As a significant subset of unmanned underwater vehicles (UUVs), autonomous underwater vehicles (AUVs) possess the capability to autonomously execute tasks. Characterized by its flexibility, cost-effectiveness, extensive operational range, and robust environmental adaptability, AUV has emerged as the primary technological apparatus for deep-sea exploration and research. In this paper, we present the design of a 10,000 m class AUV equipped with capabilities such as fixed-depth navigation, regional autonomous cruising, full-depth video recording, and temperature and salinity profiling. Initially, we outline the comprehensive design of the AUV, detailing its structural configuration, system components, functional module arrangement, and operational principles. Subsequently, we compute the hydrodynamic parameters using a spatial kinematics model. Finally, the AUV designed in this paper is tested for its functions and performance, such as fixed-depth sailing, maximum speed, and maximum diving depth, and its reliability and practicability are verified.
11 October 2023 in IEEE Transactions on Neural Networks and Learning Systems
In real classification scenarios, the number distribution of modeling samples is usually out of proportion. Most of the existing classification methods still face challenges in comprehensive model performance for imbalanced data. In this article, a novel theoretical framework is proposed that establishes a proportion coefficient independent of the number distribution of modeling samples and a general merge loss calculation method independent of class distribution. The loss calculation method of the imbalanced problem focuses on both the global and batch sample levels. Specifically, the loss function calculation introduces the true-positive rate (TPR) and the false-positive rate (FPR) to ensure the independence and balance of loss calculation for each class. Based on this, global and local loss weight coefficients are generated from the entire dataset and batch dataset for the multiclass classification problem, and a merge weight loss function is calculated after unifying the weight coefficient scale. Furthermore, the designed loss function is applied to different neural network models and datasets. The method shows better performance on imbalanced datasets than state-of-the-art methods.
1 May 2025 in Journal of Atmospheric and Oceanic Technology
Sea surface temperature (SST) is a key variable in understanding air–sea interactions, and changes in SST can have profound effects on the global climate and may lead to extreme events such as marine heat wave and floods. Accurate prediction of SST is, therefore, essential for detecting extreme events and mitigating the negative consequences they cause. Numerical model-based predictions are limited by seafloor topography data, parameterization schemes, comprehensive preparation of input data, and computing resources. Recently, data-driven methods bypassing these limitations have merged, but they still face difficulties with nonlinearities in the data. Many traditional models rely on linear assumptions, making it challenging to capture complex climate patterns and rendering them sensitive to data noise and external disturbances, which can lead to inaccurate predictions. Additionally, these methods often exhibit low robustness when dealing with highly nonstationary time series, struggling to adapt to sudden climate events. In this study, we combine empirical mode decomposition (EMD) and the gated recurrent unit (GRU) network to improve the SST prediction based on daily observations. By decomposing complex data into several intrinsic mode functions, EMD can help identify and analyze the underlying nonlinear dynamics, allowing the GRU to better learn the underlying trends and patterns in the signals. A case study in the Bohai Sea shows that the EMD–GRU model significantly outperforms the GRU model. Furthermore, prediction errors indicate that the EMD–GRU model can improve SST prediction by an average of about 45% compared to the Hybrid Coordinate Ocean Model (HYCOM). This study demonstrates that a deep learning model combined with EMD is highly promising for nonlinear and nonstationary data predictions. Significance Statement: In this paper, an accurate and effective prediction method is proposed to overcome the challenges of existing models for sea surface temperature predictions. Extensive experiments in the Bohai Sea show that the proposed prediction method can achieve accurate prediction of sea surface temperature up to half a month. In addition, compared with the traditional numerical models such as HYCOM, the proposed method can greatly improve the prediction performance of sea surface temperature. This study reflects the potential of proposed prediction methods in ocean intelligent prediction and provides an effective tool for accurate and efficient prediction of ocean variables. In short, this study has academic value and practical significance.