Abstract: To address the bottleneck issues of insufficient load monitoring accuracy and poor stability caused by the complex and variable working conditions during wheat harvesting robot operations, this study quantifies the load status using the throughput and proposes an online load monitoring method based on the fusion of multi-source power flow sensing and intelligent algorithms. First, focusing on the energy transfer paths of key harvester operations, power sensors for the feeding auger, conveyor, and threshing drum were developed. Combined with engine CAN bus data, a highly synchronous multi-channel load monitoring terminal was integrated and developed based on virtual instrument technology, enabling synchronous acquisition and real-time monitoring of multiple key power flow parameters, including feeding auger power, conveyor power, threshing drum power, instantaneous engine fuel consumption, and engine torque percentage. Second, through the design and implementation of multi-condition field trials covering various wheat varieties, yields, and moisture contents, a high-fidelity correlation dataset between power flow parameters and throughput was constructed. A moving average algorithm was used to smooth the raw data, effectively suppressing random noise and impulse interference during field operations. Regarding the core modeling method, to overcome the insufficient stability of single-signal prediction, this study innovatively proposes a hybrid machine learning prediction model that integrates multi-sensor information. This model uses support vector regression (SVR) as the core predictor and deeply integrates the aforementioned multi-source heterogeneous sensor signals to comprehensively capture the intrinsic dynamic characteristics of throughput changes. To further break through the model's performance ceiling, an advanced nutcracker optimization algorithm (NOA) is introduced to automatically and globally optimize key hyperparameters such as the penalty coefficient and kernel function parameters in SVR, thereby significantly enhancing the model's prediction accuracy and generalization ability while ensuring computational efficiency. Experimental results fully demonstrate the superiority of the proposed method: sensor calibration tests show that the reference error of all self-made power sensors does not exceed 0.5%, meeting engineering accuracy requirements. Univariate linear regression analysis shows that the threshing drum power has the highest correlation with the throughput, with a coefficient of determination (R²) of 0.86; followed by engine torque percentage, instantaneous fuel consumption, conveyor power, and feeding auger power. The SVR-NOA hybrid model constructed in this study performs excellently on the independent test set, with an R² as high as 0.96 and mean squared error (MSE) and mean absolute error (MAE) as low as 0.23 and 0.39, respectively. Horizontal comparative analysis shows that compared with the unoptimized baseline SVR model, the absolute value of R² is improved by 0.06 (from 0.90 to 0.96); compared with the optimal univariate linear model based on threshing drum power (R²=0.86), the absolute value of R² is significantly improved by 0.10, and MSE and MAE are reduced by 66.7% and 40.9% respectively, fully demonstrating the significant advantages of the multi-source information fusion model. In the comparison of optimization algorithms, the SVR-NOA model also performs best, significantly outperforming the SVR-GWO model (R² leading by 0.05), and further improving the R² of the second-best SVR-PSO model by 0.02. The final online system verification results show that the monitoring system has a maximum absolute error of 0.37 kg/s and an average absolute error of 0.15 kg/s for real-time estimation of throughput, with the relative average error remaining stable within 2.5%, demonstrating excellent real-time monitoring accuracy and robustness. This study, through systematic innovation of "precise sensing-algorithm fusion-intelligent optimization", has achieved high-precision online perception of the operating load of combine harvesters, providing direct technical support for intelligent speed regulation, efficiency optimization and fault early warning of harvesters. It has important theoretical value and practical significance for promoting the development of grain harvesting machinery towards intelligence and precision.