Study On Image Reconstruction Parallelism For Electrical Capacitance Tomography

In the late 1980s, a novel technology of computerized tomography (CT) which is named electrical capacitance tomography (ECT) was proposed by the researchers of UMIST of U. K. It has the advantages of low-cost, wide application, simple structure, non-invasive, good security performance and so on. Now it has become the mainstream development direction and research focus of the process tomography technique. In fact, the image reconstruction algorithm is a key factor to test various parameters of multiphase Flow in ECT system research. It also gives a feasible method to solve the multi-parameter visualization measurement of the multiphase flow. The BP neural network is chosen as the basic image reconstruction algorithm in the paper to do research for Image reconstruction parallelism.In the process of the ECT image reconstruction with BP neural network, there are two issues such as low convergence rate and easily falling into local minimum value, a new SA-BP algorithm for image reconstruction is presented to solve these issues. Firstly, Adding momentum term in the weight updating formula of the BP neural network and applying adaptive method adjust learning rate, so this algorithm can improve the convergence speed of neural network. Secondly, the introduction of simulated annealing algorithm based on the BP neural network, which is improved in last step, plays an important part in guiding the process of network training. Using this optimization algorithm can avoid trapping in the local minimum value in the BP neural network training process.The parallel BP neural network for ECT image reconstruction is a very good research direction due to the inherent parallelism of neural networks. In the section of parallelism analysis of image reconstruction, the parallel algorithms and ideas are introduced. After comparing with the structure parallel strategy and data parallel strategy in detail, the data parallel strategy is used in the ECT image reconstruction parallel experiments. Then, the principle of grid computing and GT4 which is a tool to build a grid platform is introduced. In the experiment, the neural network is designed and analyzed detailedly, and a large number of experimental data is collected by finite element analysis method. Finally, in a built grid platform, the training using SA-BP algorithm for the four kinds of typical flow pattern is given and the target network is generated. Comparing the results of this experiment with the other two ECT image reconstruction algorithms, it is proved that the SA-BP algorithm based on parallel image reconstruction method has s certain degree of improvement both in the reconstruction speed and accuracy.