무인잠수정의 동력학 식별 및 LSTM 신경망 적용을 위한 검토

Abstract

This paper explores a potential dynamic system identification based on Long-Short Term Memory (LSTM) neural network for the remotely operated vehicle Hemire by using data sequences of input and motion responses. Before implementing the method, we have examined a classical least square error method to discern the major factors of dynamic system identification and to review the dynamic coefficients of the ROV’s mathematical model using the data sequences. Outliers and temporal blackouts included in the measured velocities should be removed. Those noises were caused by too short bottom distance during close bottom-following survey and irregular reflection of the DVL’s acoustic beam from the sea floor. Lever-armeffect also should be compensated to get precise motion responses at the center of the vehicle. After filtering and smoothing the velocity data, dynamical coefficients were estimated with the least square method. The dynamic system identification of the ROV with LSTM neural network, however, has yet been completed. Further works are needed to learn the network with sufficient available data.