基于多任务学习的有源内腔混合气体反演算法研究(特邀)  

Active Intracavity Mixed Gas Inversion Algorithm Based on Multi-Task Learning(Invited)

作　　者：刘琨[1,2,3] 尹慧 江俊峰[1,2,3] 刘铁根[1,2,3] 赵成伟 Liu Kun;Yin Hui;Jiang Junfeng;Liu Tiegen;Zhao Chengwei(School of Precision Instruments and Opto-Electronics Engineering,Tianjin University,Tianjin 300072,China;Key Laboratory of Opto-Electronics Information Technology,Ministry of Education,Tianjin University,Tianjin 300072,China;Institute of Optical Fiber Sensing of Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学精密仪器与光电子工程学院,天津300072 [2]天津大学光电信息技术教育部重点实验室,天津300072 [3]天津大学光纤传感研究所,天津300072

出　　处：《激光与光电子学进展》2024年第3期32-41,共10页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61922061,61735011,61775161);国家重大科学仪器设备开发专项(2013YQ030915);天津市自然科学基金杰出青年科学基金(19JCJQJC61400)。

摘　　要：针对目前深度学习在气体检测领域多聚焦于学习单个任务即气体定性分类或气体体积分数定量回归,忽略了相关任务间的信息关联性,降低了模型学习精度与效率等问题,提出了一种基于一维卷积神经网络和长短期记忆网络的多任务学习模型,即MTL-1DCNN-LSTM,并行实现了混合气体种类定性识别与体积分数定量回归。利用掺铥光纤,搭建了二级放大掺铥环腔光纤激光器,基于有源内腔吸收光谱法探测了CO_(2)和NH_(3)混合气体的吸收光谱数据。将实验数据放入多任务学习模型中训练,并进行超参数优化后,对测试集数据进行测试得到气体识别准确率为100%,NH_(3)体积分数预测决定系数为99.84%,CO_(2)体积分数预测决定系数为99.62%,优于单任务模型与传统的气体反演算法如反向传播神经网络和支持向量机。所提出的深度学习算法与有源内腔法相结合的方法,为吸收光谱型混合气体反演技术的进一步研究提供了新思路。Deep learning methods used in the field of gas detection mostly focus on learning a single task,such as the qualitative classification of gas or the quantitative regression of gas concentration.However,training a model in this way ignores the correlation of information between related tasks,reducing the accuracy and efficiency of training.This paper proposes a multi-task learning(MTL)model that combines a one-dimensional convolutional neural network(1DCNN)and a long short-term memory(LSTM)network to realize qualitative identification of mixed gas species in parallel with a quantitative regression prediction of gas concentrations.Using a thulium-doped fiber,a two-stage amplified thulium-doped ring-cavity fiber laser was constructed,and the absorption spectral data of mixed gases,comprising CO_(2) and NH_(3),were detected based on the active intracavity absorption spectroscopy method.The experimental data were put into the MTL model to train until the model performance was optimized.The trained model achieves a gas classification accuracy rate of 100%,while the coefficient of determination of NH_(3) and CO_(2) are 99.86%and 99.62%,respectively.These values are superior to the equivalent values obtained using conventional single-task models and gas inversion algorithms such as the backpropagation neural network and support vector machine.By combining the deep learning algorithm with the active intracavity spectroscopy method,a superior absorption spectroscopy-based gas inversion technology is developed.

关 键 词：掺铥光纤激光器 有源内腔法 多任务学习 一维卷积神经网络 长短期记忆网络 

分 类 号：TN29[电子电信—物理电子学]