Error Compensation for Dead Reckoning Based on SVM

In the use of machine learning methods for error compensation in dead reckoning of an autonomous undersea vehicle(AUV), the neural network algorithm is commonly used. However, neural networks require a large number of training samples to achieve stable training results. To solve this problem, research was conducted on the application of support vector machine(SVM) for error compensation in dead reckoning. By utilizing SVM, an error compensation model was trained to correct the errors in dead reckoning, thereby improving navigational accuracy. The error compensation model takes seven parameters as input: pitch angle, roll angle, course angle, forward, right, and upward velocity of the Doppler velocity log(DVL) relative to the ground, and dead reckoning time of the AUV. The difference in latitude and longitude provided by the global positioning system(GPS) and inertial navigation system(INS) + DVL combination compared with latitude and longitude obtained from dead reckoning serves as the output of the model. The SVM trained model and the neural network trained model show a relative error of 0.28% and 0.93%, respectively, when the amount of data is limited. Through lake tests, it is concluded that the model trained by SVM can control the relative error of dead reckoning within 0.5%.