现代化海洋牧场智能投喂技术与装备发展综述

Recent Advances and Development Trends in Intelligent Feeding Technologies and Equipment for Modern Marine Ranches

摘要:

深远海养殖对精准投喂的迫切需求与复杂海况下感知受限、作业困难之间的矛盾,驱动投喂系统向智能化转型。本文界定了现代化海洋牧场智能投喂系统的内涵,沿“感知—决策—执行”闭环链路梳理了各环节的研究进展和瓶颈:感知层面,综述了养殖环境多参数监测以及基于视觉与声学融合的生物量估算和摄食强度量化方法;决策层面,对比了生长曲线、生物能量等机理模型与模糊控制、深度学习驱动的动态决策模型;控制与执行层面,阐述了“云-边-端”控制架构及气水联合输送、精准计量与撒布装置的优化策略。分析探讨了现代化海洋牧场投喂系统的发展路径,多模态时空对齐融合、大语言模型边缘部署及无人机集群投喂等发展趋势,旨在为海洋牧场智能化、无人化发展提供参考。

Abstract:

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.