A multi-modal large model fish intelligent monitoring and early warning system for deep-sea net cage culture

Abstract

This invention relates to the field of intelligent monitoring technology for deep-sea aquaculture, and discloses a multimodal large-scale model intelligent monitoring and early warning system for deep-sea cage aquaculture, comprising an underwater data acquisition module, an MCP protocol communication module, a multimodal large-scale model analysis module, an edge computing processing module, and an intelligent early warning decision module. This multimodal large-scale model intelligent monitoring and early warning system for deep-sea cage aquaculture effectively solves the image acquisition challenges in turbid water and uneven lighting environments through the synergistic innovation of polarized light imaging and dark channel defogging technology, significantly improving the accuracy of fish characteristic recognition. Simultaneously, based on the dynamic background suppression function of the MCP protocol communication module, it can filter out net shadows and water flow interference in real time, ensuring the purity of data acquisition and transmission stability. The system deeply integrates visual, underwater acoustic, and water quality multi-source data, combined with YOLOv11 target detection and multimodal fusion algorithms, effectively reducing aquaculture risks such as fish hypoxia and disease transmission.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for deep-sea aquaculture, specifically to an intelligent monitoring and early warning system for multimodal large-scale fish species in deep-sea cage culture. Background Technology

[0002] Deep-sea cage aquaculture is a core direction for the modernization of marine fisheries. However, its aquaculture environment is plagued by problems such as turbid water, uneven lighting, cage obstruction, and unstable marine network signals, posing numerous challenges to fish monitoring. Existing monitoring methods mostly rely on single visual imaging or manual inspection, which are not only susceptible to underwater environmental interference leading to low accuracy in fish characteristic identification, but also suffer from low monitoring efficiency and poor real-time performance. In addition, traditional systems often use cloud-based centralized data processing, which is prone to data transmission delays in deep-sea scenarios, making it difficult to detect risks such as fish hypoxia, disease transmission, and cage damage in a timely manner, thus failing to meet the intelligent and refined monitoring needs of large-scale deep-sea aquaculture.

[0003] With the large-scale development of deep-sea aquaculture, real-time monitoring of fish growth status and the aquaculture environment has become crucial for improving the industry's quality and efficiency. Currently, most mainstream fish monitoring systems are developed based on single-modal technologies, only capable of collecting visual image data of fish schools. They cannot integrate multi-source information such as water quality and underwater sound, making it difficult to comprehensively reflect changes in fish feeding, health, and the aquatic environment. Furthermore, existing systems lack communication protocols and data processing architectures optimized for deep-sea scenarios, resulting in weak data anti-interference capabilities, high processing latency, and delayed identification of abnormal behaviors such as fish aggregation and stagnation, and sudden movements. This hinders accurate prediction and timely warning of aquaculture risks, thus restricting the intelligent development of deep-sea cage aquaculture. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a multimodal large-scale intelligent monitoring and early warning system for deep-sea cage aquaculture. It has advantages such as multi-source data fusion, low-latency processing, accurate intelligent early warning, and high scene adaptability, solving the problems of existing deep-sea fish monitoring systems, such as single modality, low monitoring accuracy, large data processing delay, untimely risk warning, and poor adaptability.

[0006] (II) Technical Solution

[0007] To achieve the aforementioned goals of multi-source data fusion, low-latency processing, accurate intelligent early warning, and high scene adaptability, this invention provides the following technical solution: a multimodal large-scale model intelligent monitoring and early warning system for deep-sea cage aquaculture, comprising an underwater data acquisition module, an MCP protocol communication module, a multimodal large-scale model analysis module, an edge computing processing module, and an intelligent early warning decision module. Each unit interacts with data at each level and forms a closed-loop control. The underwater data acquisition module is equipped with a polarized light imaging component and a dark channel defogging processing unit for clearing images of turbid water. The MCP protocol communication module integrates a dynamic background suppression unit, combining data transmission and data enhancement functions to filter out net shadows and water flow interference, and to perform image defogging and noise reduction. The multimodal large-scale model analysis module uses a YOLOv11 target detection model and a multimodal fusion algorithm to achieve full-dimensional identification of fish school status. The edge computing processing unit is deployed at nearby nodes in the aquaculture area to reduce data transmission latency. The intelligent early warning decision module uses a visual automated workflow engine to achieve real-time risk early warning and coordinated response.

[0008] The underwater data acquisition module is bidirectionally connected to the MCP protocol communication module. The MCP protocol communication module is connected to the multimodal large model analysis module and the edge computing processing module respectively. The multimodal large model analysis module is interactively connected to the edge computing processing module. The edge computing processing module is connected to the intelligent early warning decision module. The intelligent early warning decision module feeds back control signals to the underwater data acquisition module.

[0009] Preferably, the underwater data acquisition module includes an underwater camera equipped with a polarized light imaging component, an underwater acoustic sensor, and a water quality monitoring sensor. The polarized light imaging component is integrated with the dark channel defogging unit for enhancing and defogging underwater turbid images.

[0010] Preferably, the MCP protocol communication module has a built-in dynamic background suppression unit, which is used to filter out fishing net shadows and water flow interference from the visual image data transmitted by the underwater data acquisition module. At the same time, it realizes low-latency and high-stability data interaction between modules through the MCP protocol, adapting to the communication needs of complex network environments in deep sea.

[0011] Preferably, the multimodal large model analysis module integrates a YOLOv11 target detection sub-model and a multimodal fusion sub-model. The YOLOv11 target detection sub-model is used to identify the fish population size, individual size, swimming trajectory, and tail fin wagging frequency. The multimodal fusion sub-model integrates visual, underwater acoustic, and water quality data from multiple sources to analyze the fish's feeding activity, body color, abnormal scars, and density distribution characteristics.

[0012] Preferably, the edge computing processing module is deployed at a nearby node in the deep-sea aquaculture area, and includes a data preprocessing unit and a model inference acceleration unit. The data preprocessing unit normalizes and reduces noise in multi-source collected data, and the model inference acceleration unit performs localized offloading of the computing tasks of the multimodal large model analysis module to reduce data transmission latency and improve analysis and inference efficiency.

[0013] Preferably, the intelligent early warning decision module includes an early warning model library, an automated workflow unit, and an early warning output unit. The early warning model library is constructed based on data on fish population fluctuations, abnormal behavior, feeding frequency, and water quality changes, and presets early warning thresholds for oxygen deficiency risk, disease transmission, cage damage, and water quality deterioration.

[0014] Preferably, the automated workflow unit is built on a visual automated workflow engine and can trigger corresponding automated handling processes based on the analysis results of the early warning model library; the early warning output unit supports three early warning methods: audible and visual early warning, mobile terminal message push, and background system pop-up window, to realize multi-terminal synchronous push of risk information.

[0015] Preferably, the polarized light imaging component has an imaging resolution of not less than 4K and an adjustable shooting frame rate of 15-60fps. The dark channel dehazing processing unit adopts an underwater image enhancement and descattering processing algorithm based on dark channel priors to complete the underwater image dehazing and sharpening processing.

[0016] Preferably, the multimodal fusion sub-model adopts an attention mechanism fusion architecture to extract and fuse features from visual, underwater acoustic, and water quality multimodal data. The model training dataset consists of real fish monitoring data in deep-sea cage aquaculture scenarios, covering sample types of different fish species, water environments, and aquaculture states.

[0017] Preferably, the system also includes a remote cloud platform management module, which is wirelessly connected to the edge computing processing module and the intelligent early warning decision module. It is used to store historical monitoring data, display the real-time status of fish schools, enable manual intervention in the early warning decision process, and support remote debugging and updating of parameters of each module.

[0018] (III) Beneficial Effects

[0019] Compared with existing technologies, this invention provides an intelligent monitoring and early warning system for multimodal large-scale fish species in deep-sea cage culture, which has the following beneficial effects:

[0020] 1. This intelligent monitoring and early warning system for large-scale multimodal fish in deep-sea cage aquaculture effectively solves the challenges of image acquisition in turbid waters and uneven lighting environments through the synergistic innovation of polarized light imaging and dark channel defogging technology. This significantly improves the accuracy of fish feature recognition. Simultaneously, the dynamic background suppression function based on the MCP protocol communication module can filter out net shadows and water flow interference in real time, ensuring the purity and stability of data acquisition and transmission. The system deeply integrates visual, underwater acoustic, and water quality data, combined with YOLOv11 target detection and multimodal fusion algorithms, to achieve full-dimensional recognition of fish populations, swimming trajectories, feeding activity, and body surface status. This solves the problems of incomplete information and high misjudgment rate associated with traditional single-modal monitoring, providing accurate real-time monitoring capabilities for deep-sea aquaculture and effectively reducing aquaculture risks such as fish hypoxia and disease transmission.

[0021] 2. This intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage aquaculture significantly reduces reliance on unstable marine network environments through the localized deployment of edge computing processing modules. It enables rapid near-field processing of multi-source data and accelerated model inference, greatly improving the system's response speed and operational reliability in deep-sea scenarios. The intelligent early warning decision module incorporates a visual automated workflow engine, which can automatically trigger multi-terminal early warning methods such as audible and visual warnings and mobile push notifications based on preset thresholds for fish population fluctuations, abnormal behavior, and water quality changes. It also links with a remote cloud platform to achieve historical data tracing and remote parameter debugging, forming a complete closed-loop management system from data collection and analysis to risk warning and response. This system significantly improves the intelligence and automation level of deep-sea cage aquaculture, effectively solving the problems of low efficiency and delayed early warning in traditional monitoring methods, and providing solid technical support for the construction of large-scale, refined marine ranches. Attached Figure Description

[0022] Figure 1 This is a diagram showing the overall system architecture and data flow of the present invention;

[0023] Figure 2 This is a detailed flowchart of the data acquisition and communication module of the present invention;

[0024] Figure 3 This is a flowchart of the multimodal analysis and edge computing processing of the present invention;

[0025] Figure 4 This is a flowchart of the intelligent early warning and feedback control process of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figure 1-4 A multimodal large-scale model intelligent monitoring and early warning system for deep-sea cage aquaculture includes an underwater data acquisition module, an MCP protocol communication module, a multimodal large-scale model analysis module, an edge computing processing module, and an intelligent early warning decision module. Each unit interacts with data at each level and forms a closed-loop control. The underwater data acquisition module is equipped with a polarized light imaging component and a dark channel defogging processing unit for clearing images of turbid water. The MCP protocol communication module integrates a dynamic background suppression unit, which has both data transmission and data enhancement functions, and achieves filtering of fishing net shadows and water flow interference, as well as image defogging and noise reduction. The multimodal large-scale model analysis module adopts the YOLOv11 target detection model and multimodal fusion algorithm to achieve full-dimensional identification of fish school status. The edge computing processing unit is deployed at the nearest node in the aquaculture area to reduce data transmission latency. The intelligent early warning decision module realizes real-time risk early warning and linkage response based on a visual automated workflow engine.

[0028] The underwater data acquisition module is bidirectionally connected to the MCP protocol communication module. The MCP protocol communication module is connected to the multimodal large model analysis module and the edge computing processing module respectively. The multimodal large model analysis module and the edge computing processing module are interactively connected. The edge computing processing module and the intelligent early warning decision module are connected to each other. The intelligent early warning decision module feeds back control signals to the underwater data acquisition module.

[0029] In the case implementation, the underwater data acquisition module includes an underwater camera equipped with a polarized light imaging component, an underwater acoustic sensor, and a water quality monitoring sensor. The polarized light imaging component is integrated with the dark channel defogging unit to enhance and defog underwater turbid images.

[0030] Among them, the polarization imaging component adopts a sealed and pressure-resistant structure with a protection level of no less than IP68, which can adapt to water depth environments of 0–50 meters. It can collect data synchronously from multiple sensors with a timestamp error of less than 10ms.

[0031] By integrating imaging and defogging processing, the imaging blurring problem caused by turbid water and uneven lighting in the deep sea is effectively overcome, and high-quality raw visual data of fish schools are obtained.

[0032] In the case implementation, the MCP protocol communication module has a built-in dynamic background suppression unit, which is used to filter out fishing net shadows and water flow interference from the visual image data transmitted by the underwater data acquisition module. At the same time, it realizes low-latency and high-stability data interaction between modules through the MCP protocol, adapting to the communication needs of complex network environments in the deep sea.

[0033] Among them, the dynamic background suppression unit adopts an algorithm that combines frame difference method and Gaussian mixture model, which can adaptively identify static backgrounds such as fishing nets and reefs and dynamic interferences such as water flow and bubbles, and retain only the effective dynamic feature data of fish. The transmission bandwidth of the MCP protocol communication module is adjustable from 1 to 100 Mbps and supports the function of resuming interrupted transmission.

[0034] By using dynamic background suppression preprocessing and the low-latency transmission of the MCP protocol, invalid data interference is effectively filtered out, improving the efficiency and stability of data transmission, and perfectly adapting to the communication needs of deep-sea networks with unstable signals and complex electromagnetic environments.

[0035] In the case implementation, the multimodal large model analysis module integrates the YOLOv11 target detection sub-model and the multimodal fusion sub-model. The YOLOv11 target detection sub-model is used to identify the fish population size, individual size, swimming trajectory and tail fin wagging frequency; the multimodal fusion sub-model integrates visual, underwater acoustic and water quality multi-source data to analyze the fish feeding activity, body color, abnormal scars and density distribution characteristics.

[0036] Among them, the YOLOv11 model optimizes the anchor box and feature extraction structure for fish swarm morphology, and the multimodal fusion sub-model performs feature alignment and weight allocation for multi-source data.

[0037] By using YOLOv11 for precise detection and multimodal fusion analysis, we can achieve full-dimensional and high-accuracy identification of fish school status.

[0038] In the case implementation, the edge computing processing module was deployed at a nearby node in the deep-sea aquaculture area. It includes a data preprocessing unit and a model inference acceleration unit. The data preprocessing unit normalizes and reduces noise from multi-source collected data, while the model inference acceleration unit performs localized offloading of the computing tasks of the multimodal large model analysis module, reducing data transmission latency and improving analysis and inference efficiency.

[0039] Among them, the data preprocessing unit uses Z-Score normalization algorithm and wavelet denoising algorithm to standardize and denoise multi-source heterogeneous data, and the model inference acceleration unit is equipped with embedded computing chip to support lightweight model deployment and parallel computing.

[0040] By offloading localized data processing and model inference at nearby nodes in the aquaculture area, the amount of data transmitted to the cloud is significantly reduced, and the data processing latency is controlled at the second level, ensuring the real-time nature of fish abnormality analysis and solving the industry pain point of data transmission latency in deep-sea areas.

[0041] In the case implementation, the intelligent early warning decision module includes an early warning model library, an automated workflow unit, and an early warning output unit. The early warning model library is built based on data on fish population fluctuations, abnormal behavior, feeding frequency, and water quality changes, and presets early warning thresholds for oxygen deficiency risk, disease transmission, cage damage, and water quality deterioration.

[0042] Among them, the early warning model library is trained and constructed based on the random forest algorithm, which integrates historical monitoring data of fish populations and water quality and aquaculture risk cases. Each early warning threshold can be adjusted according to the aquaculture species and aquaculture stage, and supports automatic iterative optimization of the threshold.

[0043] By constructing an early warning model library based on multi-dimensional data and setting preset graded early warning thresholds, the system enables precise quantitative judgment of aquaculture risks such as hypoxia and disease, upgrading risk early warning from experience-based judgment to data-driven judgment, and improving the scientific nature and accuracy of early warning.

[0044] In the case implementation, the automated workflow unit is built on a visual automated workflow engine, which can trigger corresponding automated handling processes based on the analysis results of the early warning model library; the early warning output unit supports three early warning methods: audible and visual warnings, mobile message push, and background system pop-ups, to achieve multi-terminal synchronous push of risk information.

[0045] Among them, the automated workflow supports visual orchestration and custom processing logic, and the early warning information is transmitted in encrypted form to ensure real-time performance and security;

[0046] By automating processes and simultaneously pushing information across multiple terminals, rapid risk response and unattended intelligent handling are achieved, reducing losses in aquaculture.

[0047] In the case implementation, the polarized light imaging component has an imaging resolution of no less than 4K and an adjustable shooting frame rate of 15-60fps. The dark channel dehazing processing unit adopts an underwater image enhancement and descattering processing algorithm based on dark channel priors to complete the underwater image dehazing and sharpening processing.

[0048] Among them, the polarization angle is continuously adjustable from 0 to 360°, and the parameters can be adaptively matched according to the turbidity of the water. The algorithm processing is synchronized with the camera shooting frame rate.

[0049] By using high-resolution imaging and underwater-specific defogging algorithms, the clarity and recognizability of fish images in turbid waters are significantly improved.

[0050] In the implementation of the case, the multimodal fusion sub-model adopts an attention mechanism fusion architecture to extract and fuse features from visual, underwater acoustic, and water quality multimodal data. The model training dataset consists of real fish monitoring data in deep-sea cage aquaculture scenarios, covering sample types of different fish species, water environments, and aquaculture states.

[0051] The architecture includes a three-layer structure of feature extraction, feature fusion and state classification, and supports online incremental learning and model iterative updates.

[0052] By deeply fusing multimodal features and training with real-world scene data, the model's generalization ability and recognition accuracy are improved.

[0053] In the implementation of the case, the system also includes a remote cloud platform management module. The remote cloud platform management module is wirelessly connected to the edge computing processing module and the intelligent early warning decision module. It is used to store historical monitoring data, display the real-time status of fish, enable manual intervention in the early warning decision process, and support remote debugging and updating of parameters of each module.

[0054] Among them, the remote cloud platform management module adopts a B / S architecture, supports seamless access from multiple terminals such as computers, mobile devices, and tablets, and uses a distributed storage method for historical monitoring data. The data retention time is no less than 3 years, and it supports data export and visualization analysis.

[0055] Through centralized management and wireless data connection via a remote cloud platform, the system enables cloud storage, real-time display, and manual intervention of aquaculture data. Staff can remotely debug the parameters of each module and correct early warning decisions, forming a dual control mode of "local automation + cloud remote control," which improves the system's ease of operation and management flexibility.

[0056] When implementing this procedure, please follow these steps:

[0057] 1) First, deploy each module to the corresponding location in the aquaculture area, start the underwater data acquisition module to synchronously collect visual, underwater sound and water quality data of the fish and complete the preliminary preprocessing;

[0058] 2) Then, the interference data is filtered out through the MCP protocol communication module, and the effective data is transmitted to the edge computing processing module for normalization and noise reduction.

[0059] 3) The multimodal large model analysis module then analyzes the fish swarm status, and the edge computing module performs local inference before transmitting the results to the intelligent early warning decision module;

[0060] 4) Finally, the early warning module compares the risk with the threshold, triggers the automated handling process and pushes the warning to multiple terminals, while synchronizing the data to the remote cloud platform.

[0061] In summary, this intelligent monitoring and early warning system for large-scale multimodal fish in deep-sea cage aquaculture effectively solves the challenges of image acquisition in turbid waters and under uneven lighting conditions through the synergistic innovation of polarized light imaging and dark channel defogging technology. This significantly improves the accuracy of fish feature recognition. Furthermore, the dynamic background suppression function based on the MCP protocol communication module can filter out net shadows and water flow interference in real time, ensuring the purity and stability of data acquisition and transmission. The system deeply integrates visual, underwater acoustic, and water quality data, combined with YOLOv11 target detection and multimodal fusion algorithms, to achieve comprehensive identification of fish populations, swimming trajectories, feeding activity, and body surface status. This solves the problems of incomplete information and high misjudgment rate associated with traditional single-modal monitoring, providing precise real-time monitoring capabilities for deep-sea aquaculture and effectively reducing aquaculture risks such as fish hypoxia and disease transmission.

[0062] Furthermore, the localized deployment of the edge computing processing module significantly reduces reliance on the unstable network environment at sea, enabling near-field rapid processing of multi-source collected data and accelerated model inference. This greatly improves the system's response speed and operational reliability in deep-sea scenarios. The intelligent early warning decision module has a built-in visual automated workflow engine that can automatically trigger multi-terminal early warning methods such as audible and visual warnings and mobile push notifications based on preset thresholds such as fish population fluctuations, abnormal behavior, and water quality changes. It also links with a remote cloud platform to achieve historical data tracing and remote parameter debugging, forming a complete closed-loop management system from data collection, analysis and processing to risk warning and disposal. This system significantly improves the intelligence and automation level of deep-sea cage aquaculture, effectively solving the problems of low efficiency and delayed early warning in traditional monitoring methods, and providing solid technical support for the construction of large-scale and refined marine ranches.

[0063] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multimodal large-scale model intelligent monitoring and early warning system for deep-sea cage aquaculture, comprising an underwater data acquisition module, an MCP protocol communication module, a multimodal large-scale model analysis module, an edge computing processing module, and an intelligent early warning decision-making module, characterized in that: The units interact with each other at each level and form a closed-loop control. The underwater data acquisition module is equipped with a polarized light imaging component and a dark channel defogging processing unit for clearing images of turbid water. The MCP protocol communication module integrates a dynamic background suppression unit, which has both data transmission and data enhancement functions, and realizes the filtering of fishing net shadows and water flow interference, as well as image defogging and noise reduction. The multimodal large model analysis module adopts the YOLOv11 target detection model and multimodal fusion algorithm to realize full-dimensional identification of fish school status. The edge computing processing unit is deployed at the nearest node in the aquaculture area to reduce data transmission latency. The intelligent early warning decision module realizes real-time risk early warning and linkage response based on a visual automated workflow engine. The underwater data acquisition module is bidirectionally connected to the MCP protocol communication module. The MCP protocol communication module is connected to the multimodal large model analysis module and the edge computing processing module respectively. The multimodal large model analysis module is interactively connected to the edge computing processing module. The edge computing processing module is connected to the intelligent early warning decision module. The intelligent early warning decision module feeds back control signals to the underwater data acquisition module.

2. The intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage aquaculture according to claim 1, characterized in that: The underwater data acquisition module includes an underwater camera equipped with a polarized light imaging component, an underwater acoustic sensor, and a water quality monitoring sensor. The polarized light imaging component is integrated with the dark channel defogging unit for enhancing and defogging underwater turbid images.

3. The intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage aquaculture according to claim 1, characterized in that: The MCP protocol communication module has a built-in dynamic background suppression unit, which is used to filter out fishing net shadows and water flow interference from the visual image data transmitted by the underwater data acquisition module. At the same time, it realizes low-latency and high-stability data interaction between modules through the MCP protocol, adapting to the communication needs of complex network environments in the deep sea.

4. The intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage aquaculture according to claim 1, characterized in that: The multimodal large model analysis module integrates the YOLOv11 target detection sub-model and the multimodal fusion sub-model. The YOLOv11 target detection sub-model is used to identify the fish population size, individual size, swimming trajectory, and tail fin wagging frequency. The multimodal fusion sub-model integrates visual, underwater acoustic, and water quality data from multiple sources to analyze the fish's feeding activity, body color, abnormal scars, and density distribution characteristics.

5. The intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage culture according to claim 1, characterized in that: The edge computing processing module is deployed at a nearby node in the deep-sea aquaculture area and includes a data preprocessing unit and a model inference acceleration unit. The data preprocessing unit normalizes and reduces noise in multi-source collected data, and the model inference acceleration unit performs localized offloading of the computing tasks of the multimodal large model analysis module to reduce data transmission latency and improve analysis and inference efficiency.

6. The intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage aquaculture according to claim 1, characterized in that: The intelligent early warning decision-making module includes an early warning model library, an automated workflow unit, and an early warning output unit. The early warning model library is constructed based on data on fish population fluctuations, abnormal behavior, feeding frequency, and water quality changes, and presets early warning thresholds for oxygen deficiency risk, disease transmission, cage damage, and water quality deterioration.

7. The intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage culture according to claim 1, characterized in that: The automated workflow unit is built on a visual automated workflow engine and can trigger corresponding automated handling processes based on the analysis results of the early warning model library. The early warning output unit supports three early warning methods: audible and visual warning, mobile message push, and background system pop-up, enabling multi-terminal synchronous push of risk information.

8. The intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage culture according to claim 1, characterized in that: The polarization imaging component has an imaging resolution of no less than 4K and an adjustable shooting frame rate of 15-60fps. The dark channel dehazing processing unit adopts an underwater image enhancement and descattering processing algorithm based on dark channel priors to complete the underwater image dehazing and sharpening processing.

9. The intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage aquaculture according to claim 1, characterized in that: The multimodal fusion sub-model adopts an attention mechanism fusion architecture to extract and fuse features from visual, underwater acoustic, and water quality multimodal data. The model training dataset consists of real fish monitoring data from deep-sea cage aquaculture scenarios, covering sample types of different fish species, water environments, and aquaculture states.

10. The intelligent monitoring and early warning system for multimodal large-scale fish in deep-sea cage culture according to claim 1, characterized in that: The system also includes a remote cloud platform management module, which is wirelessly connected to the edge computing processing module and the intelligent early warning decision module. It is used to store historical monitoring data, display the real-time status of fish schools, enable manual intervention in the early warning decision process, and support remote debugging and updating of parameters of each module.

Images