ECT Attention Reverse Mapping algorithm: visualization of flow pattern heatmap based on convolutional neural network and its impact on ECT image reconstruction

The flow pattern is one of the most basic characteristic parameters of oil-gas two-phase flow, and it has a great influence on the accurate measurement of other parameters of two-phase flow. Over the past decade, the convolutional neural network (CNN) algorithm has been widely used in flow pattern research. Unfortunately, the flow pattern research based on the CNN algorithm is more on model structure optimization, and there is still little insight into the relationship between the CNN algorithm and the physical meaning of the flow pattern. Thus, in this paper, inspired by the neural network visualization gradient-based class activation mapping (Grad-CAM) method, we propose the electrical capacitance tomography (ECT) Attention Reverse Mapping algorithm (EARM) to explore the relationship between the physical meaning of flow patterns and the CNN algorithm. Specifically, the Grad-CAM method is used to obtain heatmaps of flow patterns, and the EARM algorithm combines the hotspot information of the flow pattern heatmap with the ECT image reconstruction principle, which deeply explores the relationship between the CNN flow pattern identification and the ECT image reconstruction algorithm. Furthermore, we conduct prediction experiments based on the parameters of the flow pattern hotspot capacitance data, and the experimental results are compared with the ECT original capacitance data parameter prediction. The prediction accuracy of oil-gas two-phase flow parameters has been improved by more than 50% on average, and experiments have verified the correctness of the visualization of CNN network flow pattern identification.

Automated summary

This paper explores the relationship between the physical meaning of flow patterns and the CNN algorithm by proposing the EARM algorithm, which combines the Grad-CAM method and the ECT image reconstruction principle. The prediction accuracy of oil-gas two-phase flow parameters has been significantly improved. Experiments have verified the correctness of the visualization of CNN network flow pattern identification.