基于Jetson Nano处理器的大蒜鳞芽朝向调整装置设计与试验

为满足大蒜定向播种的农艺要求,针对现有大蒜鳞芽调整方法对杂交蒜适应性差的问题,该研究设计了一种基于Jetson Nano处理器的大蒜鳞芽朝向自动调整装置。采用双卷积神经网络模型结构,其中一个神经网络模型对大蒜是否被喂入进行实时监测,检测到大蒜喂入调整装置后,一个ResNet-18网络模型对蒜种鳞芽朝向进行判断,当鳞芽朝上时大蒜鳞芽调整机构打开Y型料斗使大蒜以鳞芽朝上的姿态直接落下,当鳞芽朝下时大蒜鳞芽调整机构翻转180°带动大蒜一起翻转后以鳞芽朝上的姿态落下,实现大蒜鳞芽朝向实时调整。神经网络模型推理及舵机控制采用英伟达边缘计算处理器Jetson Nano进行处理。利用离散元分析软件EDEM结合正交试验方法对调整装置的关键结构参数进行优化,并以杂交大蒜为试验对象进行台架试验,试验结果表明:大蒜鳞芽调整成功率为96.25%,模型推理时间0.045 s,平均每粒大蒜调整时间为0.785 s,满足大蒜播种机播种要求。该文研究结果可为解决杂交大蒜直立播种问题及边缘计算在精密播种设备中的应用提供有益参考。

Design and experiments of garlic bulbil orientation adjustment device using Jetson Nano processor

Garlic cultivation is highly demanding for a single seed to maintain upright-directional sowing with the roots down ward vertically. However, current adjustment devices for the direction of garlic cloves cannot be suitable for hybrid garlic varieties. In this study, an intelligent adjustment device was designed for the direction of garlic clove using edge computing. The device consisted of a feeding channel of garlic clove, a camera, a camera bracket, clove direction adjustment mechanism, turning servo and brackets. The adjustment mechanism of clove direction was composed of flip frame, reset spring, hopper opening and closing servo, Y-type hopper I and II. A dual convolution neural network (CNN) structure was adopted in the control system, where a custom deep learning CNN was for garlic monitoring in real time, and a ResNet-18 network was for the determination of garlic clove orientation. In monitoring, the garlic clove was distinguished from the background of images, thereby determining whether the clove was fed to the adjustment. Areal-time detection of orientation was to keep the pointy end of garlic clove facing upward, while the blunt end down into the soil. A suitable control strategy was provided to promptly adjust the direction of garlic clove. Higher identification accuracy and real-time performance were achieved in two different networks for separate detection and orientation of garlic clove. The specific procedure of orientation adjustment was as follows. An image processing was performed to determine whether the garlic clove entered the Y-shaped hopper from the feeding channel. Once the garlic clove was detected to be in the hopper, an image was real-time captured by the camera. The captured image was processed immediately through the deep learning network of detection and orientation. When the scales (blunt end)of garlic cloves were facing upward, the opening and closing servos of a hopper rotated at a certain angle to open the lower end of Y-type hopper I and II. As such, the garlic clove fell directly into the inserting with the scales facing upward. If the scales of garlic cloves were facing downward, the turning servos and adjusting mechanism individually rotated 180°, to accurately tailor the orientation of scales when the garlic clove was sliding down the guide slot of Y-type hopper I. Both theoretical and empirical data demonstrated that the structural parameters of the adjusting mechanism greatly dominated the success rate of the adjusted scale bud. A discrete element method (DEM) was performed on a commercial software EDEM to simulate the working effect of the adjusting mechanism. An orthogonal test was also utilized to optimize the key parameters of adjusting mechanism. An optimal combination of parameters was obtained, where the inclination angle of the garlic clove was 80°in the feeding channel, the radius of the Y-shaped hopper was 12.49 mm, and the height difference of the hopper was 20mm.Finallya bench test was carried out to verify the direction adjustment of garlic cloves. In scale bud, the success rate of identification was 97.25%, and the success rate of adjustment was 96.25%, where the success rate of adjustment was slightly lower than that of recognition. The reason was that the correctly identified garlic turned over unexpectedly when falling, due to the irregular center of gravity in a single seed. The mean inference time of the model was 0.045 s, indicating a small proportion of adjustment time for the scale bud. The average adjustment time was 0.785 s, where the mean value was 0.59 s when the garlic cloves were facing up and 0.98 s when facing down. There was a relatively large difference in the adjustment time when the garlic buds were faced up and down. This difference came into being because there was inconsistent movement stroke of the adjustment mechanism in two cases, particularly where the rotation speed of the drive motor was the same when the scale buds were facing down. Consequently, the adjustment time of scale buds facing up was shorter than that of the roots down ward vertically in garlic planting.