基于机器学习的ECT图像重建算法的研究

发布时间：2019-03-28 16:08

【摘要】：多相流作为自然界普遍存在的一种现象,不仅是由于被测介质的介电常数会随着温度等环境的变化而变化,而且还由于被测场域中存在其他介质,会使得测量时出现介质未知的情况,并且其流动特性十分复杂,难以用数学模型完全描述,因而给测量带来困难。多相流的在线检测成像技术和实施方案多处在实验室测试研究阶段,只有少数已商品化,并能普遍应用于在线检测中,且多适用于两相流,尚缺少一种适用于介质未知且相数自适应的图像重建算法,所以需要进一步的研究推动其发展和实用化。而学习是机器学习的一大特点,通过学习介质的不同变化,如分布变化、介电常数变化等,可以及时调整算法参数,这个特点适用于解决多相流在线检测问题。ECT技术是一种用于多相流在线检测技术,广泛存在于海上石油开采等工业化领域研究中。然而,ECT技术在应用上存在不少问题和难点,尚且还不完善。本文从微弱电容处理和介电常数未知情况下的图像重建算法研究的角度出发,基于机器学习的方法研究ECT的图像重建,主要工作和贡献如下:(1)针对ECT的“软场”特征,研究了数据预处理部分的ECT电容归一化模型。分析了电容归一化的物理特性并结合并联归一化方法,建立了一种加权值的电容归一化模型,并应用于基于SVM的图像重建中。通过与并联模型对比得出,在相数确定以及管内介质无变化的条件下,该方法不仅适用于两相流还适用于两相流以上的多相流,其图像重建与真实模型的相关性要高于并联归一化方法。(2)在ECT系统的实际应用中,两相流和多相流是最普遍的流体情况。不仅是由于被测介质的介电常数会随着温度等环境的变化而变化,而且还由于被测场域中存在其他介质,会使得测量时出现介质未知的情况。本文利用机器学习中的支持向量机方法具有良好泛化性的特点,提出采用基于SVC的电容层析成像图像重建算法对未知介电常数对象进行图像重建,仿真结果表明,在相数确定的情况,该算法能有效适应介质多样性变化,即对于不同介质,该算法都能有较高的图像重建精度。(3)现存的ECT重建算法往往只对相数确定以及介质无变化的情况进行重新构建。针对该问题,作者建立了基于SVM决策树的机器学习方法进行自适应相数预测模型,通过SVM决策树的方法实现在相数不确定的情况下,对管内介质进行预测,实验结果表明,在相数不确定的情况下,该方法能较好的区分管内相数以及管内包含的介质。最后,结合上述方法,设计了基于SVM决策树的自适应相数ECT图像重建算法,初步分析了相数未知以及介质变化时,如何进行ECT图像重建,并达到提高重建精度的目的,为ECT技术提供了一种新的研究思路。
[Abstract]:As a common phenomenon in nature, multiphase flow is not only due to the change of dielectric constant in the measured medium with the change of temperature and other environments, but also due to the existence of other media in the measured field. It will make the medium unknown in the measurement, and its flow characteristics are very complex, so it is difficult to describe it completely by the mathematical model, so it is difficult to measure. The on-line detection and imaging technology and implementation scheme of multiphase flow are mostly in the laboratory test and research stage, only a few of them have been commercialized, and can be widely used in on-line detection, and most of them are suitable for two-phase flow. There is still a lack of an image reconstruction algorithm which is suitable for unknown media and adaptive phase number, so further research is needed to promote its development and practicality. Learning is one of the characteristics of machine learning. Through the change of learning medium, such as distribution change and dielectric constant change, the algorithm parameters can be adjusted in time. ECT technology is a kind of on-line multi-phase flow detection technology, which is widely used in the field of offshore oil exploitation and other industrial research. However, there are many problems and difficulties in the application of ECT technology, which is not perfect yet. From the point of view of weak capacitance processing and image reconstruction algorithm with unknown dielectric constant, this paper studies the image reconstruction of ECT based on machine learning method. The main work and contribution are as follows: (1) aiming at the "soft field" feature of ECT, In this paper, the normalized model of ECT capacitance in the data preprocessing part is studied. Based on the analysis of the physical characteristics of capacitance normalization and the parallel normalization method, a weighted capacitance normalization model is established and applied to image reconstruction based on SVM. Compared with the parallel model, the method is applicable not only to the two-phase flow but also to the multi-phase flow above the two-phase flow under the condition that the number of phases is determined and the medium in the tube is not changed. The correlation between the image reconstruction and the real model is higher than the parallel normalization method. (2) in the practical application of ECT system, two-phase flow and multi-phase flow are the most common cases of fluid. It is not only because the dielectric constant of the measured medium changes with the change of temperature and other environments, but also because of the existence of other media in the measured field, which makes the measurement medium unknown. In this paper, the SVC-based electrical capacitance tomography image reconstruction algorithm is proposed to reconstruct the unknown dielectric constant object by using the support vector machine (SVM) method in machine learning. The simulation results show that the image reconstruction algorithm is based on the electrical capacitance tomography (ECT). In the case of phase number determination, the algorithm can effectively adapt to the change of media diversity, that is, for different media, This algorithm can have high image reconstruction accuracy. (3) the existing ECT reconstruction algorithms are usually reconstructed only when the number of phases is determined and the medium is unchanged. In order to solve this problem, the author established a machine learning method based on SVM decision tree to predict the number of phases. The method of SVM decision tree is used to predict the medium in the tube when the number of phases is uncertain. The experimental results show that: When the number of phases is uncertain, the method can distinguish the number of phases in the tube and the medium contained in the tube. Finally, based on the above methods, an adaptive phase number ECT image reconstruction algorithm based on SVM decision tree is designed. When the phase number is unknown and the medium changes, how to reconstruct the ECT image is preliminarily analyzed, and the purpose of improving the reconstruction accuracy is achieved. It provides a new research idea for ECT technology.
【学位授予单位】：上海海洋大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【参考文献】