In acoustic fatigue testing, it is necessary to monitor the acoustic load on the surface of the test specimen in real time to ensure the accuracy of the test. However, in most cases, it is not feasible to directly install microphones on the surface of the test specimen. Therefore, it is necessary to use microphones around the test specimen to effectively infer the surface sound pressure of the specimen, which is known as sound field reconstruction. Traditional methods such as plane wave decomposition and point source decomposition require a large number of microphones to achieve better accuracy within a certain spatial scale. This paper proposes a sound field reconstruction algorithm based on neural networks, attempting to establish a time domain mapping between the measured sound pressure at the measurement points and the surface sound pressure of the test specimen. Simulation results have proven the superiority of this method.