电站主蒸汽管道多尺度伤损的多模态导波定量评估方法研究

To meet the requirements of safe operation and health monitoring of the equipment in power plant, the project proposes a new method for quantitative evaluation of multi-scale defect in the main steam pipeline (MSD-in-MSP) based on high-order multi-modal guided wave (HOMMGW). This method aims to solve the problem that low-order single-mode guided wave (LOSMGW) is vulnerable to missed detection of micro defects. Through forward analysis, algorithm research and experimental verification, this study intends to explore the mode separation of HOMMGW under high frequency-thickness product (FTP), the interaction model between HOMMGW and damage as well as quantitative evaluation of defect. According to the advantage that matched transform kernel is able to obtain the time-frequency distribution with high energy concentration and no cross-term interference which is beneficial to identifying unknown multiple components, a new method of multiple scattering mode separation is investigated based on parameterized time-frequency analysis. According to the characteristics that the spectral analysis can reveal the energy distribution of each mode component, the scattering coefficients are solved, and the scattering matrix model based on spectral peak is constructed. According to the information mining ability of deep learning and the small sample advantage of transfer learning, the research focuses on how to optimize the visualization technology-based convolutional layer parameters, thus a new method for quantitative evaluation of MSD-in-MSP is proposed based on pre-trained deep convolutional neural networks. This project is expected to provide a new method for quantitative evaluation of multi-scale defect in the thick-walled pipeline so as to improve the online health monitoring system for large-scale in-service equipment.

面向电力设备安全运行及健康监测重大需求，针对当前低阶单模态导波主蒸汽管道检测微小伤损易出现漏检的问题，本项目提出基于高阶多模态导波的主蒸汽管道宏、微多尺度伤损定量评估新方法。通过正向解析、算法研究、实验验证，探索高频厚积下多模态导波的模态分离、与伤损交互作用模型、伤损定量评估等关键问题的解决途径。依据匹配的变换核能获得能量集中度高且无交叉项干扰的时频分布、利于未知多分量识别的优势，探讨基于参数化时频分析的多散射模态分离新方法。依据谱分析可表征各模态能量分布的特点，求解散射系数，构建基于谱峰值的多模态导波散射矩阵模型。依据深度学习的信息挖掘能力和迁移学习的小样本优势，重点研究如何结合可视化技术优化配置卷积层参数，提出基于预训练深度卷积神经网络的多尺度伤损定量评估新方法。项目预期形成一种厚壁管道宏、微多尺度伤损定量评估新方法，完善大型在役设备在线健康监测系统。

项目面向大型在役设备安全运行及健康监测重大需求，针对当前低阶单模态导波微小伤损易出现漏检的问题，开展基于高阶多模态导波的宏、微多尺度伤损定量评估新方法。通过正向解析、算法研究、实验验证，探索高频厚积下多模态导波的模态分离、与伤损交互作用模型、伤损定量评估等关键问题的解决途径。.主要开展研究工作如下：.1、多模态导波与大厚度板件正向研究.基于大厚度钢板与兰姆波交互的COMSOL模型，构建应力与模态的对应关系并提取缺陷信号；其次，基于模型和频散曲线开展了基于速度法的模态分离和识别方法。.2基于泛谐波调频小波变换(Generalized Warblet Transform, GWT )的导波信号分析研究。.GWT采用由傅里叶级数逼近的变换核函数，可以更好的逼近多模态信号。由GWT方法得到的时频表示不仅在瞬态分量分析方面具有出色的能力，而且受图像尺寸减小的影响较小。.3、开展了基于深度卷积网络的伤损定量分析方法研究。.首先，提出了一种面向二维DCNN架构的复合剪枝方法。提出的方法从减小输入大小、在每个卷积层和全连接层后添加批量归一化层、简化全连接层、并删除卷积层中不重要的过滤器四个方面来对经典的VGG16架构进行轻量化处理，并将该构架用于三种伤损识别。其次，自建了一种简易的DCNN架构用于伤损识别。提出的自建DCNN构架由2层卷积层、BN层、池化层和全连接层构成，其中卷积层分别包含64和128个滤波器，两层全连接层各包含512个隐含节点，该构架重要参数由粒子群(Particle Swarm Optimization，PSO)算法自适应优化配置；实验验证该架构可以达到上述剪枝DCNN构架的准确度，使得网络大为简化，降低时间和存储成本；。.已在国内外学术期刊/专业会议上发表高水平论文6篇（其中，SCI检索论文3篇），受理国家发明专利1项，申请国家发明专利2项。课题开展助力人才培养：项目执行期间；培养硕士研究生7名（含毕业生1名，在读6名），本科生科研训练多名；负责人牵头获批多项省级成果。