Autonomous underwater vehicle motion state recognition and control pattern mining

Self-awareness of its own state during autonomous operation is critical for Autonomous Underwater Vehicles (AUVs) to execute tasks and monitor their health. Automated data analysis methods appear to be the most appropriate tool for attempting to understand and elucidate the motion states of AUVs. However, research scholars often focus on the state recognition of small AUVs used for shallow water detection, and the relevant conclusions cannot be directly applied to deep-sea AUVs. As a representative deep-sea AUV, Qianlong-3 is studied in this paper. In this study, a data-driven approach was employed to recognize the motion states and control modes of the Qianlong-3 AUV while retaining the original data noise. First, the class imbalance problem in the data was addressed, and then the data features were enhanced through a sliding window statistical method. By comparing the differences in the accuracy of motion state recognition after class imbalance processing using different supervised learning methods, the Random Forest (RF) algorithm was found to be the most suitable classification algorithm. In addition, the Toeplitz Inverse Covariance-Based Clustering (TICC) algorithm was used to recognize and mine the motion states and control modes of the AUV. The TICC algorithm was effective in extracting latent information from the actual data. Finally, through correlation analysis, the three motion states of the AUV were discussed, and the control modes for each of these states were determined. Overall, this study demonstrates the feasibility of relevant algorithms in practical deep-sea applications for state recognition and control mode mining. This provides valuable reference and experience for the development of future deep-sea AUV.

Automated summary

In this study, a data-driven approach was used to recognize the motion states and control modes of the Qianlong-3 AUV. By addressing the class imbalance problem in the data and enhancing the data features, the Random Forest algorithm was found to be the most suitable classification algorithm, and the Toeplitz Inverse Covariance-Based Clustering algorithm was effective in extracting latent information from the actual data. This study demonstrates the feasibility of relevant algorithms in practical deep-sea applications for state recognition and control mode mining.