Hybrid network of convolutional neural networks and transformer for metal debris imaging with electrostatic tomography.

Saved in:

Bibliographic Details
Title:	Hybrid network of convolutional neural networks and transformer for metal debris imaging with electrostatic tomography.
Authors:	Wang, Ying¹ 15111018@bjtu.edu.cn, Liu, Xianglong², Li, Danyang², Zhang, Kun², Feng, Huilin²
Source:	Insight: Non-Destructive Testing & Condition Monitoring. Jun2026, Vol. 68 Issue 6, p380-388. 9p.
Subjects:	Convolutional neural networks, Transformer models, Tomography, Artificial neural networks, Inverse problems
Abstract:	In this paper, electrostatic tomography (EST) is applied to the charge distribution imaging of metal abrasive particles in a pipeline, which provides technical support for the detection of abrasive particles in liquid-solid/gas-solid two-phase flow in the industrial field. The measured induced charge signal of charged particles on a 16-electrode array is used as input data to an inverse problem solver to reconstruct the charge density distribution in the target area. EST is a method of solving the inverse problem to reconstruct the charge density distribution in the target region. However, the solution of the EST inverse problem is challenging due to the absence of independent measurements. This leads to the relatively low accuracy and efficiency of traditional algorithms in the solution process, making it difficult to accurately estimate the actual charge distribution. To address this problem, this paper designs a hybrid residual network (ResNet) and transformer network (RTHN) model for solving the inverse problem. For the hybrid network design, ResNet-18 and a vision transformer (ViT) are combined to employ the advantages of convolutional neural networks (CNNs) while utilising all the benefits of transformers, by fully making use of the deep feature extraction ability of CNNs and the global modelling ability of the transformer. The first two stages of ResNet-18 and an eight-head attention mechanism are used, which can capture long-distance dependency and effectively use local feature information to extract key features in images. After 250 iterations, results using simulated data show that the RTHN model achieves significant performance improvement in EST image reconstruction tasks, possessing higher accuracy and efficiency than traditional methods. It also has better performance than CNNs and ResNet models alone. In addition, three types of noise are added to the simulated data to test the anti-noise performance and the random sample reconstruction ability is tested, proving the noise resistance performance and generalisation ability of the model. A 16-electrode sensor experimental system is designed to verify the effectiveness of the proposed hybrid model for metal abrasive particle detection. The experimental results demonstrate the potential and application prospect of the proposed model in EST imaging of metal debris. [ABSTRACT FROM AUTHOR]
	Copyright of Insight: Non-Destructive Testing & Condition Monitoring is the property of British Institute of Non-Destructive Testing and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database:	Engineering Source

Description
Abstract:	In this paper, electrostatic tomography (EST) is applied to the charge distribution imaging of metal abrasive particles in a pipeline, which provides technical support for the detection of abrasive particles in liquid-solid/gas-solid two-phase flow in the industrial field. The measured induced charge signal of charged particles on a 16-electrode array is used as input data to an inverse problem solver to reconstruct the charge density distribution in the target area. EST is a method of solving the inverse problem to reconstruct the charge density distribution in the target region. However, the solution of the EST inverse problem is challenging due to the absence of independent measurements. This leads to the relatively low accuracy and efficiency of traditional algorithms in the solution process, making it difficult to accurately estimate the actual charge distribution. To address this problem, this paper designs a hybrid residual network (ResNet) and transformer network (RTHN) model for solving the inverse problem. For the hybrid network design, ResNet-18 and a vision transformer (ViT) are combined to employ the advantages of convolutional neural networks (CNNs) while utilising all the benefits of transformers, by fully making use of the deep feature extraction ability of CNNs and the global modelling ability of the transformer. The first two stages of ResNet-18 and an eight-head attention mechanism are used, which can capture long-distance dependency and effectively use local feature information to extract key features in images. After 250 iterations, results using simulated data show that the RTHN model achieves significant performance improvement in EST image reconstruction tasks, possessing higher accuracy and efficiency than traditional methods. It also has better performance than CNNs and ResNet models alone. In addition, three types of noise are added to the simulated data to test the anti-noise performance and the random sample reconstruction ability is tested, proving the noise resistance performance and generalisation ability of the model. A 16-electrode sensor experimental system is designed to verify the effectiveness of the proposed hybrid model for metal abrasive particle detection. The experimental results demonstrate the potential and application prospect of the proposed model in EST imaging of metal debris. [ABSTRACT FROM AUTHOR]
ISSN:	13542575
DOI:	10.1784/insi.2026.68.6.380