Near infrared (NIR) spectrum analysis technology has outstanding advantages such as rapid, nondestructive, pollution-free, and is widely used in food, pharmaceutical, petrochemical, agricultural products production and testing industries. Convolutional neural network (CNN) is one of the most successful methods in big data analysis because of its powerful feature extraction and abstraction ability, and it is especially suitable for solving multi-classification problems. CNN-based transfer learning is a machine learning technique, which migrates parameters of trained model to the new one to improve the performance. The transfer learning strategy can speed up the learning efficiency of the model instead of learning from scratch. In view of the difficulty in acquisition of drug NIR spectral data and high labeling cost, this paper proposes three simple but very effective transfer learning methods for multi-manufacturer identification of drugs based on one-dimensional CNN. Compared with the original CNN, the transfer learning method can achieve better classification performance with fewer NIR spectral data, which greatly reduces the dependence on labeled NIR spectral data. At the same time, this paper also compares and discusses three different transfer learning methods, and selects the most suitable transfer learning model for drug NIR spectral data analysis. Compared with the current popular methods, such as SVM, BP, AE and ELM, the proposed method achieves higher classification accuracy and scalability in multi-variety and multi-manufacturer NIR spectrum classification experiments.

中文翻译：

---

基于近红外光谱和深度迁移学习的多厂商药物识别

近红外（NIR）光谱分析技术具有快速、无损、无污染等突出优势，广泛应用于食品、医药、石油化工、农产品生产和检测等行业。卷积神经网络（CNN）因其强大的特征提取和抽象能力而成为大数据分析中最成功的方法之一，特别适用于解决多分类问题。基于 CNN 的迁移学习是一种机器学习技术，它将训练模型的参数迁移到新模型以提高性能。迁移学习策略可以加快模型的学习效率，而不是从头开始学习。鉴于药物近红外光谱数据获取难度大，标注成本高，本文提出了三种简单但非常有效的迁移学习方法，用于基于一维 CNN 的药物多制造商识别。与原始的CNN相比，迁移学习方法可以用更少的近红外光谱数据实现更好的分类性能，大大降低了对标记近红外光谱数据的依赖。同时，本文还对三种不同的迁移学习方法进行了比较和讨论，选择了最适合药物近红外光谱数据分析的迁移学习模型。与当前流行的方法，如SVM、BP、AE和ELM相比，该方法在多品种和多制造商的近红外光谱分类实验中实现了更高的分类精度和可扩展性。