다채널 시계열 신호 기반 3 차원 측위를 위한 심층강화학습

Abstract

Several fields employ a method for estimating the source position of a multichannel time series signal received via multiple sensors. This requires an algorithm to estimate the position from the propagation time difference between the signals. Moreover, the algorithm needs to be modified according to the geometric condition of the propagation space, propagation speed in the medium, and noise caused by the reflector. However, since it is difficult to supplement the algorithm to respond individually to various propagation conditions, the more complex the propagation conditions, the less precise the source localization. To overcome these limitations, this study proposes a reinforcement learning algorithm to estimate the source location using simulations that generate multi-channel timeseries signals from random locations, which are propagated through three-dimensional space and received by multiple sensors. The reinforcement learning model estimates the source position by inputting the difference between the target signals and the signals generated from the random position as well as moving the positions to a new expected source positions where the differences can be minimized. By repeating the movement of the expected source positions, if the changes are repeated within the allowable minimum value in which the position retains in a specific region, the corresponding position can be estimated to be the source position. The proposed reinforcement learning model uses 300 samples from seven sensor positions at a sampling rate of 0.5 kHz and a propagation speed of 5000 m/s in a space with dimension of 1000 mm width, 1000 mm length, and 1000 mm height. From estimation of 1000 random source positions with the condition, the error was less than 50 mm in 90% or more. Thus, by reconfiguring the simulation environment to conform to a new target situation and learning the model, source localization models can be developed.