基于可见/近红外光谱与卷积神经网络的苹果内部品质在线分选系统研究

Apple Internal Quality Online Separation System Based on Visible / Near-infrared Spectroscopy

摘要:

针对苹果的现场非破坏快检分选及品质检测需求,基于可见/近红外光谱检测分析技术,设计了链传动式苹果在线检测分选系统。该系统由动力机构、光谱采集单元、PLC控制模块及分级执行机构等硬件组成。通过传感器协同定位触发光谱采集,搭载检测模型可以实现对苹果内部品质(SSC和霉心病)的精准检测和判别。以富士苹果为对象,建立了霉心病判别和健康果SSC检测的双响应卷积神经网络(Convolutional neural network, CNN)模型,测试集的霉心病判别准确率为100% 、健康果SSC预测决定系数R2达到0. 97。最后将模型植入软件系统后对未参与训练的50个苹果进行测试,霉心病分选准确率为98% ,健康果SSC验证决定系数R2为0. 96,最佳分选速度0. 47 m/ s。结果表明,本文自主研发的链传动式苹果在线检测分选系统,建模精度高,稳定性强,具备快速实现苹果霉心病判别以及健康果SSC检测功能。

Abstract:

Based on the visible / near-infrared spectroscopic analysis technology, a chain-driven online apple inspection and sorting system was designed to meet the needs of non-destructive on-site quick inspection and quality testing of apples. The system consisted of hardware such as power mechanism, spectral acquisition unit, PLC control module and grading actuator. The sensor co-location triggered spectral acquisition, and the prediction model enabled rapid, non-destructive online discrimination of apple mouldy core disease and accurate prediction of healthy fruit SSC. Taking Fuji apples as the object, a dual-response convolutional neural network (CNN) model was established to discriminate mouldy heart disease and predict the SSC of healthy fruits, and the accuracy of the model prediction set for discriminating mouldy heart disease was 100% , and the fit of the SSC prediction of healthy fruits was 97% . Finally, the model was tested by applying apple random validation data after implanting the model into the software system, with 98% accuracy for mouldy core sorting, 96% fit for healthy fruit SSC prediction, and an optimal detection and sorting speed of 0. 47 m/ s. In addition, the decrease in discrimination accuracy was due to the intersection of the internal characteristics of healthy and slightly mouldy kernels, which was what we would focus on in the subsequent study. The results showed that the chain-driven online apple inspection and sorting system developed independently had high modelling accuracy and stability, and had the functions of quickly realizing the discrimination of apple mouldy heart disease and the non-destructive detection of healthy fruit SSC.