Scholarlink

<p>The classification of spherical cap bubble (SCB) flow in pipes is a major challenge when gas and liquid flow simultaneously in many industrial processes. Moreover, several conventional methods are used to classify SCB flows but are restricted by operating conditions and limited databases. Furthermore, real-time automated monitoring of gas–liquid flow is challenging due to inadequate image processing techniques. This study uses conventional, convolutional neural network (CNN) and hybrid methods to classify air and silicone oil SCB flow in an upward vertical pipe. Furthermore, three experimental runs were used to obtain images for the SCB flow (fine, fine-coarse, and coarse SCB bubbles) using advanced wire mesh sensor (WMS) instrumentation. Additionally, the WMS time-averaged velocity measurements for the three experimental runs were validated using electrical capacitance tomography (ECT) instrumentation. Monotonic and nonmonotonic pressure profiles were observed along the pipe yielding pressure drops across the entire pipe length. The testing supervised CNN classification results showed an accuracy (97.52%), sensitivity (97.32%), specificity (98.80%), precision (97.21%), and F1 score (97.30%) compared to the state-of-the-art supervised machine learning (ML) models, with SVM having an accuracy of 49.60% being the best. CNN was combined with SVM as a hybrid classifier, which resulted in an improved accuracy (98.57%), sensitivity (97.73%), specificity (99.17%), precision (97.62%), and F1 score (97.66%). Additionally, local interpretable model-agnostic explanations, gradCAM, and occlusion sensitivity as explainable artificial intelligence (XAI) techniques integrated into the CNN model were used to explain SCB image results that contributed to the total classification scores. Furthermore, the receiver operating characteristic (ROC) curve, area under the curve (AUC), and confusion matrix (CM) also showed that both the CNN and hybrid CNN-SVM classifiers performed well on the testing datasets. Last, the SCB classifications with the standalone CNN and hybrid CNN-SVM models’ generalization capability have improved SCB classification accuracy compared to conventional techniques with certainty and confidence for industrial applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Modeling of spherical cap bubble flow patterns in a vertical pipeline using experimental wire mesh sensor data

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Eric Thompson Brantson,
  • Mukhtar Abdulkadir,
  • Emmanuel Epelle,
  • Francis Adjei,
  • Fuseini Naziru Issaka,
  • Zainab Ololade Iyiola,
  • Adu-Awuku Joel,
  • Nannan Liu,
  • Laliri Bright Ntebu

摘要

The classification of spherical cap bubble (SCB) flow in pipes is a major challenge when gas and liquid flow simultaneously in many industrial processes. Moreover, several conventional methods are used to classify SCB flows but are restricted by operating conditions and limited databases. Furthermore, real-time automated monitoring of gas–liquid flow is challenging due to inadequate image processing techniques. This study uses conventional, convolutional neural network (CNN) and hybrid methods to classify air and silicone oil SCB flow in an upward vertical pipe. Furthermore, three experimental runs were used to obtain images for the SCB flow (fine, fine-coarse, and coarse SCB bubbles) using advanced wire mesh sensor (WMS) instrumentation. Additionally, the WMS time-averaged velocity measurements for the three experimental runs were validated using electrical capacitance tomography (ECT) instrumentation. Monotonic and nonmonotonic pressure profiles were observed along the pipe yielding pressure drops across the entire pipe length. The testing supervised CNN classification results showed an accuracy (97.52%), sensitivity (97.32%), specificity (98.80%), precision (97.21%), and F1 score (97.30%) compared to the state-of-the-art supervised machine learning (ML) models, with SVM having an accuracy of 49.60% being the best. CNN was combined with SVM as a hybrid classifier, which resulted in an improved accuracy (98.57%), sensitivity (97.73%), specificity (99.17%), precision (97.62%), and F1 score (97.66%). Additionally, local interpretable model-agnostic explanations, gradCAM, and occlusion sensitivity as explainable artificial intelligence (XAI) techniques integrated into the CNN model were used to explain SCB image results that contributed to the total classification scores. Furthermore, the receiver operating characteristic (ROC) curve, area under the curve (AUC), and confusion matrix (CM) also showed that both the CNN and hybrid CNN-SVM classifiers performed well on the testing datasets. Last, the SCB classifications with the standalone CNN and hybrid CNN-SVM models’ generalization capability have improved SCB classification accuracy compared to conventional techniques with certainty and confidence for industrial applications.