钠瞳 Sodium Pupil：让野外相机先算账，再花钠电

摘要

野外生态监测设备长期工作在无电网、无人维护、低温与低光照并存的极端环境中。它们的失效原因往往不在感知能力不足，而在缺乏对有限能量的主动分配：设备只能"要么睡、要么用力看"，无法区分值得花电的事件与大量误触发。

钠瞳把野外视觉节点的感知动作拆成四档——睡、看、认、传，让一次低成本的"瞥一眼"（约 1 mWh）去决定是否值得唤醒昂贵的识别与回传（约 124 mWh），并由一个在端侧运行的决策器，按剩余电量、温度、光照预测与事件价值分配能量，同时守住"不能把自己饿死到日出"的底线。

我们在同一光照、温度、触发与事件序列下，对三种做法做了受控对照，并做了电池化学消融、参数敏感性分析与极寒压力测试。结果表明：在日日照 6.27 Wh 的冬季山脊条件下，钠瞳的事件捕获率为 92.5%，全天耗电 13.8 Wh，全程存活；常规相机与定时阈值策略则分别在 8.8 小时与 18.4 小时后停止工作。参数在 ±100% 范围内扰动时，20 次重复的捕获率均值为 89.8%、标准差 4.4，结论未发生翻转。

其中有一个否定性结论值得单独说明：在能量宽松的常规工况下，把钠电换成磷酸铁锂，捕获率完全不变。这说明策略才是决定存活的主导因素，钠电的价值具体成立于低温可充电与低温可用功率两个物理环节。我们把它写进了第五、六部分，而不是藏起来。

Abstract

Wildlife monitoring devices in remote areas have to work in extreme environments where there is no grid, no maintenance, and where low temperature and low light occur at the same time. They usually fail not because their perception is insufficient, but because they cannot actively allocate a limited amount of energy: a device can only either sleep or look hard, and it cannot distinguish an event worth spending energy on from thousands of false triggers.

Sodium Pupil splits the perception actions of a field vision node into four levels — Sleep, Peek, Identify and Report. A cheap "peek" (about 1 mWh) decides whether it is worth waking up the expensive identification and transmission stage (about 124 mWh). A decision engine running entirely on the device then allocates energy according to the remaining charge, temperature, solar forecast and event value, while holding one bottom line: it must not starve itself before sunrise.

Under identical solar, temperature, trigger and event sequences, we ran a controlled comparison of three approaches, together with a battery-chemistry ablation, a parameter sensitivity analysis and an extreme-cold stress test. The results are as follows. On a winter ridge day with 6.27 Wh of daily solar energy, Sodium Pupil reached an event capture rate of 92.5% with 13.8 Wh of daily consumption and survived the whole day; the conventional camera and the fixed-threshold strategy stopped working after 8.8 hours and 18.4 hours respectively. When parameters were perturbed within ±100%, the capture rate over 20 repetitions averaged 89.8% with a standard deviation of 4.4, and the conclusion did not flip.

One negative result deserves a separate note. Under energy-abundant normal conditions, replacing the sodium-ion battery with LFP leaves the capture rate completely unchanged. This shows that the strategy is the dominant factor for survival, and that the value of sodium-ion batteries is specifically located in two physical links: low-temperature charging and low-temperature available power. We wrote this into Parts 5 and 6 instead of hiding it.