基于MSCNN-LSTM的注意力机制U型管道缺陷识别模型  

Siphon defect recognition model based on the MSCNN-LSTM and attention mechanism

作　　者：朱雪峰 冯早 马军[2] 范玉刚[2] ZHU Xuefeng;FENG Zao;MA Jun;FAN Yugang(Faculty of Civil Aviation and Aeronautics,Kunming University of Science and Technology,Kunming 650500,China;Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学民航与航空学院,昆明650500 [2]昆明理工大学信息工程与自动化学院,昆明650500

出　　处：《振动与冲击》2023年第22期293-302,共10页Journal of Vibration and Shock

基　　金：国家自然科学基金(62173168)。

摘　　要：对于承担缓震功能的特异U型管道,其结构复杂使管内和管壁缺陷具有时延性和多源多征兆等特点。针对U型管道缺陷难以有效识别的问题,提出一种基于多尺度卷积神经网络–长短期记忆(multi-scale convolution neural network-long short-term memory,MSCNN-LSTM)的注意力机制U型管道缺陷识别方法。采用主动声学检测方法获取管道声学响应信号,将原始声学信号作为模型输入,训练多尺度卷积神经网络提取重要细粒度局部特征。然后,多尺度局部特征融合为一个特征向量输入至LSTM网络中抽取潜藏在时序规律的粗粒度上下文特征。下一步引入注意力机制,对提取的特征赋予不同的权重,使模型更关注于最具类别区分度的特征,滤除冗余特征,提高模型缺陷识别能力。最后,在输出端通过Softmax分类器实现U型管道缺陷识别。试验结果表明,与其他常用的分类方法相比,该方法拥有更快的收敛速度,可实现98.44%的缺陷识别准确率。此外,采用Grad-CAM类激活可视化方法对所提模型的特征学习和缺陷分类机理实现了过程分析和展示。In order to avoid physical damage to buried pipelines caused by resonance,most large buildings above ground are designed with sinking siphon.To solve the multi-source and multi-symptom defect recognition problem caused by the siphon's complex structure,the attention mechanism defect recognition method based on the multi-scale convolution neural network(MSCNN) and long short-term memory(LSTM) neural networks was proposed.The pipeline response acoustic signal was acquired through active acoustic detection method.With the original signal as a model input,the MSCNN was trained to extract important fine-grained local features.Then,multi-scale local features were fused into a feature vector which was input into the LSTM to extract coarse-grained contextual features underlying time sequence regularity.By introducing the attention mechanism and assigning different weight to each extracted feature,the features with the highest category discrimination were paid more attention,and the redundant features were filtered out,thereby improving the defect recognition ability.Finally,siphon's defect recognition was achieved through the Softmax classifier at the output.According to the experimental results,the proposed method has faster convergence speed and high defect recognition accuracy rate up to 98.44%.Moreover,a Grad-CAM class activation visualization method was adopted for the process analysis and the demonstration of feature learning and defect classification mechanism of the proposed model.

关 键 词：U型管道 缺陷识别 多尺度卷积神经网络(MSCNN) 长短期记忆(LSTM) 注意力机制 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置] TU992[自动化与计算机技术—控制科学与工程]