Research On Intelligent Device For Urban Domestic Waste Sorting

According to national statistics,two-thirds of China’s large and medium-sized cities have already been surrounded by garbage."Garbage siege" has caused serious pollution to land,water and air,threatening people’s health.Therefore,it is necessary to take effective measures to comprehensively control urban garbage.Garbage sorting treatment is an effective way to achieve harmless,reduction and resource utilization of garbage.Taking into account the weak awareness of the current national garbage classification and the unclear classification instructions of public facilities,it is very necessary to design an intelligent garbage sorting system applied in public places,which can automatically complete two categories of garbage collection and other garbage sorting,and at the same time help to establish national garbage classification awareness through classification tips.In this thesis,three aspects which include garbage-classification-dataset establishment,classification-algorithm design and the implementation of sorting system are studied.The main work is summarized as follows:Firstly,the dataset of garbage classification is established.Through the survey,it is found that the public dataset of garbage classification research is very few,and the scale is very small.The largest garbage classification public dataset is the Trash Net,the total number of pictures is only 2527,and the dataset size is too small,which is not conducive to deep learning algorithm model training.In addition,the samples collected from existing datasets do not fully take into account the morphological diversity of domestic garbage due to extrusion deformation.For this reason,this paper designs a special data collection box and software,collects 10 624 samples in the garbage collection station,and establishes garbage classification dataset called Garbage Net.This paper designs a classification and recognition algorithm for garbage images.Secondly,a classification algorithm for Garbage Net is proposed.By comparing the performance of typical convolutional neural network models(Inception,Res Net,Dense Net,Nas Net)on Garbage Net,Dense Net161 is selected as baseline,with an accuracy of 95.5%.A series of improvements have been made to the network: adding Random Crop region restriction,a hybrid training method of recovering cross-entropy loss and category cross-entropy loss is proposed;on the data augmentation mode of the network,Random Erasing and Mixup are tried.After optimization,the accuracy is increased to 97.0%.Thirdly,software and hardware of sorting system are constructed.Microcomputer Raspberry Pi as a control brain can detect or control the status of magnetic proximity switch,relay module,electromagnetic lock,steering gear,ultrasonic module,power module and other modules.In addition to the module selection,the schematic design and Printed Circuit Board(PCB)design are completed based on Cadence,and the welding tests are made.In the transmission of Raspberry Pi and server,SFTP with high security and speed is selected by comparing the speed of various file transfer protocols.Finally,we build a prototype system and test each module,which basically realizes the automatic sorting function.In the future,the appearance,mechanical structure design and electrical design of the system will be further optimized.