Frequency domain airborne EM inversion based on improved particle swarm depth neural network

The traditional gradient inversion depends on the selection of initial model and is easy to fall into local minimum, which affects the accuracy and convergence speed of inversion to a certain extent. To solve the problems, a frequency domain airborne EM inversion method based on an improved particle swarm deep neural network was proposed. Firstly, a large number of sample data sets were obtained through model forward modeling. Secondly, the basic framework of the deep neural network was established based on the data sets. The network input was the normalized vertical magnetic field component and the output was the corresponding electrical model parameters. Thirdly, an inertial weight oscillation attenuation method was proposed to improve the global optimization ability of particle swarm optimization algorithm, and the swarm algorithm optimizes the training process of the deep neural network to obtain the optimal solution of the connection weights and thresholds. Finally, the optimal weights and thresholds were used as the initial values of the network which were used to perform inversion tests on the unknown geoelectric model. In this paper, the inversion results of the improved particle swarm deep neural network algorithm(IPSO-DNN), the particle swarm deep neural network algorithm(PSO-DNN) and the single deep neural network algorithm(DNN) were tested by layered geological model, and this method was applied to inversion of measured aeromagnetic data. The results show that the improved particle swarm neural network algorithm can make full use of the global searching capability of particle swarm optimization and the local optimization of deep neural network. In the process of inversion, it can effectively prevent the inversion from falling into the local minimum, find the global optimal solution and accurately reverse the parameters of geoelectric model. Compared with the particle swarm neural network and a single neural network inversion method, this method has higher accuracy and convergence speed. © 2020, Central South University Press. All right reserved.