A Computation Method for Separate Flow Velocity Based on ERT in Dredging Engineering
2024 IEEE INTERNATIONAL INSTRUMENTATION AND MEASUREMENT TECHNOLOGY CONFERENCE, I2MTC 2024(2024)
Code
Data
The authors of this paper include Zhao Yuye, Yue Shihong, Li Kun, and Dong Fanpeng. Zhao Yuye is from the School of Electrical and Information Engineering at Tianjin University, with a research focus on Electrical Tomography. Yue Shihong's research involves Pattern Recognition and Intelligent Systems, as well as electrical-based industrial and medical testing, having chaired multiple National Natural Science Foundation of China projects. Li Kun is a faculty member in both the School of New Media and Communication and the School of Intelligence and Computing at Tianjin University, with research interests in Super-Resolution, 3D Reconstruction, Data Mining, and Feature Extraction, and has received honors such as the National Outstanding Young Scholar of Beiyang. Dong Fanpeng's research areas include Electrical Tomography, Acousto-Optic Effect, Optical Sensing, and Sensors.
1. Introduction
- Importance of flow velocity in multiphase flow detection
- Advantages and applications of Electrical Resistance Tomography (ERT)
- Limitations of flow velocity calculation methods based on Cross-Correlation (CC) principle
- Limitations of flow velocity calculation methods based on Convolutional Neural Networks (CNN)
- Method proposed in this paper: Hybrid CNN-RKSVM based on ERT
2. Related Technologies
- 2.1 Cross-Correlation Principle
- 2.2 Convolutional Neural Networks
- 2.3 Reproducing Kernel
- 2.4 Reproducing Kernel Support Vector Machine (RKSVM)
3. Hybrid Network CNN-RKSVM
- 3.1 Network Structure
- 3.2 Network Training
- 3.3 Sample Dataset
4. Results and Discussion
- 4.1 Evaluation Parameters
- 4.2 Results and Discussion
5. Conclusion
- Advantages of the CNN-RKSVM method in flow velocity detection
- Future research directions
Q: What specific research methods were used in the paper?
- Cross-correlation method (CC): The flow velocity in the pipeline was calculated by comparing the measurement sequences obtained from two adjacent ERT sensors using the cross-correlation principle.
- Convolutional Neural Network (CNN): The powerful feature extraction capability of CNN was utilized to extract key features from ERT measurement data and to predict flow velocity.
- Support Vector Machine based on Reproducing Kernel Function (RKSVM): RKSVM was used to map the feature data extracted by CNN into a high-dimensional feature space for flow velocity prediction.
- Hybrid Network (CNN-RKSVM): The CNN and RKSVM were combined, with CNN used for feature extraction followed by RKSVM for prediction, to improve the accuracy and generalization ability of the flow velocity calculation.
Q: What are the main research findings and achievements?
- The CNN-RKSVM method outperforms traditional CC, RKSVM, and CNN methods: In both low and high flow velocity conditions, the RMSE and MAPE of the CNN-RKSVM method were lower than those of other methods, indicating higher prediction accuracy.
- The CNN-RKSVM method has stronger generalization ability: The CNN-RKSVM method maintained high prediction accuracy under different types of solid particles, solid phase fractions, and flow velocity conditions.
- The CNN-RKSVM method has a faster response time: Although the computation time of the CNN-RKSVM method is slightly longer than that of the CC method, its response time is still less than 1 second, meeting the practical requirements of dredging engineering.
Q: What are the current limitations of this research?
- Limitation of the sample dataset: The sample dataset used in the paper only contains specific types of solid particles and specific fluid parameters, which cannot fully represent all working conditions in dredging engineering.
- Limitation of CNN feature extraction: The feature data extracted by CNN may lose some neighborhood information, which could affect prediction accuracy.
- Quality of ERT reconstruction images: The paper did not consider the impact of the quality of ERT reconstruction images on prediction accuracy, and it will be necessary to select an appropriate image reconstruction algorithm in the future.