基于加速度和卷积神经网络的船体板裂纹损伤检测

Hull plate crack damage detection based on acceleration and convolutional neural network

裂纹是船舶结构中最常见的损伤形式之一。由于船舶结构十分复杂而且密闭空间多,传统的依靠人力的裂纹损伤检测方法耗时长、主观依赖性强,难以满足智能船舶的需求。本文提出了一个基于加速度和卷积神经网络(convolutional neural networks,CNN)的船体板裂纹损伤实时在线无损检测方法,该方法能够自动地学习裂纹损伤特征。通过基于Python语言的Abaqus二次开发技术建立简支板损伤模型并计算其动力学响应。采集板的加速度数据用于训练CNN模型,利用数据裁剪技术对数据集进行扩充,讨论了不同CNN结构形式对船体板裂纹损伤检测的影响。与基于小波包变换的多层感知机神经网络相比,提出的CNN方法能够更好地提取裂纹位置和长度损伤特征,同时对噪声的敏感程度较低。

Crack is one of the most common damage forms of ship structures.Because the ship structure is very complicated and there are many confined spaces,the traditional human-based crack damage detection method is time-consuming and subjectively dependent,which is difficult to meet the needs of intelligent ships.A hull plate crack damage real-time online non-destructive detection method was proposed based on acceleration and convolutional neural network(CNN),which can automatically learn crack damage character⁃istics.With the Abaqus scripting interface,simply-supported plate models with preset cracks were modeled.Node accelerations at representative locations were adopted as the CNN input parameters.The data cropping method was adopted for the augmentation of samples.The effect of CNN structures on the hull plate crack damage detection was discussed.Compared with the multilayer perceptron neural network based on wavelet packet transformation,the proposed CNN method can better extract the crack location and length damage characteristics and is less sensitive to noise.