voltage.

Self-Referenced Prediction: Leveraging the spatial consistency of environmental energy, a node $i$ predicts the voltage of a neighbor $j$ ( $\hat{V}_j(t)$ ) based on its own voltage change ( $\Delta V_i$ ) and a learned offset. This allows nodes to build a local "Energy-Belief Map" and select optimal offloading partners without dedicated synchronization packets.

4. Decentralized Spatiotemporal Coordination

The algorithm employs two decentralized mechanisms to optimize coverage and computation:

Collision-Aware Desynchronization: If a node detects a neighbor's signal during its sensing phase, it identifies a spatiotemporal overlap. It then shifts its next wake-up schedule by a random backoff time to desynchronize sensing windows and eliminate redundancy.
Collaborative Offloading: During the handover window, if a node lacks sufficient energy to complete DNN inference, it processes only the feasible initial layers. Intermediate feature tensors are offloaded to the neighbor with the highest predicted energy surplus ( $\hat{V}_{neighbor} > V_{self}$ ), ensuring tasks flow toward energy-rich nodes.

Key Contributions

The paper outlines three primary contributions:

Coverage-First Scheduling: A novel intermittent-computing scheduling algorithm that minimizes sensing blind spots via lightweight state propagation. It coordinates wake-up cycles to ensure at least one node remains active in each target region, enabling the instantaneous capture of ephemeral events.
Energy-Aware Collaborative Inference: A collaborative computing algorithm for DNN inference driven by lightweight neighbor energy prediction. This enables dynamic task partitioning and offloading, allowing the network to bypass single-node limitations.
Hardware-Profiled High-Fidelity Simulation: A rigorous simulation grounded in real-world hardware profiling of the ESP32-C3-SuperMini. The simulation models a network of 100 nodes with precise energy harvesting and discharging dynamics to validate the algorithm's feasibility.

Results

The system was evaluated using a Python-based simulator modeling a hexagonal grid of 100 nodes with heterogeneous solar harvesting conditions (50% sunny, 50% shady). The simulation ran for approximately 2.78 hours, tracking 138 acoustic events.

Coverage and Stability: Compared to a baseline "Vanilla Mode" (reactive, single-node processing), the proposed Algorithm Mode achieved a 9.84% increase in average coverage (from 59.51% to 68.75%). The standard deviation of coverage dropped from 15.89 to 12.39, indicating significantly more stable and reliable surveillance.
Physical Impact: Translating these metrics to physical space, the proposed system covers an additional average area of 2,496 $m^2$ . This increased coverage implies the ability to capture approximately 9% more events (roughly 103 additional critical events per day) that would otherwise be missed due to nodes being offline or redundant sensing.
Inference Accuracy: A comparison between a standalone lightweight model (Small DNN) and the distributed split-computing approach showed a significant performance gap. The standalone model achieved 68.3% accuracy, whereas the distributed execution of a larger model achieved 98.4% accuracy, validating the necessity of split computing for high-fidelity inference in resource-constrained networks.

Significance and Claims

The authors claim that this work establishes a robust foundation for building resilient, fail-safe surveillance systems on resource-constrained nodes. By shifting the paradigm from "waiting for energy" to "continuous surveillance," the proposed algorithm addresses the critical availability gap in intermittent sensor networks.

The paper modestly notes that while the results are promising, the system has not yet reached its theoretical optimum. The current coordination algorithm relies on heuristic approaches to prioritize feasibility, leaving significant potential for further algorithmic refinement to maximize performance. Ultimately, the work aims to realize a truly ubiquitous AIoT system where high-performance DNN models are seamlessly executed across distributed environments, even under heterogeneous harvesting conditions.