index a0b83e8..9a3d9ea 100644
@@ -1,7 +1,14 @@
---
+first_captured: 2026-04-10
+sources:
+- sources/google-sheets-ideas.md
status: raw
tags:
- hardware
+- defense
+- embedded
+- ml
+- audio
title: acoustic drone detection
type: idea
updated: 2026-04-11
@@ -10,4 +17,8 @@ visibility: public
# acoustic drone detection
-audio scene classification on MCUs for defense/conservation.
\ No newline at end of file
+audio scene classification running on microcontrollers (MCUs) — the core premise is detecting drones, aircraft, or wildlife by sound without requiring a networked computer. the motivating insight is that visual detection fails at night or in dense foliage, but acoustic signatures are reliable and can be processed entirely on cheap embedded hardware. defense applications include perimeter monitoring and early warning; conservation applications include detecting illegal poaching aircraft in wildlife reserves.
+
+the hard technical problem is running a capable ML model on constrained hardware — MCUs have kilobytes of RAM and no GPU. this pushes toward techniques like quantization, knowledge distillation, and TinyML frameworks (TensorFlow Lite Micro, Edge Impulse). the classification pipeline: raw audio → spectrogram features → lightweight CNN or RNN → class label. the drone-vs-background distinction requires good negative examples (wind, insects, vehicles) to avoid false positives, which makes dataset construction a meaningful research contribution. similar embedded ML challenges appear in [[predictive-maintenance-sensors|predictive maintenance sensors]] and [[eeg-artifact-rejection|EEG artifact rejection]].
+
+this sits at the intersection of the [[cluster-hardware-wearable|hardware and wearables]] cluster and serious ML research — it is not a typical web app, which is part of why it scores highly on the spreadsheet evaluation. the defense/conservation dual-use angle gives it real-world traction. connects to [[agent-simulation|agent-based simulation]] if you want to model acoustic environments for training data generation, and to [[ppg-biomarker-wearable|PPG biomarker wearable]] as another example of novel signal → trained model on constrained hardware.
\ No newline at end of file