A CPU-Efficient Robotic System for Instance Segmentation and Control of Mineral Fertilizer Granulation

The work presents a robotic optical system for controlling the granulation process of mineral fertilizers in industrial production. The key component of the system is a custom neural network for instance segmentation, built upon UNet and MobileNet v3 architectures, designed for real-time image processing. The network achieves a detection rate of 90% of fully visible granules on a central processing unit. To enhance the prediction of inter-granular space, a combined loss function was employed during training. The developed neural network is compared with known instance segmentation architectures, including performance evaluation on both graphical and central processing unit platforms. Furthermore, the hardware features and operational principle of the robotic optical system, ensuring continuous operation and self-diagnostics in industrial conditions, were described. Results from pilot industrial trials and subsequent implementation are provided, showing a divergence of less than 10% compared to standard laboratory analyses. An interrelation between fractional composition trend lines and the average granule diameter trendline for granulation process management was discovered. These findings can be leveraged to expedite and streamline the control of the granulation production process. Possible application schemes of the developed system for achieving full automation in the classification of produced granules were outlined.