现代化海洋牧场智能投喂技术与装备发展综述OA

The escalating demand for precision feeding in offshore aquaculture,where feed typically accounts for approximately 70%of full-lifecycle production costs,stands in direct tension with the constrained sensing fidelity and operational reliability imposed by harsh sea conditions,driving feeding systems toward intelligent transformation.This review delineated the scope of intelligent feeding systems for modern marine ranching and systematically examined technological progress and remaining bottlenecks along the"perception-decision-execution"closed-loop framework.At the perception layer,five core environmental parameters(water temperature,dissolved oxygen,current velocity and direction,turbidity,and meteorological conditions)and their sensing principles were reviewed,followed by non-contact biomass estimation via stereo-vision keypoint regression and three-dimensional reconstruction,and feeding-intensity quantification that fused visual cues(residual pellets,surface splash features,and underwater behavior)with active and passive acoustic monitoring.At the decision layer,mechanistic models,including the thermal-growth-coefficient growth curve and bioenergetic balance equations were systematically compared with dynamic models driven by fuzzy control,machine learning,and deep learning architectures(CNN-GRU and Transformer variants),highlighting the trade-off between interpretability and adaptability.At the control and execution layer,the hierarchical"cloud-edge-device"architecture was examined alongside scenario-matched configurations spanning land-based recirculating systems,nearshore cages,and offshore aquaculture vessels,with detailed comparisons of volumetric versus gravimetric metering,peristaltic,hydraulic,and pneumatic conveying tailored to live,fresh-wet,and extruded pellet feeds,and centrifugal,pneumatic,and hydraulic spreading devices.Finally,the evolutionary pathway from mechanized to automated and ultimately intelligent feeding was traced,and four development trends were identified:multimodal spatiotemporal and semantic alignment of heterogeneous sensors,edge deployment of compressed large language models for explainable decision-making,UAV-swarm-based non-contact feeding supported by shipborne take-off,refueling,and battery-swap stations,and digital-twin-enabled iterative optimization of mechanistic models,providing a reference for the unmanned and intelligent development of modern marine ranching.