Abstract
The precise identification of operating conditions is a critical prerequisite for ensuring the safety of manned deep-sea submersibles, a task complicated by extreme environments and tightly coupled subsystems. Traditional methods, which often overlook the latent correlations among monitoring variables, are frequently insufficient for this high-stakes application. To address these issues, this paper proposes a novel working condition identification method tailored to the challenges of the deep-sea environment, which leverages graph neural networks to explicitly model the relationships between sensor variables. The proposed framework first dynamically constructs a graph structure from multi-sensor snapshots using the K-Nearest Neighbor (K-NN) algorithm, where edges represent high similarity in the submersible's operational state. Subsequently, a graph neural network is employed to learn from this relational data. Furthermore, a model update strategy is introduced to enable the adaptive recognition of new, emerging operational conditions. A case analysis using real operational data from a manned deep-sea submersible demonstrates that the proposed method significantly enhances identification accuracy. Moreover, interpretability analysis reveals that the model learns physically meaningful patterns consistent with the submersible's engineering principles, enhancing its trustworthiness.