Measuring the mass flow of sugarcane in real-time is essential for harvester automation and crop monitoring. Data integration from multiple sensors should be an alternative to receive more reliable, accurate, and valuable predictions than data delivered by a single sensor. In this sense, the objective was to evaluate if the fusion of different sensors installed in a sugarcane harvester improves the mass flow prediction accuracy. A harvester was experimentally instrumented, and neural network models integrated sensor data along the harvester to perform the self-calibration of these sensors and estimate the mass flow. Nonlinear autoregressive networks with exogenous input (NARX) and multiple linear regression (MLR) models were compared to predict the mass flow. The prediction with the NARX showed a significant superiority over MLR. MLR decreases the estimated mass flow variability in the harvester. NARX with multi-sensor data has an RMSE of 0.3 kg s-1, representing a MAPE of 0.7%. The fusion of sensor signals improves prediction accuracy, with higher performance than studies with approaches that used a single sensor. The mass flow approach with multiple sensors is a potential approach to replace conventional yield monitors. The system generates accurate data with high sample density within sugarcane rows.
Keywords: artificial intelligence; data fusion; machine automation; precision agriculture; self-calibration; yield monitor.