A Multihead ProbSparse Self-Attention Mechanism-Based High-Precision and High-Robustness Reconstruction Model for Missing Ocean Data

Ocean sensor networks (OSNs) can sense long-time series data, which plays a crucial role in tasks that rely on massive sensory data such as ocean ecosystem monitoring and maritime regulation. Unfortunately, ocean nodes are prone to failures, and harsh ocean channels, and natural environments can interfere with the data collection and transmission process, which can lead to large-scale loss of ocean sensory data. Based upon the comprehensive and in-depth analysis of the ocean’s time-varying dynamic environment, we proposed a multihead ProbSparse self-attention mechanism (MPSM)-based Model to predict and reconstruct missing ocean data. Specifically, MPSM can learn key features of ocean observation data on both sides of missing gaps with a dual encoder architecture. Subsequently, the MPSM can efficiently capture intersequence dependencies in parallel with lower computational complexity. Moreover, its residual connection will enhance the model’s expressive ability and alleviate gradient vanishing. Simulation experiment results demonstrate that the metrics of MPSM are lower than the selected benchmark algorithms on different ocean datasets, which means MPSM has a stronger robust and generalizability. MPSM can provide high-precision, low-latency missing data restoration capabilities for marine monitoring scenarios. It enhances the continuity and reliability of ocean observation systems.