Research On Solving Inverse Source Problem By Machine Learning

The inverse point source problem refers to the inversion of the parameters of the unknown point source such as location and intensity by using measurement data.At present,it has become an important problem in medical imaging,environmental pollution,tomography and antenna synthesis.In practice,the measured data are usually collected in a limited observed aperture and compromised by noises.With the decrease of observation information,the difficulty of point source inversion will increase significantly.Machine learning algorithms can compensate the lack of information and reduce noise through a large number of training data,and has a good approximation effect on nonlinear mapping.Therefore,this paper constructs machine learning models to solve the acoustic inverse source problem in frequency domain and time domain respectively using the limited observed aperture data.Aiming at the inverse point source problem of acoustic wave in frequency domain,a location parameter inversion model was constructed based on neural network with gating thought(LPIMNNG).The single frequency far-field measurements collected in a limited aperture and the location parameters of the point source are taken as the input and output sequences,respectively.The network state is selectively updated to preserve the specific structure of the underlying data through a gated idea of neural network and the use of the long-term memory function.The parameters of the network are updated by self-learning algorithm so as to reconstruct the location of the point source.The computational cost of the LPIMNNG are theoretically calculated by the Multiply Accumulate(MACC),and the computational complexity of the model was given.The experimental results show that the model has a good inversion effect on the measurement data collected in a limited aperture,and the model has certain robustness and generalization under noise interference.The effect of distance between point sources on the inversion results of the model is weak.Aiming at the inverse point source problem of acoustic wave in time domain,a moving trajectory inversion and prediction model(TPIMNN-Res Net)is proposed based on gating thought of neural network with Residual network(Res Net)to invert and predict the trajectory of the moving point source.Firstly,a trajectory parameter inversion model based on neural network is constructed(TPIMNN).The limited-aperture data and the trajectory parameters of the moving point source are taken as the input and output sequences,respectively.Selectively extract data features and invert trajectory parameters.The parameters of the network are updated by self-learning algorithm so as to reconstruct the trajectory of the moving point source in a limited time window.The computing complexity of the TPIMNN are calculated by the Multiply Accumulate.Then,the Residual network is used to predict the trajectory of the point source in the next period.The experimental results show that the TPIMNN-Res Net has certain robustness and generalization,and has good inversion effect on the measurement data collected in a limited aperture,and can inversion and predict the trajectory of point source in a limited time window.