现在樱桃市场上存在着大量以次充好的不良现象, 严重损害了名牌樱桃的品牌经济效益, 所以亟需一种能对不同产地樱桃实现快速无损鉴别的技术。 拉曼光谱溯源技术作为光谱溯源技术的一种, 由于具有快速、 高效、 无污染、 无损分析等优点, 逐渐得到相关研究者的重视。 长短期记忆（LSTM）网络是一种具有记忆性的反馈神经网络, 它是循环神经网络的一种变体。 LSTM网络克服了循环神经网络中梯度消失的缺点, 适合处理序列敏感的问题和任务, 目前被广泛应用在语音识别、 图像识别和手写识别等领域, 但LSTM网络在产地溯源方面的应用还有待研究。 基于此, 提出了一种LSTM网络与拉曼光谱技术结合的能对不同产地樱桃实现快速无损鉴别的技术。 将来自美国、 山东和四川的369个樱桃作为研究样本, 用拉曼光谱仪在785 nm激光下获得了不同产地樱桃的光谱数据。 并且以每条经过基线校正后的拉曼光谱数据作为网络输入数据, 基于LSTM网络构建了能对不同产地樱桃实现快速鉴别的判别模型, 并且以样本判别准确率A、 样本精确率P、 样本召回率R和样本F值作为评价指标, 探究了不同预处理方法对LSTM网络判别模型性能的影响。 结果表明: 当样本训练集和测试集的比例为85∶38时, 直接采用原始拉曼光谱数据的LSTM网络模型的产地鉴别能力不高, 鉴别准确率为79.87%。 但当使用预处理过后的拉曼光谱数据, 模型的鉴别准确率维持在92%以上。 并且光谱经过SG+MSC预处理后模型的鉴别准确度最好, 鉴别准确率达99.12%。 同时在采用SG+MSC预处理的方法下, LSTM网络鉴别模型的精确率、 召回率、 F值均较高, 表明了所提出的LSTM网络模型有较好的性能可实现对不同产地樱桃的鉴别, 为樱桃的产地溯源提供了一种新的思路。

At present, there are a lot of unhealthy phenomena in the cherry market, which have seriously damaged the economic benefit of famous cherry brands. As a kind of spectrum tracing technology, Raman spectrum tracing technology has been paid more and more attention because of its advantages of fast-speed, high efficiency, pollution-free and non-destructive analysis. And the long short-term memory (LSTM) network is a kind of feedback neural network with memory, which is a variant of the recurrent neural network. LSTM network overcomes the problem of gradient disappearance in the recurrent neural network, and is suitable for solving sequence-sensitive problems and tasks. At present, it is widely used in speech recognition, image recognition and handwriting recognition. However, there are few studies on the application of LSTM network in origin tracing. Therefore, a technology that can identify cherries of different origins quickly and non-destructively is urgently needed. Based on this, this study in this paper proposes a fast and non-destructive identification technique for cherries from different origins by using LSTM network and Raman spectroscopy. In this study, 369 cherries from the United States, Shandong and Sichuan are used to obtain the spectral data of cherries from different regions with the Raman spectrometer under the 785 nm laser. Moreover, the Raman spectral data after baseline correction is taken as the network input data, and a discriminant model is built based on the LSTM network to realize rapid identification of cherries from different origins. In addition, the sample discrimination accuracy A, sample precision P, sample recall R, and sample F values are used as evaluation standards to explore the effects of different prepossessing methods on the sample discrimination accuracy. The results showed that when the ratio of the sample training set to the test set is 85∶38, the LSTM network model that directly uses the original Raman spectral data has poor ability to identify the origin, and the identification accuracy is only 79.87% on average. But when prepossessed Raman spectral data are used, the average accuracy of the model remains above 92%. And the model has the best discrimination accuracy after using Stravinsky-Golay (SG) and multiplicative scatter correction (MSC) prepossessing methods, and the discrimination accuracy reaches 99.12%. At the same time, the accuracy rate, recall rate and F value of LSTM network discrimination model are all high when the preprocessing method named SG+MSC is used. It means that the LSTM discrimination model proposed in this paper can perform well in distinguishing cherries from different regions, which provides a new way of tracing the origin of cherries.