PUB - Publikationen an der Universität Bielefeld

Accurate measurement of combine harvester separation process efficiency - the amount of grain separated from straw and material other than grain (MOG) - is crucial to process automation and economic efficiency in agriculture. Traditional methods rely on structure-borne noise sensors for grain counting. In most combines of leading harvester manufactures there is a single sensor line at the end of the separation unit to measure the amount of grain within the straw fraction, some manufactures are already using multiple sensors. The signal processing underlies disturbance variables like varying grain, field, and weather conditions, which interferes with predictions for the separation process. This study introduces a novel approach utilizing a large grid aligned sensor network within the combine harvester's separation unit, in conjunction with a suite of machine learning regression models. The models used to predict the separation efficiency are k-nearest neighbours (k-NN), support vector regression (SVR), decision trees (DT), fully connected neural networks (FCNN), convolutional neural networks (CNN), and recurrent neural networks (RNN). Through the analysis of sensor and combine setting data represented in one-, two-, and three-dimensional spatial tensor formats, we demonstrate that FCNNs and CNNs, especially with three-dimensional data representations, yield the highest prediction accuracy, while CNNs are the best models when just sensor data is considered. Most models achieved validation accuracy, evidenced by an average R² score of up to 90 % with k-fold cross-validation on the limited recorded data, thereby proving their reliability across varying conditions. Machine learning models like k-NN or DT with an R² score of up to 79 % could not achieve such validation results. A major difference between CNNs and most other machine learning models is the inherent focus on location-dependent data due to the convolutional calculations, which indicates to be the main benefit in analysing separation processes.