SoftSensors for Aquaculture

About Model

The Problem Ammonia (NH3) and carbon dioxide (CO2) are critical, dangerous water-quality parameters in aquaculture, especially during live fish transport. Elevated NH3 damages gill tissue and drives in-transit mortality; rising CO2 suppresses oxygen uptake even when DO looks adequate, creating stress operators often miss until fish are already distressed. Yet these parameters are rarely monitored continuously, because dedicated NH3/CO2 probes are fragile, need frequent calibration, degrade fast in biologically active water, and are largely import-dependent — often costing 3.5x the price of a base IoT monitoring unit. This makes continuous monitoring commercially unviable for most small and mid-scale farmers and transporters. Our Platform: BlueSense Shuvoneel RAS Systems (Fishmongers, BlueFarm.ai) has already built and deployed BlueSense, a proprietary IoT unit for live fish transport and pond monitoring. It continuously tracks pH, DO, temperature, and TDS using rugged, low-cost sensors — but not NH3/CO2 directly. This use case closes that gap without new hardware. The Solution: AI Soft Sensing Instead of installing costly NH3/CO2 sensors, we infer these values from data BlueSense already collects. NH3 and CO2 dynamics are governed by known chemical/biological relationships with pH, temperature, TDS, and DO — non-linear patterns well suited to machine learning. We built a CNN-based model that takes real-time pH, temperature, DO, and TDS readings plus derived features to predict NH3 and CO2 levels, using the CNN's ability to capture temporal trends and rate-of-change signals that precede dangerous spikes. Feature Engineering We engineered derived features — rate-of-change trends, pH–temperature interaction terms (which determine toxic un-ionized ammonia fraction), and rolling stats reflecting biological load — enabling the model to detect the combined signature of multiple sensor trends moving together, flagging dangerous excursions before they become critical. Analytical Twin for Training Data Paired ground-truth NH3/CO2 data is expensive to collect at scale. We built an analytical (digital) twin replicating transport water chemistry, temperature dynamics, and oxygen consumption under varying stocking densities to generate chemically grounded synthetic training data. This lets the model train across far more conditions than field instrumentation alone could provide, ensuring robustness across diverse geographies and water qualities, and is continuously refined against real reference readings. Economic Impact Replacing physical NH3/CO2 sensors with AI inference eliminates hardware cost (up to 3.5x the BlueSense unit), calibration and consumables cost, breakdown-related downtime, and import dependency. We estimate ~₹9.2 lakh saved per truck per year, alongside a ~40% reduction in overall BlueSense maintenance burden. At scale — 10 trucks over 5 years — this projects to over ₹45 crore in total savings, making continuous NH3/CO2 awareness viable for operators who could never justify physical sensor hardware. Why This Matters Beyond cost, this enables fully local, failure-resistant, scalable water-quality intelligence. Since the model runs on data from hardware already deployed, there's no new failure point, calibration schedule, or import dependency — supporting predictive rather than reactive transport management, giving operators time to intervene before mortality occurs. Platform Fit This capability extends the BlueSense platform and plugs into our broader live fish transport monitoring stack, including the AI-IoT platform under the IndiaAI Innovation Challenge in West Bengal. Since BlueSense units are already field-deployed, NH3/CO2 soft sensing can reach them via software update — no hardware retrofit or added capital cost.