基于生成对抗网络的棉种裂纹损伤超声波检测方法研究

In the process of delinting cottonseeds, sulfuric acid penetrated into the cottonseeds through the cottonseeds cracks. This would cause damage to the cottonseeds and reduce the germination rate of the cottonseeds. This project intended to use air-coupled ultrasonic technology to detect cracked cottonseeds. A Generative Adversarial Networks(GAN) was designed to generate a large number of data samples of cottonseed's ultrasonic signals. The model classification accuracy could be improved by a large number of data samples generated by GAN. Because of the non-stationary and non-linear characteristics of air-coupled ultrasound signals, the method of noise filtering and feature extraction of ultrasonic signal was investigated. The feature fusion method was proposed by using deep Recursive Neural Networks(RNN) and Convolutional Neural Networks(CNN). A deep neural networks discriminant model was constructed and the deep neural network model parameters was optimized. A experimental platform of cottonseed crack detection based on air-coupled ultrasound was designed to achieve rapid detection of cottonseed cracks. The research results have important significance for exploring the methods of cottonseed ultrasonic signal data sample generation and elucidating the ultrasonic detection mechanism of cracked cottonseeds. At the same time, the research lays a theoretical foundation and method foundation for the development of high-throughput on-line detection equipment for cracked cottonseeds based on ultrasonic signal.

在棉种脱绒过程中，硫酸会通过种皮裂纹渗入棉种内部,造成对棉种的损害，进而降低棉种的发芽率。本项目拟采用空气耦合超声技术对裂纹棉种进行检测。研究利用生成对抗网络生成大量棉种超声信号数据样本的方法，通过生成的大量数据样本提高模型的分类精度。针对空气耦合超声信号非平稳、非线性的特点，探索针对超声信号的滤噪和特征提取方法。研究基于深度递归神经网络和卷积神经网络的特征融合方法。构建深度学习判别模型，优化深度神经网络模型参数。设计一套基于空气耦合超声的裂纹棉种检测实验平台，实现裂纹棉种的快速检测。研究成果对探索棉种超声信号数据样本的生成方法，阐明裂纹棉种超声检测机理具有重要意义，同时为开发基于超声信号的裂纹棉种高通量在线检测装备奠定了理论和方法的基础。

棉花种子加工过程中，经过轧花加工环节，极易造成棉花表皮破损，而破损表皮在泡沫硫酸脱绒环节存在硫酸进入棉种内部的可能，降低棉种播种后的发芽率。同时，在棉种播种后，水分会通过破损表皮进入棉种内部，进一步降低棉种的发芽率。在裂纹棉种检测方面，本项目搭建了基于空气耦合超声的裂纹棉种检测实验平台，针对裂纹棉种采用空气耦合超声技术进行检测，提出了一种裂纹棉种与正常棉种的判别方法，该方法通过变分模态分解方法（VMD），对原始空气耦合超声信号进行分解，从而获得超声信号的本征模函数序列，利用模函数序列构建超声信号特征矩阵，并进一步提出一种基于颜色的编码方法，将空气耦合超声信号的特征矩阵转换为彩色图像，从而实现从一维超声信号到二维图像的转换。将基于自注意力机制的Transformer模型与卷积神经网络进行特征融合，即可以从编码彩色图像中获得全局特征，同时又可以获得局部特征，实现全局特征与局部特征的融合，从而构建用于裂纹棉种检测的深度学习判别模型，实现裂纹棉种的快速检测。为实现对所有缺陷棉种的检测，对除裂纹棉种之外的破损棉种的检测也开展了研究。通过对不同种类农产品表面缺陷，构建深度学习网络模型，采用模型压缩，网络剪枝等技术实现了农产品的实时检测，为破损棉种的检测提供的技术储备。在上述基础研究的支持下，为开发缺陷棉种高通量在线检测装备提供了理论和方法的支撑。