The Research On Real-time Monitoring Of Glass Curtain Wall Component Quality Based On Machine Vision

Glass curtain walls have been prevalently used worldwide as a type of building material.However,the uncertainty of external environments can often lead to their quality problems and further result in frequent safety incidents.At present,the quality inspection of curtain wall glass components is manually conducted in most industrial scenarios.The manual sampling and inspection methods are usually inefficient and unstable,no longer meeting the requirement of high-volume agile production.Hence,there has been an urgent need to explore a fully automated glass quality inspection method.For the problems of slow defect localization,inaccurate classification and poor real-time performance in curtain wall glass component quality inspection,this paper proposes a real-time inspection method of curtain wall glass component quality based on machine vision.The specific research work is as follows:(1)Based on the analysis of the existing problems in the quality inspection of curtain wall glass components,a glass quality inspection method based on scaling cross-stage partial network is proposed.We construct YOLOv5s-P model suite that can be scaled up and down for networks of different sizes,based on the scaling algorithm of cross-stage partial network and YOLOv5 s model;specially,the Neck network structure is CSP-ized to improve the feature extraction capability of the model Then,Soft Pool downsampling method is used to optimize network structure and parameters of the SPP module,and the depthwise separable convolution is introduced to make the model lightweight while avoiding accuracy loss.Compared with the original model,the detection accuracy is improved by 3%,and reduces the parameter amount by 1.7%.The detection speed is 52 FPS when deployed on GPU.(2)Aiming at the adaptability of distributed detection in heterogeneous environment,we proposed q-Fed DANE,an adaptive federated optimization algorithm for heterogeneous networks.Based on the classical Fed DANE,the attenuation parameter q is introduced into the formula,attenuating negative effects of the gradient correction term and the near-end term so as to improve the model’s ability of perceiving heterogeneity.In addition,the new algorithm flow carryies out only one round of device communication for each update to reduce the communication cost and relevant overhead.Then,the random optimizer Adam is adopted on the server side,dynamically yielding an adaptive learning rate to optimize the objective with global information across edge device networks.This turns out to improve the convergence stability of the model(3)An experimental platform for quality inspection of curtain wall glass components and a Web visualization platform were established.The collected curtain wall glass images are inspected using the experimental platform and the theoretical methods in Chapters 3 and 4 are verified through experiments.In the glass inspection experiments,the experimental results of YOLOv5s-csp and different lightweight detectors in this paper are compared and analyzed,and the YOLOv5s-P series model is tested and high accuracy is obtained.In the federal optimization simulation experiments,the influence of q-value on the convergence stability and detection accuracy of the algorithm in different heterogeneous degree environments is analyzed,and the model performance in the case of low device involvement is given.