CNN-Based Hotspot Detection in Small-Scale Composting From Multistatic Microwave Time-Domain Responses

Small-scale composting scenarios are highly sensitive to localized temperature peaks, known as hotspots, which can disrupt the overall composting process and degrade its quality. To address such issues, this study presents a deeplearning method to detect such hotspots using multi-static timedomain scattering responses. A large dataset is generated with an in-house full-wave Finite-Difference Time-Domain (FDTD) electromagnetic solver. The dielectric properties of compost and hotspot regions are modeled based on existing literature, and synthetic White Gaussian noise is added to recorded time-domain responses to mimic more realistic measurement conditions. The framework includes a physics-informed feature-extraction stage that encodes the time-domain signal signatures into feature vectors. These vectors are then processed by a compact 2-Dimensional Convolutional Neural Network (2D-CNN) to classify scattered signals in three categories: no hotspot, one hotspot, or two hotspots. The results indicate that the framework achieves over 95% accuracy at signal-to-noise ratio (SNR) levels similar to those found in practical microwave sensing systems.