AI × Embedded
Edge-AI Early Stone Detection for Agricultural Machinery
A reusable signal-processing and machine-learning pipeline for early detection of stones entering a harvester header using acoustic and machine-context signals.
12 / 28
Real events detected
0.697 s
Real-data advance
~11.5 KB
MCU model
2 MB PASS
MCU fit
Problem
The goal is to detect stone uptake early enough to protect agricultural machinery. Microphone data and machine operating context are used to predict an upcoming reference event generated by a metal detector.
Engineering pipeline
MF4 measurement files are loaded into a shared pandas representation, segmented into header-on episodes, windowed, labelled and transformed into acoustic and operating-context features. Grouped validation is performed by complete measurement run to avoid leakage.
Embedded deployment
A constrained Random Forest was exported to plain C arrays and deterministic tree traversal for inference on an automotive-class microcontroller. The current export contains 36 features and 8 trees and occupies roughly 11.5 KB of model data.
Limitation
The real-data baseline is not deployment-ready: 12 of 28 reference events were detected and false alarms remain too frequent. Synthetic-data experiments are much stronger, but are treated only as pipeline validation rather than evidence that the real-world problem is solved.