Hardware implementation of photoseparator pneumatic rejection subsystem with time delay compensation

Abstract

A hardware implementation of a pneumatic sorter rejection subsystem designed to remove hard-to-separate foreign matter from a grain flow in real time is discussed. The subsystem is modular and includes an industrial CMOS camera with a global shutter (1280 × 1024 pixels, 255 fps), a computational node with a graphics processor, a microcontroller unit based on the AVR architecture with 14 solid-state relay channels, and a pneumatic system with high-speed valves. To minimize the total delay between image capture and actuator response, a compensation method based on a predictive tracker with exponential speed smoothing and a deterministic binary data exchange protocol is proposed. Asynchronous multithreaded software architecture is developed that separates machine vision (YOLOv8) and actuator control processes. Bench test results confirm the stability of end-to-end video capture without dropped frames and a neural network inference time of 65–80 fps. Bench tests were conducted on a newly developed experimental setup which differs from all existing setups in that the vibrating chute is equipped with rectangular plate guides to force grain particle orientation along the direction of movement for rapid particle recognition. The video cameras are equipped with polarizing filters to eliminate glare and are flanked by light sources. The control computer is equipped with software including a convolutional neural network trained on a dataset containing images of the main crop seeds and seeds of difficult-to-separate impurities. The experiments have shown that the jitter in control command transmission does not exceed 1 ms confirming the effectiveness of the proposed architectural solution for precision optical separation tasks.