目的：针对太赫兹技术在中药材定性和定量分析中存在的实验数据难以规范、中药材特征吸收峰不明显以及模式识别方法无法有效提取光谱特征的问题，提出一种基于多尺度卷积网络的中药材质量鉴定方法。方法：利用川白芷太赫兹时域光谱数据，先经快速傅里叶变换获得频域吸收系数并生成二维图像；再输入InceptionTime网络提取时域波形多尺度特征、ConvNeXt 提取频域图像纹理特征；两分支末端分别输出固定维度的特征向量，经通道级联（concatenate）输入全连接融合网络，再经 Softmax 得到最终质量类别概率，构成QID-MCN模型，实现中药材综合分析。结果：QID-MCN模型对川白芷的鉴定准确率达到98.6%，显著提高了药材质量鉴定的准确性和速度。结论：QID-MCN模型通过并行提取和融合时域与频域特征，有效解决了太赫兹光谱特征提取困难的问题，为中药材质量分析提供了一种高效、精准的新方法。

Objective: To address issues in terahertz technology for qualitative and quantitative analysis of medicinal herbs, such as difficulties in standardizing experimental data, indistinct characteristic absorption peaks of medicinal herbs, and the inability of pattern recognition methods to effectively extract spectral features, a quality identification method for medicinal herbs based on a multi-scale convolutional network is proposed. Method: Using the terahertz time-domain spectral data of Angelica dahurica, the frequency-domain absorption coefficients are first obtained through fast Fourier transform (FFT) and converted into two-dimensional images. The InceptionTime network is then employed to extract multi-scale features from the time-domain waveforms, while the ConvNeXt network extracts texture features from the frequency-domain images. The outputs from the two branches are fixed-length feature vectors, which are concatenated at the channel level and fed into a fully connected fusion network. Finally, the Softmax function is applied to generate the probability distribution of quality categories, forming the QID-MCN model for comprehensive analysis of medicinal herbs. Results: The QID-MCN model achieved an identification accuracy of 98.6% for Angelica dahurica, significantly improving the accuracy and speed of medicinal herb quality identification. Conclusion: The QID-MCN model effectively solves the challenge of extracting terahertz spectral features by parallel extraction and integration of time-domain and frequency-domain features, providing an efficient and precise method for medicinal herb quality analysis.

四川省重点研发计划；国家自然科学基金