【摘要】 玉露香梨因其果肉酥脆、汁多香甜而深受消费者喜爱，但在运输过程中易发生机械损伤，若未及时检测，将导致果实内部腐烂，进而影响整批果实新鲜度并造成经济损失。鉴于传统肉眼检测方法难以快速、精准识别早期损伤，提出一种基于表面增强拉曼光谱（SERS）结合深度学习模型的检测方法，探明玉露香梨在机械损伤早期发生阶段中的拉曼光学特性变化规律，通过深度学习算法挖掘早期损伤阶段的光谱特征，进行玉露香梨早期机械损伤检测。具体研究内容（1）构筑高灵敏SERS银溶胶纳米基底结合拉曼光谱仪获取不同损伤阶段的玉露香梨表面SERS光谱数据；（2）采用S-G平滑和迭代自适应加权惩罚最小二乘法进行原始光谱的预处理，消除荧光噪声和基线漂移；（3）利用数据增强技术扩展训练数据，快速傅里叶变换提取特征并构建一维卷积神经网络（1D-CNN）模型进行玉露香梨机械损伤的早期检测。结果表明，模型在准确率、精度、召回率和F1分数等多个指标均达到100%,同时梨果损伤部位蛋白质的拉曼特征峰由1 607 cm^-1向1 589 cm^-1发生了偏移。研究表明SERS结合深度学习在玉露香梨机械损伤早期检测中较强的判别能力，为果品损伤早期检测提供了一种新的研究思路，同时，为开发高灵敏性的水果品质检测传感器提供数据支撑。更多还原

【Abstract】 Yuluxiang pear is loved by consumers because of its crisp flesh and sweet juice, but it is prone to mechanical damage during transportation. Once it does not occur in time, it will lead to internal decay of the fruit, which will affect the whole batch of fruits and cause economic losses. It is difficult to detect early damage quickly and accurately depending on the human eye. Modern optical technology has been widely used in fruit tree quality inspection because of its non-contact and rapid advantages. Raman spectroscopy has emerged in fruit and vegetable quality detection due to its fast molecular fingerprint characteristics and non-sensitivity to water. To solve the problem of Yuluxiang pear being vulnerable to mechanical damage during picking and transportation, this study proposes a method based on surface-enhanced Raman scattering（SERS） technology combined with a deep learning method. The change law of Raman optical characteristics of Yuluxiang pear in the early stage of mechanical damage was explored, and the coupling relationship between the early damage stage and the spectrum was mined through the deep learning algorithm to detect the early mechanical damage of Yuluxiang pear. Specific research contents:（1） The SERS spectrum data of Yuluxiang pear surface at different damage stages were obtained by constructing a highly sensitive SERS silver sol nano substrate combined with Raman spectrometer;（2） The S-G smoothing algorithm and the iterative adaptive weighted penalized least square method are used to preprocess the original spectrum to eliminate the fluorescence noise and baseline drift.（3） Data enhancement technology was used to expand the training data, a fast Fourier transform was used to extract features, and a one-dimensional Convolutional neural network（1D-CNN） model was constructed to detect the mechanical damage of Yuluxiang pear. The results showed that the model achieved ideal accuracy, precision, recall, and F1 score, especially when the injury was only 4 hours old. At the same time, the Raman characteristic peak of the protein in the injured part of pome fruit shifted from 1 607 to 1 589 cm^-1. The study shows that SERS combined with deep learning has a strong discrimination ability in detecting mechanical damage of Yuluxiang pear early. This study provides a new research idea for the early detection of fruit damage and data support for developing high-sensitivity fruit quality detection sensors.更多还原