作者：林旭梅, 胡川, 朱广辉, 陈一戈, 苗芳荣
作者单位：青岛理工大学信息与控制工程学院，山东 青岛 266520
关键词：钢筋混凝土;腐蚀检测;多传感器检测;残差神经网络;adam算法;学习率衰减
摘要：
针对钢筋混凝土腐蚀检测中单一传感器或检测方式获取锈蚀特征信息不足及准确率不高等问题，提出一种改进神经网络模型结构下对集成阳极梯、应变力、温度传感器的多传感器数据融合检测方式。首先将一维多传感器数据二维化，采用卷积核对特征信息滤波提取，提取后的信息平展后连接bp残差神经网络层，增强浅层低非线性度特征信息向深层网络的直接传递和重复利用，提高网络模型的拟合及泛化能力。针对adam优化算法在模型训练后期学习率可能震荡不收敛问题，引入分段学习率衰减策略抑制后期震荡，同时对二阶矩估计梯度变化进行调整，提高迭代收敛效率。仿真结果表明，改进后的adam-cnn算法模型具有更好的分类性能，在钢筋腐蚀样本测试集上的平均准确率为96.2%。
an improved adam-cnn model for corrosion detection of reinforced concrete
lin xumei, hu chuan, zhu guanghui, chen yige, miao fangrong
college of information and control engineering, qingdao university of technology, qingdao 266520, china
abstract: in order to solve the problems of insufficient information and low accuracy of single sensor or detection method in corrosion detection of reinforced concrete, a multi-sensor data fusion detection method based on improved neural network model structure is proposed, which integrates anode ladder, strain force and temperature sensors. firstly, the one-dimensional multi-sensor data is transformed into two dimensions, and the feature information is extracted by convolution check filter, the extracted information is flattened and connected with bp residual neural network layer to enhance the direct transmission and reuse of shallow low nonlinear feature information to deep network, and improve the fitting and generalization ability of network model. aiming at the problem that the learning rate of adam algorithm may fluctuate and not converge in the later stage of training, the piecewise learning rate attenuation strategy is introduced to suppress the later stage of oscillation, and the gradient change of the second moment estimation is adjusted to improve the iterative convergence efficiency. the simulation results show that the improved adam-cnn algorithm model has better classification performance, with an average accuracy of 96.2% on the test set of reinforcement corrosion samples.
keywords: reinforced concrete;corrosion detection;multi-sensor detection;residual neural network;adam algorithm;decay of learning rate
2023, 49(2):8-14 收稿日期: 2021-06-03;收到修改稿日期: 2021-08-31
基金项目: 国家重点基础研究发展计划“973”项目（2015cb655100）
作者简介: 林旭梅(1971-),女,安徽桐城市人,教授,博士,研究方向为智能检测技术
参考文献
[1] 侯保荣, 张盾, 王鹏. 海洋腐蚀防护的现状与未来[j]. 中国科学院院刊, 2016, 31(12): 1326-1331
[2] 艾志勇, 舒小平, 荣耀, 等. 混凝土钢筋锈蚀防治技术研究与应用进展评述[j]. 材料保护, 2020, 53(9): 107-113 132
[3] zeng d, hao b h, zeng q h. the study for non-destructive quantification method of reinforcement corrosion degree based on electrochemical detection and finite analysis technology[j]. applied mechanics & materials, 2014, 527: 31-36
[4] 王潇舷, 金祖权, 姜玉丹, 等. 基于dic与应变测试的混凝土中钢筋锈胀应力分析[j]. 材料导报, 2019, 33(16): 2690-2696
[5] 申家玮, 彭建平, voelker c, 等. 多传感器钢筋锈蚀无损检测数据的机器学习[j]. 无损检测, 2019, 41(11): 59-64
[6] gotoh y, hirano h, nakano m, et al. electromagnetic nondestructive testing of rust region in steel[j]. ieee transactions on magnetics, 2005, 41(10): 3616-3618
[7] 彭伟康, 陈爱军, 吴东明, 等. 基于改进faster r-cnn的水准泡缺陷检测方法[j]. 中国测试, 2021, 47(7): 6-12
[8] 常亮, 邓小明, 周明全, 等. 图像理解中的卷积神经网络[j]. 自动化学报, 2016, 42(9): 1300-1312
[9] he k, zhang x, ren s, et al. identity mappings in deep residual networks [c]. european conference on computer vision , 2016, 9908: 630-645.
[10] 王晓东, 宁静, 陈春俊. 1d cnn和lstm高速列车横向稳定性状态识别研究[j]. 中国测试, 2020, 46(11): 25-30
[11] kingma d, ba j. adam: a method for stochastic optimization[j]. computer science, 2014: 6980
[12] reddi s j, kale s, kumar s. on the convergence of adam and beyond[c]// iclr18, 2018.
[13] 贺昱曜, 李宝奇. 一种组合型的深度学习模型学习率策略[j]. 自动化学报, 2016, 42(6): 953-958
[14] 蒋文斌, 彭晶, 叶阁焰. 深度学习自适应学习率算法研究[j]. 华中科技大学学报(自然科学版), 2019, 47(5): 79-83
[15] yang y w, hao x j, zhang l l, et al. application of scikit and keras libraries for the classification of iron ore data acquired by laser-induced breakdown spectroscopy (libs)[j]. sensors, 2020: 32143315
