少样本场景下的气液两相流动状态识别  

State identification of gas-liquid two-phase flow in few-shot scenario

作　　者：李凌涵 张淑美 董峰 LI Linghan;ZHANG Shumei;DONG Feng(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China;School of Information and Intelligent Engineering,Tianjin Renai College,Tianjin 301636,China)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072 [2]天津仁爱学院信息与智能工程学院,天津301636

出　　处：《化工进展》2025年第4期1794-1805,共12页Chemical Industry and Engineering Progress

基　　金：国家自然科学基金(62373277,51976137)。

摘　　要：气液两相流具有复杂多变的流动状态,包括典型状态和过渡状态。流动状态的准确识别与监测对深入理解两相流机理和保证工业过程的安全运行至关重要。然而,传感器部署成本的限制和实际流动过程数据采集的难度导致了气液两相流的数据欠缺。本文系统地讨论了气液两相流的少样本流动状态识别问题,采用电导传感器信息处理、多尺度域时频特性分析、原型网络小样本元学习框架。首先,利用响应快、安全、成本低的电导传感器进行测量,获取反映气液两相流动态特性和流动结构的含水率信息。然后,对电导传感器的响应信号进行时频分析,通过改进的经验小波变换法获得电导信号在多尺度域的波动信息,实现不同流动状态的联合表征。最后,将所提取特征嵌入原型网络中,在元学习框架下进行模型训练,解决少样本场景下的气液两相流动状态识别问题。利用6种典型流动状态和4种过渡状态进行了少样本状态识别实验,在目标流动状态只采用3个或5个样本进行训练的情况下,平均识别准确率超过80%,验证了方法的有效性。Gas-liquid two-phase flow has complex and changeable flow states,including typical states and transition states.The accurate identification and monitoring of flow state are crucial for understanding the mechanism of two-phase flow and ensuring the safe operation of industrial processes.However,the limitation of deployment cost of sensors and the difficulty of data acquisition in actual two-phase flow processes lead to the data deficiency in gas-liquid two-phase flow.In this paper,the problem of few-shot flow state identification of gas-liquid two-phase flow was systematically discussed,which was researched under the framework of information processing of conductance sensor,time-frequency analysis in multi-scale domain and few-shot learning by the prototypical network.Firstly,the water holdup information reflecting dynamic characteristics and flow structure of gas-liquid two-phase flow was obtained by the conductance method with fast response,safety and low cost.Then,the response signal of the conductance sensor was analyzed by time-frequency analysis.The fluctuation information of conductance signal in multi-scale domain was obtained by the improved empirical wavelet transform method,which characterized different flow states jointly.Finally,the extracted features were embedded into the prototypical network,and the model was trained under the meta-learning framework to solve the problem of gas-liquid two phase flow state identification in few-shot scenario.The few-shot state identification experiments were carried out through 6 typical flow states and 4 transition states.When only 3 or 5 samples were available for target flow states during training,the comprehensive recognition accuracy exceeded 80%,which verified the effectiveness of the method.

关 键 词：气液两相流 仪器仪表 神经网络 状态识别 少样本学习 原型网络 

分 类 号：TP212[自动化与计算机技术—检测技术与自动化装置] TQ021[自动化与计算机技术—控制科学与工程]