Marine ranch net cage fish body individual monitoring system

Abstract

The invention relates to the technical field of fishery breeding, and discloses a marine ranch net cage fish body individual monitoring system, which comprises a frame, a batch feeder, a sonar detection module and a movable bracket, the batch feeder is fixedly mounted at the center position of the frame, the sonar detection module and the movable bracket are fixedly mounted at the midpoints of four center separation structures of the frame, and the sonar detection module is connected with the movable bracket. The cameras are fixedly installed at the top ends of the four movable supports, the multiple sets of radio frequency identification devices are evenly and additionally installed on each center separation structure of the frame in the length direction, the swimming robot is arranged in the offshore net cage, and the data processing terminal is integrally installed on a rack of the batch feeder. And the passive radio frequency chip is implanted into the fish body. According to the method, an original split visual observation index is converted into a quantifiable and comprehensive individual health risk coefficient, the early recognition accuracy and reliability of the sick, weak or abnormal fish body are remarkably improved, and missing report or false report caused by misjudgment of a single feature is reduced.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, and in particular to a monitoring system for individual fish in marine ranch cages. Background Technology

[0002] In recent years, with increasing pressure on marine fishery resources, developing deep-sea aquaculture has become a national strategic direction for ensuring food security and building a "blue granary." Traditional near-shore aquaculture models are rapidly transforming into offshore, mechanized, and intelligent deep-sea marine ranching models. In this process, leveraging information technology to achieve refined and intelligent management of farmed fish is crucial for improving aquaculture efficiency, reducing disease risks, and ensuring product quality and safety. Currently, the industry generally recognizes that relying solely on manual experience for feeding and inspection is no longer sufficient to meet the management needs of large-scale, deep-sea aquaculture; the application of intelligent monitoring and management systems is an inevitable trend.

[0003] To address these needs, industry and academia have conducted numerous technological explorations and developed a series of patented solutions aimed at improving the automation and intelligence of aquaculture monitoring. These existing technologies mainly focus on the following areas:

[0004] Firstly, in terms of macro-monitoring and feeding control of fish schools, integrated monitoring systems combining multiple sensors have emerged. For example, the invention patent with publication number CN114847210B, entitled "An Intelligent Three-Dimensional Monitoring System for Large-Scale Deep-Sea Aquaculture Farms," ​​utilizes the synergy of a fish population monitoring system, a water environment monitoring system, an automatic feeding system, and a cloud platform. Based on environmental parameters such as water temperature, water flow, and dissolved oxygen, as well as an estimated fish population, it intelligently calculates and controls the feed weight G using a specific formula. Such systems achieve data acquisition and feedback control from the environmental perspective to the fish population, representing progress in macro-management.

[0005] Secondly, in visual analysis techniques targeting individual fish, existing patents focus on improving the accuracy and robustness of image processing. For example, CN117809331B's "Method, Apparatus, Electronic Device, and Storage Medium for Detecting Fish Swimming Posture" utilizes an infrared camera on the water surface to acquire images, locates fish using a target detection neural network, performs precise segmentation using an unsupervised image segmentation network, and finally obtains the fish's swimming posture through ellipse fitting. Another patent, CN117994534B's "Method for Detecting Feature Points of Aquatic Animals," improves the robustness of the feature point detection network and solves the problem of outlier outputs by constructing a fish shape model, designing a prior fish set, and a joint cost function. These technologies provide fundamental tools for image-based individual behavior analysis.

[0006] Furthermore, in terms of the deployment and inspection of monitoring equipment, some patents aim to expand the monitoring range. For example, the patent CN218734495U, "An Underwater Monitoring Device for Net Cage Aquaculture," solves the problem of blind spots in fixed underwater monitoring devices by setting a horizontal moving component, a rotating motor, and a lifting component on the top of the net cage to drive the underwater camera to move horizontally, lift, and rotate within the net cage. In addition, using underwater robots (AUVs) or intelligent inspection systems to inspect netting, automatically detecting damage and returning data, is also a feasible solution to improve the efficiency of safety inspections.

[0007] Finally, in the area of ​​using sonar for population assessment, methods for estimating fish populations by combining artificial intelligence algorithms have emerged. For example, the paper "A Method for Estimating Fish Population in Deep-Sea Cage Culture Based on Image Sonar" (publication number CN117830814A) employs forward-view image sonar for continuous detection, combined with a YOLO target detection model with added attention mechanism and a BP neural network, to achieve real-time automatic estimation of fish populations, providing a basis for feeding and catch planning.

[0008] Despite the significant progress made in the aforementioned existing technologies, there are still obvious technological gaps and deficiencies in achieving true "individual fish monitoring" and "precise treatment," which may lead to the following problems in practical work:

[0009] First, existing systems lack continuous and reliable methods for tracking the identity of individual fish. Current solutions often focus on group behavior (such as overall population size and macroscopic swimming movements) or single visual identification, failing to reliably link specific behavioral anomalies (such as poor feeding or unusual swimming posture) to a single, uniquely identified fish in the complex environment of fish cages. For example, a camera may detect an abnormal fish, but once it rejoins a school of fish, the system cannot accurately identify it again and track its subsequent movements. This makes early detection and continuous isolation monitoring of diseased fish difficult, increasing the risk of disease transmission within the group.

[0010] Second, multimodal perception data has failed to be effectively integrated and coordinated around the "individual." Subsystems such as environmental monitoring, swarm sonar, visual analysis, and feeding control often operate independently or only perform simple data parallelization. When an AI visual algorithm identifies an abnormal fish in a certain area, the system cannot automatically direct specialized equipment in that area to accurately identify the suspect, nor can it direct a mobile robot to the precise location for close-up confirmation. This fragmentation of perception, identification, location, and response prevents early warnings from automatically translating into precise intervention actions, still requiring significant manual intervention for judgment and operation, resulting in low efficiency and slow response.

[0011] Third, existing monitoring solutions lack a complete closed-loop system for handling abnormal events. Most systems stop at "monitoring" and "alarming." Even if an abnormal individual is identified and its approximate location is known, there is a lack of an automated mechanism to guide mobile equipment to the site, collect high-definition diagnostic images, and integrate all information into the fish's individual file. This fragments valuable health management data, making it impossible to form a complete health profile of an individual throughout its life cycle, which is detrimental to disease analysis and aquaculture optimization.

[0012] In summary, existing marine ranching cage monitoring technologies face bottlenecks in transitioning from "group management" to "precise individual management." This stems from the failure to construct an integrated monitoring system centered on individual fish identities, deeply coupling functions such as zoned video AI monitoring, full-coverage RFID arrays, precise mobile robot scheduling, and automatic multi-source data integration and archiving. Therefore, an innovative system solution is urgently needed to address the aforementioned technical problems of fragmented perception, decoupled identification, and incomplete response loops. This would enable early detection of anomalies in individual fish, precise identification, rapid on-site confirmation, and complete record-keeping, ultimately improving the health management level and risk control capabilities of deep-sea aquaculture.

[0013] Therefore, we propose an individual monitoring system for fish in marine ranch cages. Summary of the Invention

[0014] The present invention mainly addresses the technical problems existing in the prior art and provides a monitoring system for individual fish in marine ranch cages.

[0015] To achieve the above objectives, the present invention adopts the following technical solution: a marine ranching cage fish individual monitoring system, comprising:

[0016] The frame is fixedly connected to the offshore gabion frame by anti-corrosion anchor piles, and the frame is divided into four independent material feeding zones.

[0017] The feeding machine is fixedly installed at the center of the frame, and the bottom of the feeding machine is equipped with a sealed, waterproof, adjustable rotating base.

[0018] A sonar detection module, which is fixedly installed above the frame of the feeding machine;

[0019] There are four movable supports, which are fixedly installed one-to-one at the midpoint of the four central partition structures of the frame. The movable supports are made of stainless steel and the tilt angle is adjustable.

[0020] Each camera is fixedly installed on the top of one of the four movable supports, and each camera is aimed at a feeding section of the frame.

[0021] Radio frequency identification (RFID) devices, wherein multiple sets of RFID devices are uniformly mounted on each central partition structure of the frame along the length direction, and the multiple sets of RFID devices form a full-coverage RFID array on the frame;

[0022] A towed robot is configured inside the marine cage, and the marine cage has a pre-set docking point for the towed robot. The towed robot is equipped with a radio frequency identification module with dual functions of counting and ID confirmation, as well as a high-definition waterproof visual acquisition module. Its radio frequency identification module performs the counting function when it receives the position information of the non-surfacing fish from the sonar detection module, and performs the ID confirmation function when it receives abnormal fish information from the data processing terminal.

[0023] A data processing terminal is integrated and installed on the frame of the feeding machine. The data processing terminal establishes real-time signal connections with the feeding machine, the sonar detection module, the camera, the radio frequency identification device, and the roaming robot.

[0024] And passive radio frequency chips implanted in fish.

[0025] Preferably, the data processing terminal is equipped with customized data processing software, which integrates an AI anomaly detection algorithm module. This AI anomaly detection algorithm module establishes a continuous signal connection with the camera, receiving continuous video stream data collected in real time from cameras in each zone as input. The module first performs frame-by-frame decoding and target detection on the input video stream, segmenting the independent image region of each fish in the image. Then, it extracts a color histogram from each segmented independent image region of the fish and converts it into a standardized body color feature vector. This body color feature vector is input into a pre-trained deep neural network for fish disease diagnosis. Combined with body color feature samples of various fish diseases in the training set, it determines whether each fish is diseased and the specific type of disease. Afterwards, a multi-dimensional feature extraction process is executed. The feature extraction process includes calculating the gray-level co-occurrence matrix of the fish image region to quantify texture features, extracting the color histogram of the image to characterize body color distribution features, and calculating the fish's centroid displacement vector field through continuous inter-frame difference and optical flow methods to quantify swimming posture and trajectory features. Simultaneously, the AI ​​anomaly detection algorithm module, combined with the feeder's operating status signal, counts the number of fish entering the preset feeding area below the feeder and their dwell time within a unit of time during the feeding period, generating a feeding activity time series. Finally, the AI ​​anomaly detection algorithm module inputs the extracted swimming posture feature vector, body color feature vector, and feeding activity sequence into a pre-trained multi-input deep neural network classifier for fusion analysis and joint inference; the classifier outputs a probability value for judging the fish's health status. ,when Exceeding the preset threshold If the fish is found to be abnormal, an abnormal event record is generated, which includes the timestamp of the abnormal fish's appearance, the camera number where it is located, the pixel coordinates in the image, the disease determination result, and the suspected disease. This abnormal event record is sent to the central control unit of the data processing terminal in real time, and at the same time, the data processing terminal sends an instruction to the instruction scheduling module to dispatch the tow robot.

[0026] Preferably, upon receiving an anomaly event record from the AI ​​anomaly detection algorithm module, the data processing terminal immediately triggers an instruction scheduling process. The central control unit of the data processing terminal parses the camera number and pixel coordinate information in the anomaly event record and maps them to the corresponding physical feeding zone. Subsequently, the central control unit sends a control signal containing a high-power scanning mode instruction and a specific scanning frequency set to all RFID device groups responsible for the mapped zone. Upon receiving the instruction, the RFID device groups simultaneously activate a high-intensity, directional radio frequency field to perform intensive scanning of the target zone's water body. After the RFID device groups read the radio frequency signal carrying a globally unique identification code reflected by the passive radio frequency chip inside the abnormal fish, they upload a scan data packet containing the identification code, a reading timestamp, and the device's own serial number to the data integration module of the data processing terminal. The data integration module performs time and space correlation matching between the received scan data packet and the previously received anomaly event record. When the time difference between the two is... When the value is less than the set threshold and the physical location is logically consistent, the identity information of the abnormal fish is confirmed and locked, namely its unique chip identification code. The identification code is then bound to the time when it was first identified as abnormal and the partition information it is located in, creating a complete abnormal fish tracking file.

[0027] Preferably, the customized data processing software of the data processing terminal integrates a data integration module; the data integration module continuously receives raw scan logs periodically uploaded from all RFID devices on the frame, the logs containing the read chip identification code, the reading time, and the device number; the data integration module first performs a data cleaning step, removing data with signal strength below a threshold. Invalid read records were then removed; subsequently, a deduplication algorithm based on time windows and device proximity was applied to the cleaned data. This algorithm is defined as follows: for the same chip identification code, within a time window... Within a given set of records reported by multiple physically adjacent RFID devices, only the record with the strongest signal strength is retained as a valid read. After deduplication, the module calculates statistics for each RFID device in each statistical period based on the preset partition affiliation information. The number of unique chip identification codes appearing in each feeding zone indicates the number of fish aggregated in that zone during that period. ,in This represents the partition index; ultimately, the data integration module, according to the time series, integrates the data from each partition. In addition, the chip identification code list is integrated to generate a structured partitioned fish distribution data table, which is stored in the terminal database and used for subsequent analysis and visualization.

[0028] Preferably, the customized data processing software of the data processing terminal integrates an instruction scheduling module; the first function of the instruction scheduling module is to process the missed detection signal of the floating fish uploaded by the sonar detection module, the signal containing the depth information of the floating fish and its polar coordinate position relative to the sonar origin; the module calls the built-in three-dimensional space model of the net cage and converts the polar coordinate position into a rectangular coordinate system coordinate system inside the net cage with the docking point as the origin. The system plans a collision-free underwater path from the current coordinates or default docking point coordinates of the tow robot to the target coordinates. After planning, the instruction scheduling module generates a first navigation instruction set, which includes the target coordinates and a sequence of key points along the path, and sends it to the motion controller of the tow robot via a wireless communication link. The second function of the instruction scheduling module is to respond to the abnormal fish identification information and its location partition locked by the data integration module. Based on the partition where the abnormal fish is located, the module retrieves the coordinates of the pre-stored center point or typical patrol point of that partition. Similarly, a navigation path is planned and a second navigation instruction set is generated and sent to the tow robot, instructing it to navigate to the area where the target fish is most likely to appear. After the tow robot arrives at the target area, it activates the ID confirmation function of the radio frequency identification module, reads the unique identification code of the passive radio frequency chip in the suspected abnormal fish, and compares it in real time with the abnormal fish ID pre-transmitted by the data processing terminal. Only when the two match completely will the tow robot activate the high-definition waterproof vision acquisition module to capture a close-up video stream or high-resolution still image of the abnormal fish.

[0029] Preferably, the customized data processing software of the data processing terminal also integrates an image management module. This module continuously monitors and receives data packets transmitted back by the tow robot after completing its task. These data packets contain compressed, encoded close-up video streams or high-resolution still images of abnormal fish captured by the robot's high-definition camera, along with an associated task identifier. The image management module first retrieves the corresponding abnormal fish tracking file from the database based on the task identifier. Then, it decodes and enhances the received video stream or image, including adaptive contrast equalization and noise reduction filtering. The processed image data is categorized according to the fish's unique chip identification code and date, and stored in a dedicated image storage warehouse. Simultaneously, the module creates an image record index within the corresponding abnormal fish tracking file. This index contains the image file's storage path, acquisition time, and shooting angle information, thereby achieving a strong association between the image data and the individual fish's identity file. Furthermore, the image management module provides a query interface, allowing users to retrieve and replay stored image data based on the chip identification code or time range, and to visualize it in a timeline or gallery format.

[0030] Preferably, the AI ​​anomaly detection algorithm module consists of a video input interface, a preprocessing submodule, a parallel feature extraction submodule, a feature fusion submodule, and a classification decision submodule connected in sequence. The video input interface is responsible for receiving the original H.264 or H.265 format video stream from the camera. The preprocessing submodule decodes the stream, converts it into an RGB image sequence, and applies scale normalization processing to each frame to adjust all images to a fixed resolution. Simultaneously, Gaussian filtering is performed to suppress water surface ripple noise; the parallel feature extraction submodule contains three independently running threads or computational kernels, which respectively execute texture and morphology feature extraction, color feature extraction, and motion feature extraction algorithms; the feature fusion submodule receives feature vectors from the three parallel threads, and uses fully connected layers or attention mechanisms to perform multimodal feature fusion to generate a unified deep feature representation vector. The classification decision submodule will The input is fed into a multi-input deep neural network model that has been loaded into memory, and forward propagation calculation is performed. Finally, the decision probability value is output. .

[0031] Preferably, the feature extraction and fusion submodule of the AI ​​anomaly detection algorithm module is specifically implemented based on a multi-branch convolutional neural network architecture; for the texture and morphological feature branches, the input is a preprocessed single-frame RGB image. This branch uses the first few layers of a lightweight convolutional neural network, such as MobileNetV2, as the backbone network to extract spatial feature maps of the image, and further generates texture morphology feature vectors through global average pooling layers. For the color feature branch, the input is also an image. This branch first converts the image from the RGB color space to the HSV color space and calculates its... Channel and The channel's two-dimensional histogram is then vectorized and encoded into a color feature vector through a fully connected network. For the motion feature branch, the input is continuous. Image sequence of frames This branch first calculates the dense optical flow field between adjacent frames, and then stacks the optical flow fields in the time dimension to form a... A tensor is input into a specially designed 3D convolutional neural network block for spatiotemporal feature extraction, and the output is a motion feature vector. The feature fusion submodule will , and Perform a concatenation operation to obtain the concatenated vector. The concatenated vector is then fed into a fusion network containing two fully connected layers. This network learns the weights of different feature modalities through training and outputs the unified deep feature representation vector. .

[0032] Preferably, the training and optimization of the multi-input deep neural network model in the classification decision submodule follows these steps: First, initialize the weights of the entire multi-branch network, where the backbone network of the texture morphology branch is loaded with pre-trained weights from the ImageNet dataset; second, normalize the training set... Data is input into the network in batches, and forward propagation is used to obtain predicted probabilities. Then, a loss function is calculated between the predicted probabilities and the true labels, using weighted cross-entropy loss. ,in It is the standard cross-entropy loss. It is a triplet loss, used to shorten the feature distance between similar samples and widen the feature distance between dissimilar samples. and The hyperparameters are balanced; then, the adaptive moment estimator optimizer Adam is used with a learning rate of For loss function Perform backpropagation to update all trainable parameters in the network; after each training epoch, use the validation set. Evaluate model performance, monitoring accuracy and recall metrics; when validation set metrics no longer improve, initiate learning rate decay or early stopping strategies; finally, select the model weights that perform best on the validation set and apply them to the independent test set. The final evaluation is conducted, and the qualified models are solidified and deployed to the AI ​​anomaly recognition algorithm module of the data processing terminal for operation.

[0033] Beneficial effects

[0034] This invention provides a system for monitoring individual fish in marine ranching cages. It has the following beneficial effects:

[0035] (1) This marine ranch cage fish individual monitoring system proposes an AI anomaly recognition technology based on multi-dimensional feature fusion and collaborative judgment to solve the problem that existing visual analysis technologies only focus on single morphological or movement features and cannot comprehensively and accurately determine the health status of individuals. The core of this technology is that the system does not analyze video images in isolation, but forcibly collects and processes three-dimensional data streams simultaneously: swimming posture and trajectory anomalies quantified by optical flow analysis based on continuous video frames; body surface color and texture anomalies extracted based on color histograms and texture analysis of single-frame images; and feeding behavior activity time-series anomalies statistically obtained by combining the working signals of the feeder within a specific spatiotemporal range. Subsequently, a deep neural network classifier designed specifically for multimodal data is used to fuse and jointly infer the above feature vectors, and finally output a comprehensive health status anomaly probability. The technical effect achieved by this technology is to transform the originally fragmented visual observation indicators into a quantifiable and comprehensive individual health risk coefficient, which significantly improves the early identification accuracy and reliability of sick, weak, or abnormal fish, and reduces missed or false alarms caused by misjudgment of single features.

[0036] (2) This marine ranch cage fish individual monitoring system proposes an event-triggered "visual perception-identity locking" linkage response technology to solve the problem of the disconnect between video surveillance and individual identity recognition systems in the existing technology, which are independent of each other and cannot immediately determine the individual's identity when abnormal behavior is detected. The core of this technology is to construct a closed-loop response logic controlled by a central processing unit: when the above-mentioned AI anomaly recognition module determines that a specific fish is abnormal, it will immediately generate a trigger command containing its precise two-dimensional pixel position and timestamp. This command is not only used for alarm, but is automatically mapped by the system to the corresponding physical feeding zone and directly drives the full-coverage RFID array deployed on the zone framework to start a high-power, high-frequency key scanning mode. The array performs intensive scanning of the water area of ​​the zone to read the unique identification code of the passive chip implanted in the identified abnormal fish. Through time and space correlation algorithms, the system binds the abnormal behavior event with the RFID code. The technical effect achieved by this technology is to realize a seamless connection and automated conversion from "discovering a fish with abnormal behavior" to "determining who it is (unique ID)". It has overcome the technical challenge of real-time identification of specific suspected individuals in intensive aquaculture environments, laying the foundation for subsequent precise management.

[0037] (3) This marine ranch cage fish individual monitoring system proposes a closed-loop operation technique of "command scheduling - movement confirmation" for individual handling, in order to solve the problem of existing monitoring systems that "emphasize monitoring but neglect handling" and the inability of early warning information to be automatically converted into precise on-site actions. The core of this technique is to introduce a towed robot with autonomous navigation capabilities as a mobile execution terminal and design an intelligent command scheduling module. This module receives two types of key command inputs from the system: one is the three-dimensional coordinates of suspected dead or unconscious floating fish reported by the sonar module; the other is the identity of the abnormal fish that has been locked and its location information. The core work of the scheduling module is to perform task planning and path calculation. It converts the above inputs into a specific and executable set of navigation commands, including the three-dimensional coordinates of the target point, the optimal travel path and the task type (such as detecting floating objects or observing specific fish at close range), and sends them to the towed robot in real time via a wireless link. The robot navigates to the target area according to the command, uses its onboard high-definition camera unit to observe, photograph or confirm at close range, and transmits the on-site image data back. The technical effect achieved by this approach is to construct a complete automated operation closed loop of "monitoring-identification-scheduling-confirmation", which dynamically extends static alarm information to precise intervention in the physical space, greatly reduces the burden of manual inspection and response delay, and realizes on-site and rapid verification and handling of abnormal events.

[0038] (4) This marine ranch cage fish individual monitoring system proposes a full life cycle data fusion management method with individual fish identity information as the core. The overall technical means of this method lies in designing and implementing an integrated architecture that runs through data perception, identity binding, action scheduling and file collection. The system requires that all collected data about the fish must be associated with a unique radio frequency identification code. Whether it is the abnormal behavior record initially judged by the AI ​​algorithm, the identity information confirmed by the radio frequency array scan, the high-definition diagnostic image transmitted back by the tow robot after scheduling, or even the record of the fish appearing in different feeding zones at different times, they will be automatically gathered, cleaned and associated by the data integration module and the image management module, and finally included in an independent individual electronic file with the chip identification code as the main key. The technical effect of this overall method is that it completely changes the fragmented mode of data unfolding around "group" or "event" in traditional aquaculture management, and for the first time constructs a full-dimensional health log for each fish that runs through its aquaculture cycle. This not only provides unprecedented data support for tracking individual health status and tracing the source of diseases, but also creates a solid data foundation for advanced management applications such as precision feeding and insurance claims based on individual differences, realizing a paradigm shift from extensive group farming to refined individual management from the perspective of information management. Attached Figure Description

[0039] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0040] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0041] Figure 1 This is a system structure diagram of the present invention;

[0042] Figure 2 This is a diagram showing the installation structure of the system of the present invention;

[0043] Figure 3 This is a data flow diagram of the abnormal fish identification and processing method of the present invention;

[0044] Figure 4This is a structural diagram of the AI ​​anomaly detection algorithm module of the present invention;

[0045] Figure 5 This is a flowchart of the AI ​​model training and deployment process of this invention.

[0046] Legend:

[0047] 1. Frame; 2. Feeder; 3. Sonar detection module; 4. Movable support; 5. Camera; 6. Radio frequency identification equipment; 7. Towing robot; 8. Data processing terminal; 9. Net cage. Detailed Implementation

[0048] 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.

[0049] like Figure 1 As shown, a marine ranching cage fish individual monitoring system includes a frame 1, which is shaped like a grid and made of 304 stainless steel. It is fixedly connected to the frame of the marine cage 9 by anti-corrosion anchor piles. The total side length of the frame 1 is 20m and the height is 0.5-0.8m. The whole is divided into four independent feeding zones (each zone has a side length of 10m). The frame 1 only blocks the spread of feed through the closed structure of the four sides and the middle partition. The bottom is completely open, which can ensure that the fish can freely enter from below the frame, while confining the floating feed within each zone.

[0050] At the midpoint of the four central partition structures of frame 1, a movable support 4 is installed. The movable support 4 is made of stainless steel, with a height of 2-2.5m and an tilt angle that can be adjusted between 0-30°. A camera 5 is fixed at the top of each movable support 4. The four sets of cameras 5 together form a visual recognition module, which is aimed at the four feeding zones of frame 1. It can capture the feeding scenes of fish in each zone. With the help of AI anomaly recognition algorithm, it can identify abnormal fish based on the characteristics of fish swimming posture, body color, activity level, etc. It can also realize the function of fish disease diagnosis through body color feature extraction and special neural network reasoning, accurately determine whether the fish is sick and the specific type of disease, and trigger the swimming robot to take pictures for evidence.

[0051] At the center of frame 1, a feeder 2 is installed. The bottom of feeder 2 is equipped with an adjustable rotating base (10-15 rpm). The rotating base adopts a sealed waterproof design. The main body of feeder 2 integrates a timed and quantitative feeding unit, which can evenly feed floating feed into four feeding zones under the drive of the rotating base. A sonar detection module 3 is also fixed above the frame of feeder 2. This module is a high-frequency side-scan sonar. The scanning parameters can cover the entire bottom area of ​​the cage 9, and the working sequence is synchronized with feeder 2.

[0052] Multiple sets of radio frequency identification devices 6 are evenly installed along the length of each central dividing structure of frame 1 to form a full-coverage radio frequency array; the radio frequency array can synchronously sense and read the passive radio frequency chip (with built-in unique identification code) implanted in the fish and upload the identification data to the data processing terminal 8.

[0053] The net cage 9 is also equipped with a tow robot 7. The tow robot 7 is made of high-strength engineering plastic and is equipped with a high-definition waterproof visual acquisition module, an RFID module, an underwater positioning module, a power propulsion unit, and a signal receiving module. Its docking point is preset inside the net cage 9. The tow robot 7 can automatically navigate to the target area after receiving commands. When it receives sonar-related commands, it uses the counting function of the RFID module to replenish the count of fish that have not yet surfaced. When it receives commands related to abnormal fish, it completes tasks such as high-definition image capture of the abnormal fish after RFID ID verification. After completing the tasks, it automatically returns to the docking point to recharge. The RFID module of the tow robot 7 has dual core functions: counting and ID verification. When it receives the location information of fish that have not surfaced from the sonar detection module, the RFID module starts the counting function to accurately count the number of fish that have not surfaced in the target area. When it receives information about abnormal fish from the data processing terminal, the module starts the ID verification function for subsequent identification of the target fish.

[0054] On the frame of the feeder 2, a data processing terminal 8 is also integrated. The data processing terminal 8 is an industrial-grade control terminal, equipped with customized data processing software and supporting 4G / 5G wireless communication. It establishes real-time signal connections with the feeder 2, radio frequency identification device 6, camera 5, sonar detection module 3, and towing robot 7, enabling real-time data transmission, equipment status monitoring, data storage, and command issuance. Finally, it generates the total number of fish in the net cage 9 and the file of abnormal fish through algorithms such as deduplication and merging, and supports data query, export, and image playback.

[0055] The data processing terminal is equipped with customized data processing software, which integrates an AI anomaly recognition algorithm module. The AI ​​anomaly recognition algorithm module establishes a continuous signal connection with the camera and receives continuous video stream data collected by cameras in each zone as input in real time. The module first performs frame-by-frame decoding and target detection on the input video stream to segment the independent image region of each fish in the image. The AI ​​algorithm module automatically performs precise processing on each segmented independent image region of the fish, extracts the color histogram of the image and converts it into a standardized body color feature vector. This body color feature vector is input into a pre-trained deep neural network for fish disease diagnosis for specialized analysis and reasoning. Combined with body color feature samples of various fish diseases (such as skin ulcers, water mold, red spot disease, etc.) covered in the training set, it accurately determines whether each fish is diseased and the specific type of disease. Subsequently, a multi-dimensional feature extraction process is executed in parallel. This process includes calculating the gray-level co-occurrence matrix of the fish image region to quantify texture features, extracting the color histogram of the image to characterize body color distribution features, and calculating the fish's centroid displacement vector field through continuous inter-frame difference and optical flow methods to quantify swimming posture and trajectory features. Simultaneously, this module combines the feeder's operating status signal to count the number of fish entering the preset feeding area below the feeder and their dwell time within a unit of time during the feeding period, generating a feeding activity time series. Finally, the AI ​​anomaly detection algorithm module inputs the extracted swimming posture feature vector, body color feature vector, and feeding activity sequence into a pre-trained multi-input deep neural network classifier for fusion analysis and joint inference. By fusing the fish disease diagnosis-specific inference results with other feature dimension information, the classifier outputs a probability value for judging the fish's health status. ,when Exceeding the preset threshold If the fish is found to be abnormal, an abnormal event record is generated, which includes the timestamp of the abnormal fish's appearance, the camera number where it is located, the pixel coordinates in the image, the disease determination result, and the suspected disease. This abnormal event record is sent to the central control unit of the data processing terminal in real time, triggering the data processing terminal to send a dispatch command to the instruction scheduling module, which dispatches the tow robot to the target location to take close-up photos of the suspected diseased fish for evidence collection, providing intuitive image support for subsequent diagnosis.

[0056] Upon receiving an anomaly event record from the AI ​​anomaly detection algorithm module, the data processing terminal immediately triggers an instruction scheduling process. The central control unit of the data processing terminal parses the camera number and pixel coordinate information in the anomaly event record and maps them to the corresponding physical feeding zone. Subsequently, the central control unit sends a control signal containing a high-power scanning mode instruction and a specific scanning frequency set to all RFID device groups responsible for that mapping zone. Upon receiving the instruction, the RFID device groups simultaneously activate a high-intensity, directional radio frequency field to perform intensive scanning of the target zone's water body. After reading the radio frequency signal carrying a globally unique identification code reflected by the passive radio frequency chip inside the abnormal fish, the RFID device groups upload a scan data packet containing the identification code, a reading timestamp, and the device's own serial number to the data integration module of the data processing terminal. The data integration module performs time and space correlation matching between the received scan data packet and the previously received anomaly event record. When the time difference between the two is... When the value is less than the set threshold and the physical location is logically consistent, the identity information of the abnormal fish is confirmed and locked, namely its unique chip identification code. The identification code is then bound to the time when it was first identified as abnormal and the partition information it is located in, creating a complete abnormal fish tracking file.

[0057] The customized data processing software of the data processing terminal integrates a data integration module; the data integration module continuously receives raw scan logs periodically uploaded from all RFID devices on the frame, the logs containing the read chip identification code, the reading time, and the device number; the data integration module first performs a data cleaning step, removing data with signal strength below a threshold. Invalid read records were then removed; subsequently, a deduplication algorithm based on time windows and device proximity was applied to the cleaned data. This algorithm is defined as follows: for the same chip identification code, within a time window... Within a given set of records reported by multiple physically adjacent RFID devices, only the record with the strongest signal strength is retained as a valid read. After deduplication, the module calculates statistics for each RFID device in each statistical period based on the preset partition affiliation information. The number of unique chip identification codes appearing in each feeding zone indicates the number of fish aggregated in that zone during that period. ,in This represents the partition index; ultimately, the data integration module, according to the time series, integrates the data from each partition. In addition, the chip identification code list is integrated to generate a structured partitioned fish distribution data table, which is stored in the terminal database and used for subsequent analysis and visualization.

[0058] The customized data processing software of the data processing terminal integrates an instruction scheduling module. The first function of the instruction scheduling module is to process the missed detection signal of the floating fish uploaded by the sonar detection module. The signal contains the depth information of the floating fish and its polar coordinate position relative to the sonar origin. This module calls the built-in three-dimensional spatial model of the net cage 9 and converts the polar coordinate position into a rectangular coordinate system inside the net cage 9 with the docking point as the origin. The system plans a collision-free underwater path from the current coordinates or default docking point coordinates of the tow robot to the target coordinates. After planning, the instruction scheduling module generates a first navigation instruction set, which includes the target coordinates and a sequence of key points along the path, and sends it to the motion controller of the tow robot via a wireless communication link. The second function of the instruction scheduling module is to respond to the abnormal fish identification information and its location partition locked by the data integration module. Based on the partition where the abnormal fish is located, the module retrieves the coordinates of the pre-stored center point or typical patrol point of that partition. Similarly, a navigation path is planned and a second navigation command set is generated and sent to the tow robot, instructing it to navigate to the area where the target fish is most likely to appear. After arriving at the target area, the tow robot first activates the ID confirmation function of the radio frequency identification module to perform a close-range radio frequency scan of the fish in the area, reads the unique identification code of the passive radio frequency chip in the suspected abnormal fish, and compares the identification code with the abnormal fish ID pre-transmitted by the data processing terminal in real time; only when the two match perfectly will the tow robot trigger subsequent actions, activate the high-definition waterproof vision acquisition module to capture a close-up video stream or high-resolution still image of the abnormal fish, ensuring that the subject of the photograph is the accurately locked abnormal fish.

[0059] The customized data processing software of the data processing terminal also integrates an image management module. This module continuously monitors and receives data packets transmitted back by the tow robot after completing its task. These data packets contain compressed, encoded close-up video streams or high-resolution still images of abnormal fish captured by the robot's high-definition camera, along with an associated task identifier. The image management module first retrieves the corresponding abnormal fish tracking file from the database based on the task identifier. Then, it decodes and enhances the received video stream or image, including adaptive contrast equalization and noise reduction filtering. The processed image data is categorized according to the fish's unique chip identification code and date, and stored in a dedicated image storage warehouse. Simultaneously, the module creates an image record index within the corresponding abnormal fish tracking file. This index contains the image file's storage path, acquisition time, and shooting angle information, thereby achieving a strong association between the image data and the individual fish's identity file. Furthermore, the image management module provides a query interface, allowing users to retrieve and replay stored image data based on the chip identification code or time range, and to visualize it in a timeline or gallery format.

[0060] The AI ​​anomaly detection algorithm module is implemented at the hardware level by a dedicated AI acceleration computing unit integrated within the data processing terminal. This computing unit includes a graphics processing unit (GPU) or a neural network processing unit (NPU). At the software architecture level, the module consists of a video input interface, a preprocessing submodule, a parallel feature extraction submodule, a feature fusion submodule, and a classification decision submodule, all connected in sequence. The video input interface receives raw H.264 or H.265 format video streams from the camera. The preprocessing submodule decodes the streams, converts them into RGB image sequences, and applies scale normalization to each frame, adjusting all images to a fixed resolution. Simultaneously, Gaussian filtering is performed to suppress water surface ripple noise; the parallel feature extraction submodule contains three independently running threads or computational kernels, which respectively execute texture and morphology feature extraction, color feature extraction, and motion feature extraction algorithms; the feature fusion submodule receives feature vectors from the three parallel threads, and uses fully connected layers or attention mechanisms to perform multimodal feature fusion to generate a unified deep feature representation vector. The classification decision submodule will The input is fed into a multi-input deep neural network model that has been loaded into memory, and forward propagation calculation is performed. Finally, the decision probability value is output. .

[0061] The training process of the multi-input deep neural network model in the AI ​​anomaly detection algorithm module relies on a pre-constructed, large-scale labeled fish behavior image dataset. The input sources for this dataset consist of two parts: the first part is video clips of normal and abnormal fish bodies collected over a long period from the system's historical video library; the second part is relevant fish images from publicly available marine biological research databases. For the first part of the video source, a semi-automatic annotation tool is used, with aquatic pathologists labeling each fish body frame-by-frame with its health status. These labels include at least "healthy," "abnormal swimming," "surface lesions," and "poor feeding." For the second part of the image source, morphological and color features are further annotated. Subsequently, data augmentation techniques are applied to all labeled data, including random horizontal and vertical flipping and random rotation angles. Brightness and contrast were randomly adjusted within the range of [0.8, 1.2], and random Gaussian noise was added to augment the original dataset. This multiplies the dataset, creating a standardized dataset for model training, validation, and testing. .

[0062] The specific implementation of the feature extraction and fusion submodule of the AI ​​anomaly detection algorithm module is based on a multi-branch convolutional neural network architecture; for the texture and morphological feature branches, the input is a preprocessed single-frame RGB image. This branch uses the first few layers of a lightweight convolutional neural network, such as MobileNetV2, as the backbone network to extract spatial feature maps of the image, and further generates texture morphology feature vectors through global average pooling layers. For the color feature branch, the input is also an image. This branch first converts the image from the RGB color space to the HSV color space and calculates its... Channel and The channel's two-dimensional histogram is then vectorized and encoded into a color feature vector through a fully connected network. For the motion feature branch, the input is continuous. Image sequence of frames This branch first calculates the dense optical flow field between adjacent frames, and then stacks the optical flow fields in the time dimension to form a... A tensor is input into a specially designed 3D convolutional neural network block for spatiotemporal feature extraction, and the output is a motion feature vector. The feature fusion submodule will , and Perform a concatenation operation to obtain the concatenated vector. The concatenated vector is then fed into a fusion network containing two fully connected layers. This network learns the weights of different feature modalities through training and outputs the unified deep feature representation vector. .

[0063] The training and optimization of the multi-input deep neural network model in the classification decision submodule follows these steps: First, initialize the weights of the entire multi-branch network, where the backbone network of the texture morphology branch is loaded with pre-trained weights from the ImageNet dataset; second, normalize the training set... Data is input into the network in batches, and forward propagation is used to obtain predicted probabilities. Then, a loss function is calculated between the predicted probabilities and the true labels, using weighted cross-entropy loss. ,in It is the standard cross-entropy loss. It is a triplet loss, used to shorten the feature distance between similar samples and widen the feature distance between dissimilar samples. and The hyperparameters are balanced; then, the adaptive moment estimator optimizer Adam is used with a learning rate of For loss function Perform backpropagation to update all trainable parameters in the network; after each training epoch, use the validation set. Evaluate model performance, monitoring accuracy and recall metrics; when validation set metrics no longer improve, initiate learning rate decay or early stopping strategies; finally, select the model weights that perform best on the validation set and apply them to the independent test set. The final evaluation is conducted, and the qualified models are solidified and deployed to the AI ​​anomaly recognition algorithm module of the data processing terminal for operation.

[0064] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A monitoring system for individual fish in marine ranching cages, characterized in that, The utility model relates to a kind of marine net cage feeding system, including: Frame (1), the frame (1) is fixedly connected with offshore net cage (9) frame by anticorrosive anchor pile, the frame (1) is divided into four independent feeding subarea; Feeding machine (2), the feeding machine (2) is fixedly installed in the center position of the frame (1), the bottom of the feeding machine (2) is equipped with adjustable rotating base of sealed waterproof; Sonar detection module (3), the sonar detection module (3) is fixedly installed on the rack of the feeding machine (2) top; Movable support (4), four, four movable supports (4) are fixedly installed in the midpoint of four center separation structure of the frame (1) one by one, the movable support (4) is stainless steel material and the adjustable angle of inclination; Camera (5), the camera (5) is fixedly installed in the top of four movable supports (4) one by one, each camera (5) is aligned with a feeding subarea of the frame (1) respectively; Radio frequency identification device (6), multiple groups of radio frequency identification device (6) are evenly installed on each center separation structure of the frame (1) along the length direction, and multiple groups of radio frequency identification device (6) form full-coverage radio frequency array on the frame (1); Towing robot (7), the towing robot (7) is configured in the offshore net cage (9), and the offshore net cage (9) is pre-provided with the stop point of the towing robot (7);The towing robot (7) is carried with radio frequency identification module with double functions of counting and ID confirmation and high-definition waterproof visual acquisition module, and its radio frequency identification module executes counting function when receiving the position information of the fish body not floating of sonar detection module, and executes ID confirmation function when receiving the abnormal fish body information of data processing terminal; Data processing terminal (8), the data processing terminal (8) is integrally installed on the rack of the feeding machine (2), and the data processing terminal (8) is respectively connected with the feeding machine (2), sonar detection module (3), camera (5), radio frequency identification device (6) and data processing terminal (8) real-time signal; And passive radio frequency chip implanted in fish body.

2. A system for monitoring individual fish in a net pen of an ocean farm according to claim 1, wherein: The data processing terminal is equipped with customized data processing software, and the customized data processing software is integrated with an AI abnormality recognition algorithm module; the AI abnormality recognition algorithm module is in continuous signal connection with the camera, and continuously receives the continuous video stream data collected by each partition camera as input in real time; the AI abnormality recognition algorithm module first decodes and detects the target of the input video stream frame by frame, and divides the independent image area of each fish body in the picture; the AI abnormality recognition algorithm module first decodes and detects the target of the input video stream frame by frame, and divides the independent image area of each fish body in the picture; then the color histogram of each fish body independent image area after segmentation is extracted and converted into a standardized body color feature vector, and the body color feature vector is input into the fish disease diagnosis deep neural network trained in advance, combined with the body color feature samples of various fish diseases in the training set, to determine whether each individual is diseased and the specific disease type; then a multi-dimensional feature extraction process is performed, which includes calculating the gray level co-occurrence matrix of the fish body image area to quantify the texture feature, extracting the color histogram of the image to represent the body color distribution feature, and calculating the fish centroid displacement vector field by continuous frame difference and optical flow method to quantify the swimming posture and trajectory feature; at the same time, the AI abnormality recognition algorithm module combines the feeding machine working state signal to count the number of fish entering the preset feeding area under the feeding machine and the residence time per unit time within the feeding period, and generates a feeding activity time sequence; finally, the AI abnormality recognition algorithm module inputs the extracted swimming posture feature vector, body color feature vector and feeding activity sequence into a pre-trained multi-input deep neural network classifier for fusion analysis and joint reasoning; the classifier outputs a judgment probability value about the fish health status When exceeds the preset threshold , it is determined that the fish body is an abnormal fish body, and an abnormal event record containing the abnormal fish body appearance timestamp, the camera number, the pixel coordinates in the picture, the disease determination result and the suspected disease is generated, which is sent to the central control unit of the data processing terminal in real time, and at the same time triggers the data processing terminal to send the instruction to the instruction scheduling module.

3. A system for monitoring individual fish in a net pen of an ocean farm according to claim 2, wherein: The data processing terminal triggers an instruction scheduling process immediately after receiving the abnormal event record from the AI anomaly identification algorithm module; the central control unit of the data processing terminal parses the camera number and pixel coordinate information in the abnormal event record and maps it to the corresponding physical feeding partition; then, the central control unit sends a control signal containing a high-power scanning mode instruction and a specific scanning frequency set to all radio frequency identification device groups responsible for the physical feeding partition; the radio frequency identification device groups that receive the instruction start a high-intensity, directional radio frequency field simultaneously and densely scan the target partition of the water body; after the radio frequency identification device groups read the radio frequency signals reflected by the passive radio frequency chips in the abnormal fish body, which carry a globally unique identification code, they upload a scanning data packet containing the identification code, the reading timestamp, and the device's own number to the data integration module of the data processing terminal; the data integration module performs time and space correlation matching on the received scanning data packet and the previously received abnormal event record, and when the time difference between the two is less than a set threshold and the physical locations are logically consistent, the identity information of the abnormal fish body, i.e., its chip unique identification code, is confirmed and locked, and the identification code is bound with the time when it was first determined to be abnormal and the partition information, creating a complete abnormal fish body tracking profile. The data processing terminal triggers an instruction scheduling process immediately after receiving the abnormal event record from the AI anomaly identification algorithm module; the central control unit of the data processing terminal parses the camera number and pixel coordinate information in the abnormal event record and maps it to the corresponding physical feeding partition; then, the central control unit sends a control signal containing a high-power scanning mode instruction and a specific scanning frequency set to all radio frequency identification device groups responsible for the physical feeding partition; the radio frequency identification device groups that receive the instruction start a high-intensity, directional radio frequency field simultaneously and densely scan the target partition of the water body; after the radio frequency identification device groups read the radio frequency signals reflected by the passive radio frequency chips in the abnormal fish body, which carry a globally unique identification code, they upload a scanning data packet containing the identification code, the reading timestamp, and the device's own number to the data integration module of the data processing terminal; the data integration module performs time and space correlation matching on the received scanning data packet and the previously received abnormal event record, and when the time difference between the two is less than a set threshold and the physical locations are logically consistent, the identity information of the abnormal fish body, i.e., its chip unique identification code, is confirmed and locked, and the identification code is bound with the time when it was first determined to be abnormal and the partition information, creating a complete abnormal fish body tracking profile.

4. The system according to claim 1, wherein the system is characterized by: The customized data processing software of the data processing terminal is integrated with a data integration module; the data integration module continuously receives original scanning logs periodically uploaded from all the RFID devices on the framework, the original scanning logs containing read chip identification codes, reading time and device number; the data integration module first performs a data cleaning step to eliminate invalid reading records with signal strength lower than a threshold ; subsequently, a deduplication algorithm based on time window and device proximity is applied to the cleaned data, the algorithm being defined as: for the same chip identification code, only the record with the strongest signal strength is kept as a valid read within a time window and reported by multiple RFID devices adjacent in physical position; after deduplication, the data integration module counts the number of unique chip identification codes appearing in each feeding partition within each statistical period according to the preset partition attribution information of each RFID device, which is the fish aggregation number of the preset partition in the period , wherein represents the partition index; finally, the data integration module integrates the and chip identification code list of each partition in time sequence to generate a structured partition fish distribution data table, which is stored in the terminal database and used for subsequent analysis and visual display.

5. The system according to claim 1, wherein: The customized data processing software of the data processing terminal is integrated with an instruction scheduling module; a first function of the instruction scheduling module is to process a missed floating fish body missed detection signal uploaded by a sonar detection module, the missed floating fish body missed detection signal containing depth information and polar coordinate position relative to a sonar origin point of the missed floating fish body; the instruction scheduling module calls a built-in net cage three-dimensional space model, converts the polar coordinate position into a rectangular coordinate system coordinate with a docking point as the origin point in the net cage and plans a collision-free underwater path from a current coordinate of the swimming robot or a default docking point coordinate to a target coordinate; after the planning is completed, the instruction scheduling module generates a first navigation instruction set, the instruction set containing the target coordinate and a path key point sequence, and transmits the instruction set to a motion controller of the swimming robot through a wireless communication link; a second function of the instruction scheduling module is to respond to abnormal fish body identity information and a partition where the abnormal fish body is located locked by a data integration module; The instruction scheduling module retrieves the pre-stored coordinates of the center point or typical patrol point of the partition where the abnormal fish body is located according to the partition Likewise, the navigation path is planned and the second set of navigation instructions is generated and sent to the swimming robot, instructing it to navigate to the area where the target fish body is most likely to appear. After the swimming robot arrives at the target area, the ID confirmation function of the radio frequency identification module is started, the unique identification code of the passive radio frequency chip in the suspected abnormal fish body is read, and a real-time comparison is made with the abnormal fish body ID pre-transmitted by the data processing terminal. Only when the two are completely matched, the swimming robot starts the high-definition waterproof visual acquisition module to shoot the close-up video stream or high-resolution static picture of the abnormal fish body.

6. A system for monitoring the individual fish in a net cage of an ocean farm according to claim 4, characterized in that: The customized data processing software of the data processing terminal is also integrated with image management module;The image management module continuously listens to and receives the data packet returned after the task is completed by the towing robot, and the data packet includes the close-up video stream or high-resolution static picture of abnormal fish body photographed by the high-definition camera carried by robot after compression encoding and the task identifier associated therewith; The image management module first retrieves the corresponding abnormal fish tracking file from the database according to the task identifier; then, the received video stream or picture is decoded and enhanced, including contrast adaptive equalization and noise reduction filtering; the processed image data is classified according to the fish chip unique identification code and date, and stored in a dedicated image storage warehouse; at the same time, the image management module creates an image record index in the corresponding abnormal fish tracking file, which contains the storage path of the image file, the collection time and the shooting angle information, thereby realizing the strong association of the image data and the fish individual identity file; in addition, the image management module provides a query interface, which supports users to search, play back and visually display the stored image data in the form of timeline or gallery according to the chip identification code or time range.

7. A system according to claim 2, wherein: The AI abnormality recognition algorithm module is composed of a video input interface, a preprocessing submodule, a parallel feature extraction submodule, a feature fusion submodule and a classification decision submodule connected in sequence; the video input interface is responsible for receiving the original H.264 or H.265 format video stream from the camera; The preprocessing submodule decodes the code stream, converts it into an RGB image sequence, and applies scale normalization to each frame of image to adjust all images to a fixed resolution Meanwhile, Gaussian filtering is performed to suppress water surface ripple noise; the parallel feature extraction submodule includes three independently running threads or computing cores that respectively execute texture and morphological feature extraction, color feature extraction, and motion feature extraction algorithms; The feature fusion submodule receives feature vectors from three parallel threads, adopts a full connection layer or an attention mechanism for multi-modal feature fusion, and generates a unified deep feature representation vector ; the classification decision submodule inputs into a multi-input deep neural network model loaded into the memory, performs forward propagation calculation, and finally outputs the decision probability value .

8. A system according to claim 7, wherein: The specific implementation of the feature extraction and fusion submodule of the AI anomaly recognition algorithm module is based on a multi-branch convolutional neural network architecture; for the texture and morphology feature branch, the input is a single frame of pre-processed RGB image The texture and morphology feature branch uses the first several layers of a lightweight convolutional neural network as the backbone network to extract the spatial feature map of the image, and further generates a texture and morphology feature vector through a global average pooling layer ; for the color feature branch, the input is also an image The image is converted from the RGB color space to the HSV color space, the two-dimensional histogram of the channel and the channel is calculated, and the histogram is vectorized and encoded into a color feature vector through a fully connected network ; for the motion feature branch, the input is a sequence of consecutive frames of images The dense optical flow field between adjacent frames is calculated, then the optical flow field is stacked in the time dimension to form a tensor, which is input into a specially designed three-dimensional convolutional neural network block for spatio-temporal feature extraction, and the motion feature vector is output ; The feature fusion sub-module performs a concatenation operation on , and to obtain a concatenated vector , and then inputs the concatenated vector into a fusion network comprising two fully connected layers, which learns the weights of different feature modalities through training and outputs the unified deep feature representation vector .

9. A system according to claim 8, wherein: The training and optimization of the multi-input deep neural network model in the classification decision sub-module follow the following steps: first, initializing the weights of the entire multi-branch network, wherein the backbone network of the texture morphology branch loads the pre-training weights on the ImageNet dataset; second, inputting the data in the standardized training set to the network in batches, and obtaining the prediction probability by forward propagation; then, calculating the loss function value between the prediction probability and the real label, wherein the loss function adopts a weighted cross-entropy loss , wherein is a standard cross-entropy loss, is a triplet loss for pulling the feature distance of the same class samples and pushing the feature distance of the different class samples, and is a balance hyperparameter; Then, the adaptive moment estimation optimizer Adam is adopted with a learning rate The loss function is back-propagated to update all trainable parameters in the network; at the end of each training epoch, the validation set is used to evaluate the model performance and monitor the accuracy and recall metrics; when the validation set metrics no longer improve, the learning rate decay or early stopping strategy is initiated; Finally, the model weight with the best performance on the validation set is selected for the final evaluation on the independent test set and the qualified model is solidified and deployed to the AI anomaly identification algorithm module of the data processing terminal for running.

Images