Sensitivity matrix update method based on Residual Attention Fusion Network

Electrical resistance tomography (ERT) is a visualization technique that can be used to monitor gas phase distribution in GAS-LIQUID-SOLID three-phase fluidized beds. The image reconstruction algorithm, based on the sensitivity matrix, takes the uniform distribution of the liquid phase as the initial condition to obtain a general sensitivity matrix, which is numerically different from the theoretical sensitivity matrix under the real gas phase distribution, resulting in a large image reconstruction error. To solve this problem, a residual attention fusion network (RAF-Net) is proposed in this paper. It can update the sensitivity matrix by correcting the field strength distribution according to the boundary measurement value. A new channel fusional residual spatial attention block is proposed to focus more attention to the channels containing medium spatial information where we calculate the spatial attention weight parameters. And, a new loss function, based on the gas phase distribution, is proposed to make the model fit the field strength distribution better. In this paper, 49841 samples are generated through simulation and divided into training and test sets. The method proposed in this paper can obtain a theoretical sensitivity matrix with high accuracy on the test set. The Landweber algorithm, using the sensitivity matrix obtained by RAF-Net model, has higher reconstruction quality than the general sensitivity matrix. When Gaussian white noise with SNR of 50 dB, 40 dB, 30dB and 20dB is added to the test set, our method can still obtain a quite accurate sensitivity matrix and obtain the approximate position of the gas phase distribution through the reconstruction algorithm, which proves the noise immunity of the method.