מיכאל פייר

Tiny machine learning (TinyML) enables trained models to run on microcontrollers, allowing inference at the sensing node rather than only in the cloud. This can reduce latency and bandwidth usage by enabling local estimation and event reporting. However, complete pipelines that combine continuous sensing, embedded validation, and decision-relevant outputs remain limited in aquaculture systems. Here, we developed an embedded TinyML soft-sensor workflow in a biofloc-based recirculating aquaculture system (B-RAS) comprising a fish tank and a side-stream tank (biofloc tank). Multiple low-cost water-quality sensors were integrated into an Internet of Things (IoT) microcontroller node to continuously measure physical parameters, while standard laboratory tests provided complex water-quality parameters. Using paired sensor and laboratory data, we predicted laboratory parameters (i.e., soft sensors). We compared several models (linear regression, decision trees, random forests, AdaBoost, k-nearest neighbors, and XGBoost) and selected the best model for each target using cross-validation. The chosen models were exported to embedded C and deployed on the microcontroller for real-time edge inference. Performance in the biofloc tank was generally moderate to strong, while on-device validation remained target-dependent. Fish-tank prediction was limited due to a weaker sensor-target link and model compression. These results show that TinyML can support practical on-device estimation of selected laboratory-dependent water-quality parameters, particularly when online sensor signals are strongly coupled to the target process. However, deployment reliability varied by target and location, with weak fish-tank performance indicating that embedded soft sensors should be used as decision-support tools rather than replace laboratory analysis.