Research On Defect Detection Of Rake Blade Of Agricultural Machinery Based On Deep Learning

Since this century,due to the needs of agricultural development,the agricultural machinery industry has achieved rapid development.In the development process of agricultural machinery industry,the development of agricultural machinery parts industry is an important driving force,and the quality of parts affects the performance of the entire agricultural machinery.Circular disk harrow blades are a key component of agricultural machinery that is consumed in large quantities.During the production process of the harrow blades,various defects may appear on the surface,which can affect the service life of the harrow blades and the performance of agricultural machinery.During the production process of the harrow blades,it is necessary to conduct defect testing.Currently,manufacturers still rely on manual testing of the surface defects of the rake blade,which is not only inefficient,but also costly.With the continuous development of deep learning technology,object detection algorithms based on deep learning have been widely studied and applied in the field of defect detection.Starting from designing a method that is simple to implement,low in cost,and high in accuracy to replace manual detection and achieve automatic detection of surface defects on disk harrows,this thesis conducts research on two mainstream target detection algorithms based on deep learning.Two algorithms,Fast R-CNN and YOLOv5,are applied to the defect detection task of agricultural machinery disk harrows.The main work of this article is as follows:(1)Create a disc rake defect dataset.Acquire images of disk harrows with defects at the harrow production site,then label them,create the required dataset,and expand the dataset through image enhancement.(2)Research on Flaw Detection of Agricultural Machinery Rake Based on Improved Faster R-CNN Algorithm.The Faster R-CNN algorithm was studied and applied to the defect detection task of agricultural machinery disk harrows.Aiming at the shortcomings and characteristics of the algorithm itself,two basic improvement measures,k-means and ROI Align,were introduced into the Faster R-CNN algorithm,and four improvement strategies were introduced into the basic Faster R-CNN algorithm: for the problem of many small size defects on the surface of disk harrows and large size differences,Using Res Net50 as a feature extraction network and using FPN to fuse feature maps for multiscale prediction;Add a SENet attention mechanism to the model to enhance important feature channel information while suppressing unimportant feature channel information;Use better CIo U NMS to replace traditional NMS for algorithm post processing and better screen out redundant prediction frames;SIo U Loss is introduced as the boundary box regression loss function during model training.Experimental results show that these improved strategies can effectively improve the accuracy of the Faster RCNN algorithm on disk rake data sets.The improved algorithm has a m AP@0.5 value of82.67% and an FPS value of 8.0.(3)Research on Flaw Detection of Agricultural Machinery Rake Based on Improved YOLOv5 Algorithm.Aiming at the complex model and slow detection speed of twostage target detection algorithm,the YOLOv5 algorithm was applied to the defect detection task of agricultural machinery disc rake.Three improvement strategies were introduced based on YOLOv5s: adding a CBAM attention mechanism after each C3 module in the backbone network;Using CIo U NMS to replace traditional NMS for algorithm post processing;In order to solve the problem of unbalanced high and low quality samples in the training process of the model’s bounding box regression,this thesis uses Focal EIo U Loss for reference.By integrating SIo U Loss and Focal L1 Loss,we obtain Focal SIo U Loss,which is used as the bounding box regression loss function during model training.Experimental results show that these three improved strategies can effectively improve the accuracy of the YOLOv5 s algorithm on the disc rake dataset.The improved algorithm has a m AP@0.5 value of 74.87% and a FPS value of 58.8%.Although the improved Faster R-CNN algorithm has relatively high accuracy in disc rake data sets,its parameter quantity and detection speed are significantly inferior to the improved YOLOv5 s algorithm.Finally,the architecture of the defect detection system was analyzed,and a visual interface was developed for the future practical application of the algorithm in this thesis.