Fishing ground management system and method based on lightweight anomaly detection and intelligent agent collaborative decision-making

Abstract

The invention discloses a fishery management system and method based on lightweight anomaly detection and intelligent agent collaborative decision, and belongs to the technical field of fishery management, and the system comprises a visual information collection and processing module which is used for obtaining and processing the visual data of a breeding area; the lightweight anomaly detection module is used for periodically scanning the breeding area and identifying and classifying potential anomaly conditions in real time; the intelligent decision agent module is used for starting an inference engine when an abnormal condition is detected, and generating and executing a corresponding decision instruction; and the message notification Internet of Things communication module is used for realizing real-time transmission of information and equipment control. According to the invention, the problems of high energy consumption, low efficiency and difficulty in realizing intelligent monitoring in a resource-limited scene of existing fishery management are solved. According to the invention, by constructing a cascaded technical architecture of low-consumption inspection, accurate triggering and dynamic decision, efficient abnormal response can be realized in a resource sensitive scene, the computing power consumption is reduced, and the high-precision abnormal recognition rate is ensured.

Description

Technical Field

[0001] This invention relates to the field of fishery management technology, specifically to a fishery management system and method based on lightweight anomaly detection and intelligent agent collaborative decision-making. Background Technology

[0002] As the aquaculture industry accelerates its transformation towards intensification and intelligentization, traditional models relying on manual inspections and single-point sensor monitoring are no longer sufficient to meet the refined management needs of modern fish farms. The industry generally faces systemic challenges such as weak multi-dimensional sensing capabilities, delayed anomaly response, and low equipment collaboration efficiency. Especially in complex aquatic environments, the combined effects of varying lighting, equipment reflections, and dynamic interference from biological communities significantly reduce the stability of existing visual inspection systems. Meanwhile, large-scale aquaculture scenarios are highly sensitive to energy costs; how to optimize the allocation of computing resources while ensuring detection accuracy has become a key challenge restricting the practical application of this technology.

[0003] Traditional fish farm monitoring technologies have the following limitations: 1) Uncontrolled development and maintenance costs: Each type of anomaly detection requires the independent construction of a dedicated model (such as fish disease identification, feeding machine malfunction, sudden drop in dissolved oxygen, etc.), resulting in repetitive investment in the entire process of data collection, labeling, and training. Taking a typical fish farm as an example, maintaining 20 independent models consumes more than 3,000 hours of manpower per year, and compatibility conflicts are prone to occur during version iterations; 2) Defects in detection robustness under dynamic environments: Water surface reflections form high light noise on sunny afternoons, leading to a high misjudgment rate of fish density by traditional algorithms. The occlusion effect of aerator bubbles on underwater cameras causes local area detection failures, with a single frame image information loss rate reaching 40%. The generalization performance of visible light models on nighttime infrared images is also limited. The cost of the system is reduced by 2.8 times, requiring additional deployment of thermal imaging analysis modules. The false negative rate of micron-level lesions (such as fish gill parasites) in turbid water under low light conditions soars to 58%. 3) There is a serious imbalance in resource efficiency. Under the continuous operation mode of large models, the daily effective utilization rate of edge devices is less than 8%, and 90% of computing power is consumed in idle waiting during periods without abnormalities. 4) There is a double dilemma of economic feasibility. Traditional solutions have high procurement costs for a single set of equipment, and the investment payback period is as long as 3-5 years. The lack of a professional technical team leads to a mean time to repair (MTTR) of 16 hours for system failures. The full network synchronization time during model updates exceeds 72 hours. Parallel inference of multiple models causes resource competition, and the instruction delay during peak periods exceeds 8 seconds, missing the critical handling window.

[0004] These long-standing unresolved problems have kept fisheries production trapped in a cycle of "high input, low output." On the one hand, over-reliance on human experience leads to insufficient scientific decision-making, resulting in persistently high economic losses due to misjudgments of fish behavior. On the other hand, resource utilization efficiency continues to decline; the energy cost per ton of aquatic product under traditional models is significantly higher than that of intelligent aquaculture, and the cost of water pollution control continues to rise due to a lack of precise monitoring. More seriously, ecological and environmental pressures are intensifying, with frequent incidents of aquatic ecosystem damage caused by excessive discharge of aquaculture wastewater. Against this backdrop, developing intelligent and integrated fish farm monitoring and management technologies has become an essential path to break free from the constraints of traditional models and promote the green and low-carbon transformation of fisheries. This is not only about improving the economic benefits of the industry but also a crucial guarantee for achieving sustainable development in fisheries. Summary of the Invention

[0005] The purpose of this invention is to provide a fish farm management system and method based on lightweight anomaly detection and intelligent agent collaborative decision-making. By constructing a cascaded technical architecture of low-consumption inspection, precise triggering and dynamic decision-making, it achieves efficient anomaly response in resource-sensitive scenarios, significantly reduces computing power consumption, and ensures a high accuracy of anomaly identification, thus solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making includes:

[0008] The visual information acquisition and processing module is used to monitor the fish behavior, water quality environment and equipment operation status in the aquaculture area in real time using a high-resolution camera, acquire visual data of the aquaculture area, and perform cropping, scaling, rotation, flipping and color transformation processing on the acquired visual data of the aquaculture area, remove noise and filter the data.

[0009] The lightweight anomaly detection module uses a zero-shot anomaly detection model based on AnomalyCLIP to periodically scan the aquaculture area according to an adaptive frequency adjustment scheme, maintaining high-precision monitoring in resource-constrained environments and identifying and classifying potential anomalies in real time.

[0010] The intelligent decision agent module is used to start the inference engine when an abnormal situation is detected, comprehensively analyze the current environmental state, generate and execute corresponding decision instructions. Among them, when abnormal fish behavior and water quality are detected, the aerator is activated first to improve water quality conditions.

[0011] The message notification IoT communication module is used to realize real-time information transmission and device control using IoT communication modules based on LoRaWAN technology, enabling stable data interaction between the aquaculture area and the fish farm management center.

[0012] Preferably, the visual information acquisition and processing module includes:

[0013] High-resolution cameras are deployed at key locations in aquaculture areas to capture real-time fish behavior, water quality, and equipment operating status, thereby determining visual data of the aquaculture area.

[0014] The video frame preprocessing unit is used to crop and scale the visual data of the aquaculture area, extract key areas, perform color correction to compensate for color deviation caused by underwater light attenuation, and timestamp the visual data of the aquaculture area so that each frame corresponds to its acquisition time.

[0015] The data augmentation unit is used to enhance the visual data of the breeding area by randomly rotating and flipping it to simulate different perspectives and directions, and to perform color transformation, adjust brightness, contrast and saturation to simulate different lighting conditions.

[0016] Preferably, the visual information acquisition and processing module further includes:

[0017] The hardware energy-saving optimization unit is used to reduce the energy consumption of the visual information acquisition and processing module;

[0018] Among them, a dynamic frame rate adjustment scheme was formulated, and the sampling frequency was set according to the day and night cycle. During the daytime period from 6:00 to 18:00, the frame rate was 15FPS, and during the nighttime period from 18:00 to 6:00, the frame rate was 5FPS. A low-power high-resolution camera chip was adopted, with a single device standby power consumption of ≤ 0.5W and working power consumption of ≤ 1.5W.

[0019] The data storage and transmission unit is used to cache the raw data of the past 7 days using a high-speed SSD, compress the video stream based on H.265 encoding, and upload it to the cloud via AES encryption protocol. It automatically saves the untransmitted data when the network is interrupted and continues uploading after the connection is restored.

[0020] Preferably, the lightweight anomaly detection module includes:

[0021] The periodic anomaly detection unit uses an improved AnomalyCLIP architecture to scan the aquaculture area every Δt minutes and calculates the anomaly confidence level Panomaly by comparing it with a preset normal state feature vector.

[0022] The trigger-based agent activation unit is used to send a trigger signal containing the coordinates and timestamp of the abnormal area to the smart agent when the anomaly confidence level Panomaly ≥ a set threshold θ.

[0023] The inspection frequency dynamic adjustment unit is used to automatically set the inspection interval according to the day and night cycle. During the daytime period from 6:00 to 18:00, the inspection interval Δt = 5 minutes, and during the nighttime period from 18:00 to 6:00, the inspection interval Δt = 15 minutes.

[0024] Anomaly confidence calculation unit, used to calculate anomaly confidence and assess the degree of anomaly;

[0025] The hardware power-saving unit is used to save power by using a dedicated NPU chip in the Raspberry Pi 5, with a single detection power consumption of ≤1.2W.

[0026] Preferably, the threshold θ of the triggered agent activation unit is set as follows:

[0027] Base threshold When the confidence level is abnormal At that time, a regular proxy response is triggered;

[0028] Emergency threshold When the confidence level is abnormal When this occurs, the following enhanced responses will be activated:

[0029] The detection interval is temporarily shortened to Δt=1 minute. When an anomaly is detected, an alert SMS is sent to the administrator's mobile phone via Alibaba Cloud SMS service.

[0030] Preferably, the intelligent decision-making agent module includes:

[0031] The visual description generation unit is used to perform pixel-level segmentation of abnormal region images, extract visual elements of color distribution and shape features, and generate structured descriptions.

[0032] The equipment control instruction generation unit is used to generate equipment control instructions. It has a built-in instruction template library and controls the aerator and feeder.

[0033] The security verification unit is used to check whether the current state of the device allows operation before sending device commands, and to verify whether the commands comply with security specifications.

[0034] Preferably, the message notification IoT communication module includes:

[0035] Low-power wide-area network communication unit, used for long-distance, low-power data interaction based on LoRaWAN protocol wireless transmission technology;

[0036] It consists of an ESP32 main control chip and an SX1278 RF module, and supports multiple spreading factors and bandwidth settings to adapt to different communication distance and speed requirements.

[0037] Using a star topology, a LoRa gateway deployed at the center of the aquaculture area centrally receives data from each monitoring node and forwards it to a local server or cloud platform.

[0038] Data is encapsulated using JSON encoding format, including timestamps, device IDs, data types, and encrypted fields;

[0039] The TLS / SSL encryption protocol is used during data transmission to prevent data leakage and unauthorized access. At the same time, a unique identifier is assigned to each device for access authentication and access control.

[0040] Preferably, the message notification IoT communication module further includes:

[0041] The remote command receiving and execution unit is used to receive and parse operation commands from the fishery management center;

[0042] Supports lightweight command delivery via the MQTT protocol, enabling stable command reception in weak network environments;

[0043] The instruction types cover equipment start / stop control and parameter adjustment, and the received instructions are verified for origin and integrity to prevent malicious attacks or misoperations.

[0044] Preferably, the message notification IoT communication module further includes:

[0045] The multi-channel anomaly alarm notification unit is used to proactively send alarm information to administrators via email notification and SMS push service after an anomaly event is detected.

[0046] Among them, an alarm email is sent to a preset email address based on the SMTP protocol. The content includes the anomaly type, the time of occurrence, the location coordinates, and the preliminary analysis results.

[0047] Integrate Alibaba Cloud SMS service API to call telecom operator resources via HTTP interface and send concise and clear SMS notifications to specified mobile phone numbers;

[0048] Develop a multi-contact configuration and resend scheme to automatically attempt to resend the notification if the first notification fails, ensuring that critical information is not lost, and allow users to customize the notification content format according to different levels of anomalies;

[0049] The communication status monitoring and optimization unit is used to monitor the stability of the communication link and the operating status of the equipment in real time.

[0050] During signal strength detection, the received signal strength indication value of the LoRa module is periodically reported to assist in evaluating communication quality. During power monitoring, the current battery voltage and remaining power of the device are collected. When the power level is lower than the set threshold, a low power warning notification is triggered. During network interruption, data to be sent is buffered and retransmitted in priority order after the connection is restored. The module's working mode is dynamically adjusted according to the communication load.

[0051] According to another aspect of the present invention, a fish farm management method based on lightweight anomaly detection and intelligent agent collaborative decision-making is provided, implemented based on the fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making as described above, comprising:

[0052] Step 1: Use high-resolution cameras to monitor the status of the fish farm in real time, continuously capture images and video data of the fish farm, and perform preliminary processing on the captured visual data of the aquaculture area. The processed visual data of the aquaculture area is then transmitted to the anomaly detection system via the network.

[0053] Step 2: The anomaly detection system performs in-depth analysis of visual data from the aquaculture area, identifies various anomalies in the images, including abnormal fish behavior, water quality changes, and site anomalies, and generates an anomaly analysis map based on the analysis results, which intuitively displays the specific location and type of the anomaly.

[0054] Step 3: Based on the anomaly analysis chart and the current state of the fish farm, take corresponding measures, automatically control the IoT devices, adjust the water temperature and the operating status of the aerator to deal with the abnormal situation, and at the same time, notify the fish farm management personnel of the abnormal situation for timely intervention and handling.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] 1. This invention combines the feature representation capabilities of text encoders and image encoders to perform high-precision anomaly detection on input video frames. By using similarity calculation methods, it can not only identify image-level anomalies but also accurately locate specific anomaly regions, thereby achieving effective monitoring of anomalies in complex scenes. The introduction of subject focus loss makes the model more focused on abnormal changes of the subject during training, reducing the influence of background factors on the anomaly detection results and greatly improving the accuracy and practicality of anomaly detection, especially in application scenarios that require precise monitoring of the health status of specific objects.

[0057] 2. The implementation of the InternVedio2.5 module of this invention provides powerful real-time tracking capabilities, which can continuously monitor and record changes in abnormal areas. The intelligent feedback mechanism generated based on the tracking results can guide immediate operation or adjustment of IoT device settings, enhancing the system's response speed and automation level. With the help of IoT technology, it can be seamlessly integrated with various sensors and actuators in the breeding site to realize real-time data acquisition, status monitoring and automatic control, which greatly improves management efficiency and decision support capabilities.

[0058] 3. This invention adopts a modular design approach, which not only facilitates the integration of existing advanced technologies but also supports the seamless integration of future new technologies. This flexibility provides a guarantee for responding to ever-changing practical needs and promotes the continuous optimization and development of the system. By comprehensively recording and analyzing abnormal events, it provides managers with detailed reports and intuitive visualization tools, which helps to quickly understand the problem and make scientific and reasonable decisions. This undoubtedly improves the overall efficiency and service quality of fish farm management. At the same time, the integration of IoT devices allows for remote monitoring and management, enabling managers to obtain the latest aquaculture information no matter where they are, further enhancing the timeliness and accuracy of management decisions. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the fishery management method based on lightweight anomaly detection and intelligent agent collaborative decision-making according to the present invention;

[0060] Figure 2 The flowchart of the fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making of the present invention is shown below.

[0061] Figure 3 This is a schematic diagram illustrating the principle of the anomaly detection model provided by the present invention;

[0062] Figure 4 This is a schematic diagram illustrating the interaction between the anomaly model and Agent information provided by the present invention;

[0063] Figure 5 This is a schematic diagram illustrating the interaction between the Agent provided by this invention and an IoT device;

[0064] Figure 6 This is a schematic diagram of the network structure of the Internet of Things system provided by the present invention. Detailed Implementation

[0065] 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.

[0066] To address the challenges of achieving intelligent monitoring in existing fish farm management scenarios characterized by high energy consumption, low efficiency, and limited resources, please refer to [link / reference needed]. Figures 1-6 This embodiment provides the following technical solution:

[0067] The fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making includes: a visual information acquisition and processing module, a lightweight anomaly detection module, an intelligent decision-making agent module, and a message notification IoT communication module.

[0068] The visual information acquisition and processing module is used to monitor the fish behavior, water quality and equipment operation status in the aquaculture area in real time using a high-resolution camera, acquire visual data of the aquaculture area, and perform cropping, scaling, rotation, flipping and color transformation processing on the acquired visual data of the aquaculture area to remove noise and filter the data.

[0069] In this embodiment, the visual information acquisition and processing module includes:

[0070] High-resolution cameras are deployed at key locations in aquaculture areas to capture real-time fish behavior, water quality, and equipment operating status, thereby determining visual data of the aquaculture area.

[0071] The video frame preprocessing unit is used to crop and scale the visual data of the aquaculture area, extract key areas to reduce the amount of computation, perform color correction to compensate for color deviation caused by underwater light attenuation, and timestamp the visual data of the aquaculture area so that each frame corresponds to its acquisition time.

[0072] The data augmentation unit is used to enhance the visual data of the breeding area by randomly rotating and flipping it to simulate different perspectives and directions, and performing color transformation, adjusting brightness, contrast and saturation to simulate different lighting conditions to adapt to various lighting conditions. It also injects noise to add appropriate random noise to improve the robustness of the model.

[0073] Specifically, in the fish farm management system, information collection is a key link to ensure that the system can accurately identify and respond to abnormal situations. Therefore, high-resolution cameras are used as the main information collection devices to monitor various conditions in the aquaculture area in real time. Specifically, high-resolution cameras are responsible for capturing the following key information: 1) Fish behavior monitoring: The high-resolution cameras continuously capture the behavior patterns of the fish, including but not limited to swimming speed, aggregation density, and activity range. These behavioral characteristics are important indicators for judging the health status of the fish; 2) Water quality environment observation: The underwater high-resolution cameras monitor changes in water quality conditions, such as turbidity and color changes, indirectly reflecting the changing trends of water quality parameters (such as oxygen content and pH value).

[0074] To improve the training effectiveness and robustness of the model, the original visual data from the aquaculture area needs to undergo a series of data augmentation processes. Specific data augmentation methods include, but are not limited to: 1) Cropping and scaling: Randomly cropping different parts of the image and adjusting their size to increase sample diversity and help the model better adapt to target detection tasks at different scales; 2) Rotation and flipping: Rotating the image at random angles and flipping it horizontally or vertically to simulate different perspectives and directions, enabling the model to learn a wider range of feature representations; 3) Color transformation: Changing the color attributes of the image, such as brightness, contrast, and saturation, to simulate different lighting conditions and enhance the model's recognition ability under various lighting environments; 4) Noise addition: Adding appropriate random noise to the image to simulate imperfections in the real world, thereby improving the model's anti-interference ability.

[0075] In this embodiment, the visual information acquisition and processing module further includes:

[0076] The hardware energy-saving optimization unit is used to reduce the energy consumption of the visual information acquisition and processing module;

[0077] Among them, a dynamic frame rate adjustment scheme was formulated, and the sampling frequency was set according to the day and night cycle. During the daytime period from 6:00 to 18:00, the frame rate was 15FPS, and during the nighttime period from 18:00 to 6:00, the frame rate was 5FPS. A low-power high-resolution camera chip was adopted, with a single device standby power consumption of ≤ 0.5W and working power consumption of ≤ 1.5W.

[0078] The data storage and transmission unit is used to cache the raw data of the past 7 days using a high-speed SSD (Solid State Disk), compress the video stream based on H.265 encoding, and upload it to the cloud via AES encryption protocol. It automatically saves the untransmitted data when the network is interrupted and continues uploading after the connection is restored.

[0079] The lightweight anomaly detection module employs a zero-shot anomaly detection model based on AnomalyCLIP to periodically scan the aquaculture area according to an adaptive frequency adjustment scheme. It dynamically adjusts the scanning frequency based on the environment to reduce invalid inspections. It extracts microscopic anomaly features through a multi-scale feature fusion layer and generates anomaly confidence scores. When the anomaly confidence score reaches a preset threshold, it triggers subsequent processing procedures.

[0080] In this embodiment, the lightweight anomaly detection module includes:

[0081] The periodic anomaly detection unit uses an improved AnomalyCLIP architecture to scan the aquaculture area every Δt minutes and calculates the anomaly confidence level Panomaly by comparing it with a preset normal state feature vector.

[0082] The trigger-based agent activation unit is used to send a trigger signal containing the coordinates and timestamp of the abnormal area to the smart agent when the anomaly confidence level Panomaly ≥ a set threshold θ.

[0083] The inspection frequency dynamic adjustment unit is used to automatically set the inspection interval according to the day and night cycle. During the daytime period from 6:00 to 18:00, the inspection interval Δt = 5 minutes, and during the nighttime period from 18:00 to 6:00, the inspection interval Δt = 15 minutes.

[0084] Anomaly confidence calculation unit, used to calculate anomaly confidence and assess the degree of anomaly;

[0085] The degree of abnormality is assessed using the following formula:

[0086] ("Normal fishing ground conditions")

[0087] Where It represents the currently captured image;

[0088] The hardware power-saving unit is used to save power by using a dedicated NPU chip in the Raspberry Pi 5, with a single detection power consumption of ≤1.2W.

[0089] In this embodiment, the threshold θ of the triggered agent activation unit is set as follows:

[0090] Base threshold When the confidence level is abnormal At that time, a regular proxy response is triggered;

[0091] Emergency threshold When the confidence level is abnormal When this occurs, the following enhanced responses will be activated:

[0092] The detection interval is temporarily shortened to Δt=1 minute. When an anomaly is detected, an alert SMS is sent to the administrator's mobile phone via Alibaba Cloud SMS service.

[0093] Specifically, in the fish farm management system, the anomaly detection model is one of the core components, used to identify and classify potential anomalies in real time. It adopts a zero-shot anomaly detection model based on AnomalyCLIP, which significantly improves the system's generalization performance and adaptability by leveraging its ability to identify unknown anomalies without additional labeled data. In order to adapt to the specific needs of fish farm management, the AnomalyCLIP model has undergone the following optimization steps: 1) Domain adaptation: Although AnomalyCLIP is a pre-trained model, in order to improve its performance in the fishery scenario, it is fine-tuned with a small number of labeled normal samples to make it more in line with the actual application environment; 2) Robustness enhancement: By introducing adversarial training and noise injection strategies, the model's detection ability under complex lighting conditions or blurred images is further improved; 3) Addition of subject focus loss: In order to further enhance the model's ability to detect anomalies of specific subjects in fish farm management, a subject focus loss function is designed and optimized specifically for key objects in the fish farm, so that the model can more accurately identify whether these objects have abnormal behavior or state.

[0094] like Figure 2 As shown, high-resolution cameras are deployed at key locations in the fish farm to capture images or video data of the fish farm environment in real time. In order to ensure effective monitoring of the fish farm conditions and to take into account the efficient use of computing resources, the cameras are set to automatically capture an image every few minutes. This timed image capture method can not only effectively reduce the amount of data processing, but also ensure the timeliness and accuracy of monitoring.

[0095] The anomaly detection model analyzes the captured images. In this crucial step, timed images captured from the fishing grounds are fed into a specially designed anomaly detection model for analysis. This model is based on advanced deep learning technology and is designed to process and analyze highly complex, high-dimensional image data, aiming to identify any unusual features in the images.

[0096] The anomaly detection model, namely AnomalyCLIP Enhanced Edition, can not only accurately identify changes in fish behavior (such as changes in swimming patterns), water quality problems (such as changes in color or transparency), and other potential problems (such as the presence of floating objects or unidentified objects on the water surface), but also provide more detailed anomaly detection results through its unique algorithm structure. When an image is input into the model, it first undergoes a series of complex feature extraction processes to identify and quantify the relationships and changes between various elements in the image.

[0097] In terms of output, AnomalyCLIP Enhanced Edition provides two main types of results: anomaly assessment values ​​and anomaly heatmaps. Anomaly assessment values ​​are a quantitative evaluation of whether there are anomalies in each input image, usually represented by a value between 0 and 1, where the closer to 1, the higher the probability of an anomaly. This quantitative method allows managers to quickly gain an intuitive understanding of the overall health of the fish farm and take appropriate measures based on the set thresholds. Meanwhile, anomaly heatmaps provide a more intuitive way to display the location and severity of anomalies. On the anomaly heatmap, different colors represent different levels of anomaly severity, which allows even non-experts to quickly locate and understand the problem areas in the image.

[0098] After the anomaly detection model completes the image analysis, if an anomaly is confirmed, it generates a structured and highly informative prompt template. This template is designed to present the analysis results in a clear, concise, and operable manner, ensuring that all relevant information is accurately recorded and providing a reliable basis for subsequent decision-making. The prompt template design follows these principles:

[0099] Modular design: The template consists of multiple dynamic fields, each corresponding to different analysis dimensions or key information, such as anomaly type, severity, timestamp, location information, etc. This modular design allows the template to flexibly adapt to the needs of different scenarios while ensuring the integrity of information.

[0100] Semantic clarity: To facilitate human-computer interaction and subsequent automated processing, the prompt word template adopts a combination of natural language and structured data. Natural language is used to describe the specific manifestations and possible impacts of the anomaly, while structured data is used to provide quantitative indicators and precise time and location information.

[0101] Visualization support: In addition to plain text information, the prompt word template also embeds thumbnails of the anomaly heatmap, allowing users to intuitively view the problem area and its severity. This combination of text and visual information greatly improves the efficiency of problem understanding.

[0102] Specifically, our prompt word template includes the following core components:

[0103] Title: Briefly summarize the nature of the abnormal situation;

[0104] Timestamp: Records the specific time when the anomaly occurred in order to track the development trend of the problem;

[0105] Anomaly Location: By analyzing the anomaly heatmap, the areas where the anomalies occurred are marked, such as: Anomaly Area: Southeast corner of the fish farm, close to the coverage area of ​​camera No. 3;

[0106] Recommended actions: Based on the type and severity of the anomaly, some preliminary response suggestions are automatically generated, such as: it is recommended to immediately check water quality parameters and remove floating debris from the water surface;

[0107] Additional information: Includes an embedded image of the abnormal heatmap, as well as other relevant background information.

[0108] An intelligent agent will further analyze the current situation of the fishery, which may include assessing the impact of the anomaly, determining the measures that need to be taken, and regulating IoT devices. Then, the risk level will be assessed, which will help determine the priority and urgency of subsequent actions. Finally, the system will notify the fishery management of all relevant information, including the results of the anomaly detection, the conclusions of the risk assessment, and any suggested action plans.

[0109] Considering the complexity of the fishing grounds and the high false alarm rate of traditional anomaly detection models, AnomalyCLIP was improved and embedded into the workflow of the entire system.

[0110] like Figure 3 The diagram shows the principle of the anomaly detection model, and the main process is as follows:

[0111] Text Encoder: A text encoder converts normal and abnormal text descriptions into feature vectors.

[0112] ftext = TextEncoder(T), where T includes T 正常 and T 异常 ;

[0113] Image encoder: The image encoder converts the input image I into a feature vector. During the encoding process, each patch is retained, and features are saved every 6 blocks, for a total of 24 blocks, resulting in 4 sets of patch features. Each of them It contains patch features for 6 blocks;

[0114] At the same time, an image-level feature representation is also obtained: fimage = ImageEncoder(I);

[0115] Similarity calculation: In order to detect the degree of anomaly, the similarity between image-level features and text tokens, and the similarity between patch features and text tokens are calculated separately.

[0116] Image-level similarity: used to assess the overall degree of anomaly, where, ;

[0117] Patch-level similarity: used to generate anomaly maps and locate specific anomaly regions.

[0118] Introducing Subject Focus Loss: Since anomaly maps may include background anomalies, a mechanism is needed to make the model focus more on subject anomalies. A Subject Focus Loss (SF Loss) is designed to optimize this process. During model training, a mask annotation for the subject object is performed on each image. For a binary mask M, it marks the position of the subject object in the image. SF Loss can be defined as follows:

[0119] ;

[0120] Where wi is a weighting factor that is adjusted according to the actual situation, and Mi indicates whether the i-th patch belongs to the main region (0 or 1).

[0121] Total Loss Function: The total loss function combines the original loss with the newly added SF Loss. , where α is a hyperparameter that balances the two loss terms.

[0122] The intelligent decision-making agent module is used to activate the inference engine when an abnormal situation is detected, comprehensively analyze the current environmental state, generate and execute corresponding decision instructions. Specifically, when abnormal fish behavior and water quality are detected, the aerator is activated first to improve water quality conditions.

[0123] In this embodiment, the intelligent decision-making agent module includes:

[0124] The visual description generation unit is used to perform pixel-level segmentation of abnormal region images, extract visual elements of color distribution and shape features, and generate structured descriptions, such as "a dark clump-like object with a diameter of about 15cm was found in the northeast region".

[0125] The equipment control instruction generation unit is used to generate equipment control instructions. It has a built-in instruction template library and controls the aerator and feeder. For example, the aerator control is: {Equipment ID: A01, Operation: Start, Duration: 30min}, and the feeder control is: {Equipment ID: B02, Operation: Pause, Reason: Suspected disease}.

[0126] The safety verification unit is used to check whether the current status of the equipment allows operation before sending equipment instructions, and to verify whether the instructions comply with safety specifications, such as the maximum single operation time of the aerator is ≤2 hours.

[0127] Specifically, the fish farm management system employs a highly integrated intelligent agent as its core component, responsible for the entire process from anomaly detection to decision execution. This intelligent agent integrates abnormal data, analyzes abnormal situations, and generates decision instructions, achieving integrated management from perception to action. The intelligent agent integrates the following core functional modules: 1) Data analysis module: responsible for receiving data from high-resolution cameras and processing and analyzing it in real time; 2) Anomaly detection module: based on the AnomalyCLIP model, it performs anomaly detection on the data, identifies potential abnormal events, analyzes fish behavior characteristics such as swimming speed and aggregation density to determine if there are health risks, and assesses water quality change trends based on the water environment and underwater camera footage; 3) Decision execution module: based on the anomaly detection results, it comprehensively analyzes the current environmental state, generates and executes corresponding decision instructions. If abnormal fish behavior and water quality are detected, it prioritizes starting aerators to improve water quality conditions. In long-term monitoring, the intelligent agent can also predict potential risks, such as seasonal water quality changes, and take preventative measures in advance.

[0128] like Figure 4 The flowchart of the large-scale intelligent agent is shown below. The specific process is as follows:

[0129] The AnomalyCLIP model is used to detect anomalies in the input video frames, generating an anomaly prediction map. Based on the anomaly prediction map, top-k anomaly coordinates are formed. According to the anomaly coordinate information, a pre-set Prompt template is triggered. The InternVedio2.5 module receives the anomaly coordinate information and instruction information, processes the video data, and finally sends the anomaly information and coordinate information to the fish farm management personnel. The abnormal areas are tracked and marked in the video, which facilitates monitoring and automatic adjustment of IoT devices according to instruction information, such as adjusting parameters such as water quality and temperature.

[0130] InternVedio 2.5 is an advanced video processing module specifically designed to process real-time video streams from fish farm monitoring systems. Its core functions include: 1) Real-time tracking: Tracking detected abnormal areas in real time; 2) Marking and recording: Marking abnormal areas in video frames and recording relevant data for subsequent analysis; 3) Intelligent feedback: Generating intelligent feedback based on tracking results to guide further operations or adjustments to IoT devices.

[0131] Based on the received anomaly coordinates, InternVedio 2.5 initializes one or more object trackers. Each tracker is responsible for monitoring a specific anomaly area. As video frames are continuously updated, the trackers continuously calculate the new position of the anomaly area. Even if the target moves or the environment changes, the trackers can maintain their lock on the anomaly area. In addition to real-time display, InternVedio 2.5 also records relevant information for each anomaly event, including timestamps and location change trajectories. This data can be used to generate detailed reports to help managers understand the development trend of anomalies and make decisions accordingly. Based on the tracking results, InternVedio 2.5 can trigger preset actions or adjust relevant IoT device settings.

[0132] like Figure 5 The diagram shown illustrates the interaction between the Agent and IoT devices. The specific interaction process is as follows:

[0133] The agent triggers a series of operations through the Function Calling mechanism, mainly including notifying fishery management personnel, tracking abnormal areas, and controlling IoT devices.

[0134] Notifying fish farm management personnel: When an anomaly is detected, the Agent will immediately notify the fish farm management personnel. The notification methods include sending email notifications via the SMTP protocol and sending SMS notifications via the carrier API. These two notification methods ensure that the information can be delivered to the management personnel in a timely manner so that they can take appropriate measures quickly.

[0135] Tracking abnormal areas: The Agent is responsible for tracking abnormal areas detected in the video in real time. This function is closely related to the InternVedio2.5 module mentioned above, ensuring continuous monitoring and recording of abnormal situations.

[0136] Performing IoT device control: Based on tracking results and preset rules, the Agent controls various IoT devices through the MQTT protocol. Specifically, it can: 1) control the aerator switch to adjust the oxygen content in the water; 2) control the water temperature regulator to maintain a suitable water temperature environment; 3) control the feeder to automatically adjust the feed amount and time; 4) control other related IoT devices to achieve comprehensive environmental management and optimization.

[0137] Through the above interaction process, the Agent can not only detect and handle abnormal situations in a timely manner, but also intelligently adjust the working status of IoT devices, thereby providing fish farms with an efficient and intelligent management solution.

[0138] like Figure 6The diagram shows the network structure of a fish farm IoT system. This system achieves comprehensive monitoring and intelligent management of the aquaculture environment through a multi-layered network architecture. The specific structure is as follows:

[0139] Agent layer: Located at the top of the system is the Agent module, which is the control center of the entire system. The Agent is responsible for issuing control commands and receiving environmental data from each LoRa gateway. This data includes key parameters such as water quality and temperature, which are used to monitor the status of the aquaculture environment in real time.

[0140] LoRa Gateway Layer: The Agent communicates with lower-layer devices through multiple LoRa gateways. Each LoRa gateway is responsible for data transmission in a specific area. To ensure data security and integrity, LoRa gateways use TLS / SSL encryption and JSON data encapsulation technology for data exchange.

[0141] Aquaculture Area Node Layer: Below the LoRa gateway are the aquaculture area nodes, mainly including the ESP32 module and the SX1278 module. The ESP32 acts as a microcontroller, responsible for processing and forwarding data; while the SX1278 acts as a wireless communication module, ensuring that data can be stably transmitted to the LoRa gateway. This layer also connects various sensors and actuators, such as cameras, aerators, and feeders, to collect environmental data and execute control commands.

[0142] Management Center Layer: Alongside the LoRa gateway is the management center, which is responsible for functions such as data statistics, SMS push and email alarms. When an anomaly is detected, the management center will promptly notify relevant personnel and issue control commands to adjust the working status of relevant devices.

[0143] Through the above network structure, an efficient and reliable Internet of Things (IoT) system for fish farms was constructed, enabling comprehensive monitoring and intelligent management of the aquaculture environment, which greatly improves the operational efficiency and management level of fish farms.

[0144] Among them, the message notification IoT communication module is used to realize real-time information transmission and device control using IoT communication modules based on LoRaWAN technology, enabling stable data interaction between the aquaculture area and the fish farm management center.

[0145] In this embodiment, the message notification IoT communication module includes:

[0146] Low-power wide-area network communication unit, used for long-distance, low-power data interaction based on LoRaWAN protocol wireless transmission technology;

[0147] The hardware configuration consists of an ESP32 main control chip and an SX1278 RF module, supporting multiple spreading factors (SF7~SF12) and bandwidth settings (BW125kHz~BW500kHz) to adapt to different communication distance and speed requirements.

[0148] The network architecture adopts a star topology, and the data from each monitoring node is centrally received by a LoRa gateway deployed in the center of the breeding area and forwarded to the local server or cloud platform.

[0149] The data is encapsulated in JSON encoding format and includes timestamps, device IDs, data types, and encrypted fields.

[0150] The TLS / SSL encryption protocol is used during data transmission to prevent data leakage and unauthorized access. At the same time, a unique identifier is assigned to each device for access authentication and access control.

[0151] In this embodiment, the message notification IoT communication module further includes:

[0152] The remote command receiving and execution unit is used to receive and parse operation commands from the fishery management center;

[0153] Supports lightweight command delivery via the MQTT protocol, enabling stable command reception in weak network environments;

[0154] The instruction types cover equipment start-up and shutdown control (such as aerators and feeders) and parameter adjustments (such as camera resolution and detection frequency). The received instructions are verified for origin and integrity to prevent malicious attacks or misoperations.

[0155] In this embodiment, the message notification IoT communication module further includes:

[0156] The multi-channel anomaly alarm notification unit is used to proactively send alarm information to administrators via email notification and SMS push service after an anomaly event is detected.

[0157] Among them, an alarm email is sent to a preset email address based on the SMTP protocol. The content includes the anomaly type, the time of occurrence, the location coordinates, and the preliminary analysis results.

[0158] Integrate Alibaba Cloud SMS service API to call telecom operator resources via HTTP interface and send concise and clear SMS notifications to specified mobile phone numbers;

[0159] Develop a multi-contact configuration and resend scheme to automatically attempt to resend the notification if the first notification fails, ensuring that critical information is not lost, and allow users to customize the notification content format according to different levels of anomalies;

[0160] The communication status monitoring and optimization unit is used to monitor the stability of the communication link and the operating status of the equipment in real time.

[0161] During signal strength detection, the received signal strength indication value of the LoRa module is periodically reported to assist in evaluating communication quality. During power monitoring, the current battery voltage and remaining power of the device are collected. When the power level is lower than the set threshold, a low power warning notification is triggered. During network interruption, data to be sent is buffered and retransmitted in priority order after the connection is restored. The module's working mode is dynamically adjusted according to the communication load.

[0162] Specifically, in the fish farm management system, ensuring real-time information transmission and equipment control is a key aspect of achieving efficient management. A LoRaWAN-based message notification IoT communication module is adopted as the primary information transmission solution to ensure stable data interaction between the aquaculture area and the fish farm management center. Specifically, the message notification IoT communication module is responsible for the following key tasks: 1) Data upload: Through LoRa nodes deployed at each monitoring point, data generated by the visual information acquisition and processing module and the lightweight anomaly detection module are uploaded to the local server, enabling managers to obtain the latest monitoring data anytime, anywhere, and perform corresponding analysis and decision-making; 2) Remote control command reception: Supports operation commands sent from the management center or other authorized terminals, such as starting the aerator and adjusting the feeder's working status, and accurately transmits these commands to the target equipment in a timely manner. 3) Message notification function: After detecting an abnormal event, the integrated message push mechanism will proactively send alarm information to the fishery management personnel, including simultaneous notification of relevant personnel through both email (SMTP protocol) and SMS (directly through the API provided by the telecom operator), to ensure that abnormal situations can be responded to and handled in a timely manner. The email notification content covers the type of abnormality, the time of occurrence, location information and possible impact; the SMS notification provides a concise and clear summary of key information, which makes it easy for management personnel to quickly grasp the on-site situation while on the move. In addition, in order to improve the reliability and coverage of notifications, multi-level contact configuration is supported, allowing multiple receiving email addresses and mobile phone numbers to be set, and a resend mechanism is provided to deal with communication failures. All message content can be customized according to different types of abnormalities, enhancing the flexibility and applicability of the system.

[0163] To better demonstrate the fishery management process based on lightweight anomaly detection and intelligent agent collaborative decision-making, this embodiment provides a fishery management method based on lightweight anomaly detection and intelligent agent collaborative decision-making, implemented based on the aforementioned fishery management system based on lightweight anomaly detection and intelligent agent collaborative decision-making, including:

[0164] Step 1: Use high-resolution cameras to monitor the status of the fish farm in real time, continuously capture images and video data of the fish farm, and perform preliminary processing on the captured visual data of the aquaculture area. The processed visual data of the aquaculture area is then transmitted to the anomaly detection system via the network.

[0165] Step 2: The anomaly detection system performs in-depth analysis of visual data from the aquaculture area, identifies various anomalies in the images, including abnormal fish behavior, water quality changes, and site anomalies, and generates an anomaly analysis map based on the analysis results, which intuitively displays the specific location and type of the anomaly.

[0166] Step 3: Based on the anomaly analysis chart and the current state of the fish farm, take corresponding measures, automatically control the IoT devices, adjust the water temperature and the operating status of the aerator to deal with the abnormal situation, and at the same time, notify the fish farm management personnel of the abnormal situation for timely intervention and handling.

[0167] The system employs a LoRaWAN-based message notification IoT communication module as the primary information transmission solution to ensure stable data interaction between the aquaculture area and the fish farm management center. Furthermore, upon detecting anomalies, it proactively sends alarm messages to fish farm management personnel via an integrated push notification mechanism (using SMTP protocol to send emails and SMPP protocol or telecom operator API to send SMS messages), ensuring timely response and handling of abnormal situations. All data transmissions utilize TLS / SSL encryption technology to guarantee data security and integrity, making it highly suitable for applications in aquaculture and smart agriculture.

[0168] 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 process, method, article, or apparatus.

[0169] 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 fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making, characterized in that, include: The visual information acquisition and processing module is used to monitor the behavior of fish, water quality and equipment operation in the aquaculture area in real time using a high-resolution camera, acquire visual data of the aquaculture area, and perform cropping, scaling, rotation, flipping and color transformation on the acquired visual data of the aquaculture area, remove noise and filter the data. The lightweight anomaly detection module uses a zero-shot anomaly detection model based on AnomalyCLIP to periodically scan the aquaculture area according to an adaptive frequency adjustment scheme, maintaining high-precision monitoring in resource-constrained environments and identifying and classifying potential anomalies in real time. The intelligent decision agent module is used to start the inference engine when an abnormal situation is detected, comprehensively analyze the current environmental state, generate and execute corresponding decision instructions. Among them, when abnormal fish behavior and water quality are detected, the aerator is activated first to improve water quality conditions. The message notification IoT communication module is used to realize real-time information transmission and device control using IoT communication modules based on LoRaWAN technology, enabling stable data interaction between the aquaculture area and the fish farm management center.

2. The fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making as described in claim 1, characterized in that, The visual information acquisition and processing module includes: High-resolution cameras are deployed at key locations in aquaculture areas to capture real-time fish behavior, water quality, and equipment operating status, thereby determining visual data of the aquaculture area. The video frame preprocessing unit is used to crop and scale the visual data of the aquaculture area, extract key areas, perform color correction to compensate for color deviation caused by underwater light attenuation, and timestamp the visual data of the aquaculture area so that each frame corresponds to its acquisition time. The data augmentation unit is used to enhance the visual data of the breeding area by randomly rotating and flipping it to simulate different perspectives and directions, and to perform color transformation, adjust brightness, contrast and saturation to simulate different lighting conditions.

3. The fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making as described in claim 2, characterized in that, The visual information acquisition and processing module further includes: The hardware energy-saving optimization unit is used to reduce the energy consumption of the visual information acquisition and processing module; Among them, a dynamic frame rate adjustment scheme was formulated, and the sampling frequency was set according to the day and night cycle. During the daytime period from 6:00 to 18:00, the frame rate was 15 FPS, and during the nighttime period from 18:00 to 6:00, the frame rate was 5 FPS. A low-power high-resolution camera chip was adopted, with a single device standby power consumption of ≤ 0.5W and working power consumption of ≤ 1.5W. The data storage and transmission unit is used to cache the raw data of the past 7 days using a high-speed SSD, compress the video stream based on H.265 encoding, and upload it to the cloud via AES encryption protocol. It automatically saves the untransmitted data when the network is interrupted and continues uploading after the connection is restored.

4. The fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making as described in claim 3, characterized in that, The lightweight anomaly detection module includes: The periodic anomaly detection unit uses an improved AnomalyCLIP architecture to scan the aquaculture area every Δt minutes and calculates the anomaly confidence level Panomaly by comparing it with a preset normal state feature vector. The trigger-based agent activation unit is used to send a trigger signal containing the coordinates and timestamp of the abnormal area to the smart agent when the anomaly confidence level Panomaly ≥ a set threshold θ. The inspection frequency dynamic adjustment unit is used to automatically set the inspection interval according to the day and night cycle. During the daytime period from 6:00 to 18:00, the inspection interval Δt = 5 minutes, and during the nighttime period from 18:00 to 6:00, the inspection interval Δt = 15 minutes. Anomaly confidence calculation unit, used to calculate anomaly confidence and assess the degree of anomaly; The hardware power-saving unit is used to save power by using a dedicated NPU chip in the Raspberry Pi 5, with a single detection power consumption of ≤1.2W.

5. The fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making according to claim 4, characterized in that, The threshold θ of the triggered agent activation unit is set as follows: Base threshold When the confidence level is abnormal At that time, a regular proxy response is triggered; Emergency threshold When the confidence level is abnormal When this occurs, the following enhanced responses will be activated: The detection interval is temporarily shortened to Δt=1 minute. When an anomaly is detected, an alert SMS is sent to the administrator's mobile phone via Alibaba Cloud SMS service.

6. The fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making according to claim 5, characterized in that, The intelligent decision-making agent module includes: The visual description generation unit is used to perform pixel-level segmentation of abnormal region images, extract visual elements of color distribution and shape features, and generate structured descriptions. The equipment control instruction generation unit is used to generate equipment control instructions. It has a built-in instruction template library and controls the aerator and feeder. The security verification unit is used to check whether the current state of the device allows operation before sending device commands, and to verify whether the commands comply with security specifications.

7. The fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making as described in claim 6, characterized in that, The message notification to the IoT communication module includes: Low-power wide-area network communication unit, used for long-distance, low-power data interaction based on LoRaWAN protocol wireless transmission technology; It consists of an ESP32 main control chip and an SX1278 RF module, and supports multiple spreading factors and bandwidth settings to adapt to different communication distance and speed requirements. Using a star topology, a LoRa gateway deployed at the center of the aquaculture area centrally receives data from each monitoring node and forwards it to a local server or cloud platform. Data is encapsulated using JSON encoding format, including timestamps, device IDs, data types, and encrypted fields; The TLS / SSL encryption protocol is used during data transmission to prevent data leakage and unauthorized access. At the same time, a unique identifier is assigned to each device for access authentication and access control.

8. The fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making according to claim 7, characterized in that, The message notification IoT communication module also includes: The remote command receiving and execution unit is used to receive and parse operation commands from the fishery management center; Supports lightweight command delivery via the MQTT protocol, enabling stable command reception in weak network environments; The instruction types cover equipment start-up and shutdown control and parameter adjustment, and the received instructions are verified for origin and integrity to prevent malicious attacks or misoperations.

9. The fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making as described in claim 8, characterized in that, The message notification IoT communication module also includes: The multi-channel anomaly alarm notification unit is used to proactively send alarm information to administrators via email notification and SMS push service after an anomaly event is detected. Among them, an alarm email is sent to a preset email address based on the SMTP protocol. The content includes the anomaly type, the time of occurrence, the location coordinates, and the preliminary analysis results. Integrate Alibaba Cloud SMS service API to call telecom operator resources via HTTP interface and send concise and clear SMS notifications to specified mobile phone numbers; Develop a multi-contact configuration and resend scheme to automatically attempt to resend the notification if the first notification fails, ensuring that critical information is not lost, and allow users to customize the notification content format according to different levels of anomalies; The communication status monitoring and optimization unit is used to monitor the stability of the communication link and the operating status of the equipment in real time. During signal strength detection, the received signal strength indication value of the LoRa module is periodically reported to assist in evaluating communication quality. During power monitoring, the current battery voltage and remaining power of the device are collected. When the power level is lower than the set threshold, a low power warning notification is triggered. During network interruption, data to be sent is buffered and retransmitted in priority order after the connection is restored. The module's working mode is dynamically adjusted according to the communication load.

10. A fish farm management method based on lightweight anomaly detection and intelligent agent collaborative decision-making, implemented based on the fish farm management system based on lightweight anomaly detection and intelligent agent collaborative decision-making as described in claim 9, characterized in that, include: Step 1: Use high-resolution cameras to monitor the status of the fish farm in real time, continuously capture images and video data of the fish farm, and perform preliminary processing on the captured visual data of the aquaculture area. The processed visual data of the aquaculture area is then transmitted to the anomaly detection system via the network. Step 2: The anomaly detection system performs in-depth analysis of visual data from the aquaculture area, identifies various anomalies in the images, including abnormal fish behavior, water quality changes, and site anomalies, and generates an anomaly analysis map based on the analysis results, which intuitively displays the specific location and type of the anomaly. Step 3: Based on the anomaly analysis chart and the current state of the fish farm, take corresponding measures, automatically control the IoT devices, adjust the water temperature and the operating status of the aerator to deal with the abnormal situation, and at the same time, notify the fish farm management personnel of the abnormal situation for timely intervention and handling.

Images