特征提取是太赫兹光谱识别的关键处理步骤, 通常利用降维方法作为特征提取手段。 然而, 当一些化合物的太赫兹光谱曲线整体差异度较小时, 降维方法往往会缺失样本差异的重要特征信息, 从而导致分类错误。 如果不采用降维方法提取特征, 传统机器学习分类算法对维数较高的原始太赫兹光谱数据又不能很好的分类。 针对此问题, 提出了一种基于双向长短期记忆网络(BLSTM-RNN)自动提取太赫兹光谱特征的识别方法。 BLSTM-RNN作为一种特殊的循环神经网络, 利用其LSTM单元可以有效解决原始太赫兹光谱数据维数较高使得模型难以训练问题。 再结合模型的双向频谱信息利用架构模式, 可以增强模型对复杂光谱数据自动提取有效特征信息的能力。 采用三类、 15种化合物太赫兹透射光谱作为测试对象, 首先利用S-G滤波和三次样条插值对Anthraquinone, Benomyl和Carbazole等十五种化合物在0.9～6 THz内的太赫兹透射光谱数据进行归一化处理, 然后通过构建一个具有双向长短期记忆的循环神经网络对太赫兹光谱的全频谱信息进行自动特征提取并利用Softmax分类器进行分类。 通过试验优化网络结构和各项参数, 最终获得了针对复杂太赫兹透射光谱数据的预测模型, 并与传统机器学习算法SVM, KNN及神经网络算法MLP, CNN进行对比实验。 结果表明, dataset-1和dataset-2分别作为差异度较大和无明显峰值特征的五种化合物太赫兹透射光谱数据集, 其平均识别率分别为100%和98.51%, 与其他方法相比识别率有所提高; 最重要的是, dataset-3作为5种化合物谱线极为相似的太赫兹透射光谱数据集, 其平均识别率为96.56%, 与其他方法相比识别率提高显著; dataset-4作为dataset-1, dataset-2和dataset-3的透射光谱数据集集合, 其平均识别率为98.87%。 从而验证了BLSTM-RNN模型能自动提取有效的太赫兹光谱特征, 同时又能保证复杂太赫兹光谱的预测精度。 在选择模型训练优化算法方面, 使用Adam优化算法要好于RMSProp, SGD和AdaGrad, 其模型的目标函数损失值收敛速度最快。 同时随着模型训练迭代次数增加, 相似太赫兹透射光谱数据集的预测准确率也不断提升。 可为复杂太赫兹光谱数据库的光谱识别检索提供一种新的识别方法。

Feature extraction, the key process of the terahertz spectral recognition, typically uses the dimensionality reduction techniques. However, when the overall difference of terahertz spectra of some compounds is small, dimensionality reduction methods often lack important feature information of sample differences, which leads to classification errors. If the dimensionality reduction process is not performed, the traditional machine learning algorithm cannot be well classified because the original spectral data have a high dimensionality. Therefore, this paper proposes a terahertz recognition method based on bidirectional long short-term memory recurrent neural network (BLSTM-RNN), which performs automatic feature extraction with containing full spectrum information of terahertz spectrum. BLSTM-RNN is a special recurrent neural network, whose LSTM unit can be used effectively to solve the problem that the original terahertz spectral data dimension is high. Then, it becomes easier to train the model. What’s more, the architectural model combined with bi-directional spectral information can enhance the ability of the model to extract valid feature information from complex spectral data automatically. In this paper, three types and 15 compounds terahertz transmission spectra are used as test objects. The terahertz transmission spectrum samples data of 15 organic compounds such as Anthraquinone, Benomyl and Carbazole were firstly normalized in 0.9～6 THz by S-G filtering and cubic spline interpolation. Then a recurrent neural network with bidirectional Long short-term memory unit (LSTM) is constructed to automatically extract the full spectrum information of the terahertz spectrum and classify it by Softmax classifier. Through experimentation of optimizing the network structure and various parameters, the prediction model of the complex terahertz transmission spectrum data is obtained, and the comparative experiment is done by contrasting with the traditional machine learning algorithm SVM, KNN and neural network algorithm MLP, CNN. The results show that compared with other methods, the recognition accuracy of both dataset-1 and dataset-2 is improved. Dataset-1 and dataset-2 are two terahertz transmission spectral data sets of five compounds with large difference and no obvious peak characteristics, and the average recognition accuracy of the former is 100% and the latter 98.51%. Most importantly, dataset-3 is a dataset of terahertz transmission spectra with five similar spectral lines. The average recognition accuracy is 96.56%. Compared with other methods, the recognition accuracy is significantly improved. Dataset-4 as a collection of transmission spectral data sets for dataset-1, dataset-2, and dataset-3 has an average recognition accuracy of 98.87%. It is verified that the BLSTM-RNN model can automatically extract effective terahertz spectral characteristics and meanwhile ensure the prediction accuracy of complex terahertz spectra. In the selection of model training optimization algorithm, the Adam optimization algorithm is better than the RMSProp, SGD and AdaGrad optimization algorithms, and the target function loss value of the model has the fastest convergence rate. At the same time, as the number of training iterations increases, the prediction accuracy of similar terahertz transmission spectral datasets also increases. The proposed method can provide a new identification method for spectral recognition retrieval of complex terahertz spectral databases.