Seamless INS/DVL Integrated Navigation System via Online Transfer Gaussian Process Regression

In inertial navigation system (INS)/Doppler velocity log (DVL) integrated navigation systems, DVL signals are susceptible to interruptions, causing challenges to seamless navigation. In this article, a pseudo-DVL measurement construction method is proposed based on the Gaussian process regression (GPR) technique. Time-varying characteristics, such as maneuver changes, in the navigation system may cause invalid regression of the traditional GPR during long-term DVL outages. To solve this problem, a novel online transfer GPR (OTGPR) algorithm is proposed. The OTGPR utilizes a virtual dataset and a semiparametric transfer kernel to enhance the coverage of the sample distribution. In view of the computational complexity and sample distribution diversity, the Kullback-Leibler divergence-based online training dataset construction algorithm is designed in the OTGPR model. Finally, the virtual DVL measurement vector and measurement noise matrix for the adaptive Kalman filter are derived from the posterior mean and variance of OTGPR through coordinate transformation matrices. Sea trial experiment with Monte Carlo analysis indicates the effectiveness of the OTGPR-based seamless navigation system. Under diverse randomized DVL outage conditions, the reduction in the root-mean-square value of velocity errors reached 50.52%, and in the case of DVL full-beam outage for 2500 s, the reduction reached 93.6%.