基于BP神经网络的预制柱套筒灌浆缺陷检测研究

Research on Defect Detection of Sleeve Grouting in a Precast Column Based on BP Neural Network

装配式建筑节点的连接对其性能好坏起着决定性的作用,随着国家大力推广装配式建筑,节点连接处的缺陷检测变得越发重要,然而目前针对装配式混凝土结构的套筒灌浆缺陷检测相关研究较少。本文利用有限元软件ABAQUS建立了4种不同约束条件下采用套筒灌浆方式连接的预制柱的数值计算模型,通过设置不同缺陷工况,获得神经网络训练参数,从而训练BP神经网络,用于后续缺陷检测;研究了训练数据中是否含有噪声对神经网络的影响以及测试数据中含有噪声时对缺陷检测结果的影响。结果表明:相较于原始数据,由含噪数据训练得到的神经网络精度更高,性能更好,同时也能改善均方误差及各数据集的回归分析结果;对尚需检测的数据,应注意在测试前是否对数据进行滤波处理,以减小缺陷检测结果的误差。

The connection of prefabricated building joints is crucial to its performance.With the gradual promotion of precast RC structures in China,the defect detection of the joint connections become more and more important.However,there are few correlational researches on the detection of sleeve grouting defects in precast concrete structures.In this paper,the refined finite element models of four precast columns with rebars connected by sleeves at joint under different constraints were established by using software ABAQUS,and then various conditions with different grouting defects were simulated to gain the training parameters to train the BP neural network for defect detection.The influence of whether there is noise in the training data on the neural network and the influence of noise in the test data on the defect detection results were explored.The results show that(1)compared with the original data,the neural network trained from noisy data has higher accuracy and better performance and the mean square error of network training is decreased and results of regression analysis are improved;(2)for the data to be detected,attention should be paid to whether the data is filtered before the test to reduce the error of defect detection results.