Research On Intelligent Identification Algorithm And Sorting Experiment Of Low-value Recyclable Waste

The difficulty of resource utilization of low-value recyclable waste is that the recycling cost is high,the profit is low,and the manual sorting efficiency is low;the purity of traditional mechanical sorting is low,and the procurement and maintenance costs of equipment are high.Based on this background,this paper proposes an intelligent identification and sorting method for low-value recyclable waste based on deep learning,and achieves efficient sorting of target materials.The main research work of this paper includes:(1)An online image acquisition platform based on coaxial light source and line scan camera is built,which can collect high-quality target images in real time.(2)The model of the robot was selected according to key parameters such as working load and motion stroke,and the structure of the suction cup and fixed bracket was designed;the vacuum generator was selected and calculated according to the working air pressure and load,and a solenoid valve was used to control the air road on and off.(3)The TCP/IP communication protocol is used to realize the communication between the industrial computer and the PLC;the incremental encoder is used to communicate with the PLC to calculate the real-time grasping position of the target;the Ether CAT bus communication is used to realize the precise synchronization control of the servo drive by the PLC.(4)The visual detection module is used to collect and semi-automatically label2500 high-quality datasets under the Paddle Paddle deep learning framework;then the best initial instance segmentation and target detection models are tested and selected;then the recognition accuracy of the Mask Rcnn instance segmentation model and The prediction efficiency of the ppyolov2 object detection model is optimized.(5)The online detection program and the robot control program were written and deployed,the overall development of the robot prototype was completed,and the sorting success rate and sorting efficiency of the prototype were optimized.After replacing the model backbone network,data enhancement method and normalization method,the comprehensive recognition accuracy of Mask Rcnn increased from 88.28% to 92.45%,and the average prediction time was 225ms;after calculating the optimal a priori box distribution and model compression training,the average prediction time of ppyolov2 reduced from 107 ms to 89 ms,and the comprehensive recognition accuracy was 87.79%.In the acceptance test,the average recovery rate of the prototype was 97.65%,and the average sorting efficiency was 67 times/min,reaching the predetermined target.Verified by actual working conditions,the intelligent sorting system for low-value recyclable waste proposed in this paper can meet the actual production needs and achieve accurate identification and efficient sorting of target materials.