基于1D-DenseRNet的罐头真空度数据分类检测方法研究  

Research on 1D-DenseRNet-based classification and detection method for canned food vacuum data

作　　者：俞烁辰 张俊[2] 宋新杰 周锦云[2] YU Shuochen;ZHANG Jun;SONG Xinjie;ZHOU Jinyun(College of Biological and Chemical Engineering,Zhejiang University of Science and Technology,Hangzhou 310023,China;Key Laboratory of Fruit Postharvest Processing,Ministry of Agriculture and Rural Affairs,Institute of Food Science,Zhejiang Academy of Agricultural Sciences,Hangzhou 310021,China;Key Laboratory of Chemistry and Bioprocessing Technology for Agricultural Products in Zhejiang Province,Hangzhou 310023,China)

机构地区：[1]浙江科技大学生物与化学工程学院,浙江杭州310023 [2]浙江省农业科学院食品科学研究所,农业农村部果品采后处理重点实验室,浙江杭州310021 [3]浙江省农产品化学与生物加工技术重点实验室,浙江杭州310023

出　　处：《浙江农业学报》2024年第12期2846-2856,共11页Acta Agriculturae Zhejiangensis

基　　金：国家现代农业产业技术体系(CARS-26-04BY)。

摘　　要：针对食品罐头真空度检测精度低、成本高的问题,为实现食品罐头真空度的无损检测,提高检测速度与精确度,提出利用激光位移传感器测量罐头顶部曲线数据,通过1D-DenseRNet模型对曲线数据进行分类的研究。该模型包括多个改进的密集连接模块、门循环单元、注意力层和残差连接模块,提取时序电压序列特征和捕捉长期依赖关系。采用交叉验证的方法,分析了不同网络层对模型性能的影响。通过改变卷积层、结合注意力机制、不同的循环神经网络模块和残差网络结构,观察模型的准确率、模型大小等其他评价指标的变化,设计出最优的网络模型结构。实验数据表明,动量因子设定为0.9,学习率设定为0.0005时,结合了门循环单元的1D-DenseRNet模型小样本数据集上达到了最高的准确率(98.77%),且模型参数量也相对较小。对比单一的卷积神经网络及其他混合网络,展现了1D-DenseRNet模型处理食品罐头真空度检测任务的优势。Aiming at the problem of low accuracy and high cost of vacuum detection in food cans,in order to realize the nondestructive detection of vacuum in food cans and improve the detection speed and accuracy,the research of measuring the curve data on the top of cans by using laser displacement sensors and categorizing the curve data by 1D-DenseRNet model is proposed.The model includes several improved dense connectivity modules,gate cycle units,attention layers and residual connectivity modules to extract time-series voltage sequence features and capture long-term dependencies.A cross-validation approach was used to analyze the effect of different network layers on the model performance.The optimal network model structure is designed by changing the convolutional layer,combining the attention mechanism,different recurrent neural network modules and the residual network structure,and observing the changes in other evaluation indexes such as model accuracy and model size.The experimental data show that the 1D-DenseRNet model incorporating the gate recurrent unit achieves the highest accuracy(98.77%)on the small sample dataset when the momentum factor is set to 0.9 and the learning rate is set to 0.0005,and the number of model parameters is also relatively small.Comparison with a single convolutional neural network and other hybrid networks demonstrates the advantages of the 1D-DenseRNet model in handling the task of vacuum detection in food cans.

关 键 词：罐头无损检测 神经网络 真空度检测 小样本数据分类 

分 类 号：TS205.6[轻工技术与工程—食品科学]