Intelligent damage detection for deep-sea aquaculture cages using multi-sensor data fusion

Abstract

Netting damage in deep-sea aquaculture cages poses significant risks to operational safety and economic sustainability. Traditional manual inspection methods are inefficient, underscoring the necessity for real-time damage monitoring. This study proposes a damage detection method based on convolutional neural networks (CNN) and the Dempster–Shafer (D–S) evidence theory. Hydrodynamic simulations under varying wave and current conditions were conducted to construct a comprehensive dataset. Features were extracted from this dataset using continuous wavelet transform and subsequently used to train the CNN model. The output damage recognition probabilities from individual sensors were treated as basic probability assignments (BPAs). A two-level fusion strategy based on the D–S theory was developed: the first level fuses data from tension and acceleration sensors across 12 monitoring locations, and the second level aggregates these results regionally. The proposed CNN–DS model achieves a high detection accuracy of 99.07%, enabling accurate damage localization and enhancing repair efficiency. Compared to single-sensor approaches, such as the tension-based and the acceleration-based model, the proposed multi-sensor fusion model improves accuracy by 7.6% and 6.1%, respectively. This method shows promise for broader applications in monitoring other marine flexible structures and equipment.