Deep sea cage culture environment monitoring and intelligent feeding unmanned ship system

Abstract

The invention discloses a deep sea cage culture environment monitoring and intelligent feeding unmanned ship system. The system comprises a shore-based control center, a cloud server and an unmanned ship body; the unmanned ship body comprises a ship body, a power propulsion unit, a high-precision navigation positioning module, a main controller, an energy unit, an environment sensing integration module, an intelligent feeding module and a wireless communication module; the main controller of the unmanned ship body is used for controlling the power propulsion unit to execute autonomous navigation according to the intelligent feeding decision instruction and controlling the intelligent feeding module to execute feeding operation at a designated position; and the shore-based control center is used for providing a system monitoring and human-computer interaction interface. Through cooperation of the unmanned ship, the cloud server and the shore-based control center, a complete closed loop from automatic acquisition of environment data, intelligent analysis and decision making to precise feeding execution is realized, the manual dependence and the labor intensity are remarkably reduced, and the management efficiency and the operation safety are improved.

Description

Technical Field

[0001] This application belongs to the field of deep-sea aquaculture technology, specifically involving an unmanned vessel system for monitoring the environment and intelligent feeding in deep-sea cage aquaculture. Background Technology

[0002] With the rapid development of the marine economy and the near saturation of land-based aquaculture space, deep-sea cage aquaculture has become an important direction for expanding aquaculture capacity and ensuring the supply of high-quality protein. However, the complex operating environment, low level of automation, and extensive management of deep-sea aquaculture have long constrained its large-scale and intelligent development.

[0003] Specifically, the existing technology has the following main shortcomings:

[0004] Environmental monitoring methods are outdated and one-sided: Traditional cage aquaculture relies on manual sampling by boat at fixed points or the deployment of a small number of fixed sensors. Data collection is discontinuous and spatial coverage is limited, making it difficult to timely and comprehensively grasp the spatiotemporal changes of key environmental factors such as water temperature, salinity, dissolved oxygen, and ocean currents inside and outside the cages. In particular, it is impossible to effectively obtain underwater profile information and real-time fish behavior, resulting in delayed environmental early warnings and failing to provide data support for refined management.

[0005] Feeding operations are highly dependent on manual labor, resulting in extensive and inefficient practices. Feeding is typically carried out by workers operating feeding boats based on experience, which is not only labor-intensive and risky, but also makes it difficult to precisely control the amount, location, and timing of feeding. It is easily affected by weather, sea conditions, and the condition of the personnel, often leading to overfeeding or underfeeding, causing serious feed waste, water pollution (eutrophication due to uneaten feed), and uneven fish growth, directly impacting the economic benefits and ecological sustainability of aquaculture.

[0006] Therefore, there is an urgent need for a comprehensive management system for deep-sea cage aquaculture that can achieve full-process automation and intelligence, possess high-precision autonomous operation, multi-dimensional environmental perception, intelligent decision-making and precise execution capabilities, and adapt to harsh sea conditions and communication environments. Summary of the Invention

[0007] This application provides an unmanned vessel system for environmental monitoring and intelligent feeding in deep-sea cage aquaculture, aiming to solve the problems of outdated and one-sided environmental monitoring methods and the high dependence on manual labor in feeding operations, which are extensive and inefficient.

[0008] A deep-sea cage aquaculture environmental monitoring and intelligent feeding unmanned vessel system, including:

[0009] shore-based control center, cloud server, and unmanned vessel;

[0010] The unmanned vessel body includes a hull, a power propulsion unit, a high-precision navigation and positioning module, a main controller, an energy unit, an environmental perception integration module, an intelligent feeding module, and a wireless communication module.

[0011] The cloud server is used to receive and analyze the data uploaded by the environmental perception integration module, generate intelligent feeding decision instructions, and send them to the unmanned vessel body.

[0012] The main controller of the unmanned vessel is used to control the propulsion unit to perform autonomous navigation according to the intelligent feeding decision command, and to control the intelligent feeding module to perform feeding operations at a designated location.

[0013] The shore-based control center is used to provide system monitoring and human-machine interface.

[0014] Optionally, the high-precision navigation and positioning module includes: a dual-frequency multi-mode GNSS receiver for receiving satellite positioning signals;

[0015] RTK differential signal receiving unit receives differential correction data;

[0016] Inertial measurement unit (IMU) measures the angular velocity and linear acceleration of the ship's hull.

[0017] In addition, the data fusion unit fuses the data from the GNSS receiver, the RTK differential signal receiving unit, and the inertial measurement unit through a Kalman filter, and outputs the fused position, velocity, and attitude information.

[0018] Optionally, the main controller operates an adaptive fuzzy PID controller, which is configured as follows:

[0019] The PID control parameters are dynamically adjusted based on the deviation angle between the real-time heading and the target heading of the unmanned vessel, as well as the lateral position deviation between the real-time position and the target track.

[0020] The rudder angle command is calculated based on the adjusted PID parameters, and the propulsion unit is controlled to achieve precise tracking of the flight path.

[0021] Optionally, the environmental perception integration module includes a surface unit and an underwater unit;

[0022] The above-water unit integrates a small weather station for measuring wind speed, wind direction, air temperature and humidity, and light intensity.

[0023] The underwater unit includes a liftable sensor compartment, which integrates a multi-parameter water quality monitor and / or an underwater camera.

[0024] Optionally, the lifting mechanism of the liftable sensor cabin includes:

[0025] Waterproof electric winch installed on the hull;

[0026] The carrying cable is wound up and released by the electric winch, and the carrying cable is an armored coaxial cable or a multi-core waterproof cable;

[0027] The sensor housing is connected to the end of the carrying cable;

[0028] Additionally, a pressure sensor is installed on the sensor cabin to detect the water depth at which the cabin is located.

[0029] Optionally, the intelligent feeding module includes: a storage bin;

[0030] A precision screw feeder connected to the discharge port of the storage bin, the precision screw feeder being driven by a stepper motor;

[0031] A feeding distribution mechanism used for spreading feed;

[0032] In addition, a local microcontroller is used to receive feeding instructions and control the actions of the stepper motor and the scattering mechanism;

[0033] The precision screw feeder has a calibrated correspondence model between its drive pulse frequency and feed rate.

[0034] Optionally, the cloud server is deployed with an intelligent feeding decision algorithm, the algorithm comprising:

[0035] A rule-based expert system determines whether feeding is permitted based on thresholds for water quality and environmental parameters.

[0036] Based on a statistical regression model, the basic theoretical feeding rate is calculated according to parameters such as water temperature and fish body weight.

[0037] Based on a machine learning-based image analysis model, the fish school activity video is analyzed to determine the fish school aggregation degree and the intensity of feeding competition, and the real-time gain coefficient of the basic theoretical feeding rate is output.

[0038] The intelligent feeding decision algorithm integrates the outputs of the expert system, regression model, and image analysis model to generate feeding strategy instructions that include feeding amount and feeding location.

[0039] Optionally, the wireless communication module adopts a multi-mode redundancy design, including: a 4G / 5G cellular communication module as the primary link; and a long-range data transmission radio as a backup link.

[0040] Optionally, the main controller is also configured with an emergency decision-making mechanism. When the communication interruption with the cloud server exceeds a preset time limit and reaches the predetermined feeding time window, the locally stored emergency feeding model is activated to generate and execute the emergency feeding task.

[0041] Optionally, the software interface of the shore-based control center includes:

[0042] The electronic chart and situation display area is used to display the real-time position, track, and net cage boundary of the unmanned vessel;

[0043] The data dashboard is used to dynamically display the ship's attitude, power status, and water quality parameters;

[0044] The mission planning area is used to receive the cruise route points and monitoring areas set by the operator.

[0045] In addition, there is a feeding management panel, which displays the feeding strategies issued from the cloud and provides interactive options for approval, fine-tuning, or rejection.

[0046] Compared with the prior art, this application has at least the following beneficial effects:

[0047] This application achieves a complete closed loop from automatic environmental data collection and intelligent analysis and decision-making to precise feeding execution through the collaboration of unmanned vessels, cloud servers and shore-based control centers. This significantly reduces reliance on manual labor and labor intensity, and improves management efficiency and operational safety.

[0048] This application employs a high-precision navigation and positioning module with deep coupling of RTK / IMU, combined with an adaptive fuzzy PID track tracking controller, enabling the unmanned vessel to achieve centimeter-level accuracy track tracking and stable hovering under wind, wave, and current interference, ensuring safe and reliable autonomous cruising and fixed-point operation capabilities along the edge of the cage.

[0049] The cloud-deployed hybrid intelligent feeding decision algorithm of this application integrates expert rules (to ensure safety), statistical regression models (to meet basic computational needs), and machine learning image analysis (to perceive real-time intentions). It can dynamically generate the optimal feeding strategy (feeding amount, location, and timing) based on the comprehensive environmental conditions and fish behavior, thereby achieving precise nutrition supply from the source, effectively improving feed conversion rate, and reducing waste and pollution. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the module connections of an unmanned vessel system for monitoring and intelligent feeding in deep-sea cage aquaculture, provided in one embodiment of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0052] The unmanned vessel system for deep-sea cage aquaculture environmental monitoring and intelligent feeding provided in this application includes:

[0053] shore-based control center, cloud server, and unmanned vessel;

[0054] The unmanned vessel body includes a hull, a power propulsion unit, a high-precision navigation and positioning module, a main controller, an energy unit, an environmental perception integration module, an intelligent feeding module, and a wireless communication module.

[0055] The unmanned vessel serves as the system's mobile maritime execution platform, and its specific composition is as follows:

[0056] The hull is a small waterplane area catamaran or a monohull with anti-capsulation stability design. The hull material is corrosion-resistant fiberglass or aluminum alloy, and the interior has watertight compartments for housing various electronic devices. The propulsion system includes at least two jet propellers or rudder propellers driven by brushless DC motors, symmetrically arranged at the stern or sides of the hull. Differential speed control enables the hull to move forward, backward, turn, and laterally. The propellers are powered directly from the high-voltage bus of the energy system.

[0057] The high-precision navigation and positioning module integrates a dual-frequency multi-mode GNSS receiver (supporting GPS, BeiDou, and other signals), an RTK differential signal receiving unit (which can receive differential correction data via 4G network or radio), and a nine-axis inertial measurement unit. This module communicates with the main controller via a serial port or CAN bus, and can output latitude, longitude, and altitude coordinates with centimeter-level accuracy, as well as fused three-dimensional attitude angles, heading angles, and velocity information, providing precise attitude feedback for autonomous cruise and hovering.

[0058] Specifically, the high-precision navigation and positioning module is the core unit that ensures the unmanned surface vessel (USV) can accurately track complex trajectories and maintain stable attitude control. This module employs a scheme that deeply couples real-time dynamic differential (RTK) technology with an inertial measurement unit (IMU), and specifically includes:

[0059] The RTK subsystem consists of a rover station, a base station, and a data link between them. The rover station, integrated into the unmanned surface vessel (USV), includes a dual-frequency, multi-mode GNSS receiving antenna and processing unit, capable of simultaneously receiving signals from GPS L1 / L2, BeiDou B1 / B2, and other frequencies. The base station is typically deployed at a fixed point on shore with known precise coordinates, or it connects to a regional satellite-based / network-based RTK service (such as CORS). The rover station receives carrier phase differential correction data from the base station via a 4G / 5G cellular network or a UHF data radio, and uses this data to perform real-time differential processing on the satellite signals it receives, ultimately outputting latitude, longitude, and altitude coordinates with centimeter-level accuracy (typically better than ±(10 mm + 1 ppm) for horizontal positioning). To adapt to operations in cage-like areas, the rover station antenna is typically installed near the longitudinal and transverse center of gravity of the vessel to reduce the impact of vessel sway on the antenna phase center.

[0060] The inertial measurement unit (IMU) is tightly coupled near the RTK rover station and typically includes a three-axis microelectromechanical system (MEMS) gyroscope, a three-axis accelerometer, and is integrated as a six-axis or nine-axis sensor (with an additional three-axis magnetometer). The IMU independently measures the angular velocity and linear acceleration of the ship's three axes at high frequencies (e.g., 100-200 Hz). Through inertial navigation algorithms, it provides continuous and smooth attitude (roll angle, pitch angle), heading angle, and velocity / displacement change information in a short period of time, but its positioning error accumulates over time.

[0061] Data fusion and deep coupling processing are key to the module's high performance. The module incorporates or operates a tightly coupled or deeply coupled Kalman filter via the main controller. This filter uses the high-precision position and velocity information calculated by RTK as observations to periodically correct the position, velocity, and attitude states calculated by the IMU, thereby suppressing IMU error divergence. Simultaneously, during periods when the RTK signal is lost due to bridge obstruction, multipath effects, or temporary communication link interruptions (typically several seconds to tens of seconds), the system can automatically switch to pure inertial navigation or a calculation-based navigation mode combined with odometer (if applicable), maintaining continuous and usable navigation output based on IMU data until the RTK signal is restored. This fusion mechanism ensures that navigation information has absolute accuracy at the centimeter to decimeter level for most of the time, and extremely high relative accuracy and continuity during short-term signal loss, meeting the stringent requirements for collision avoidance accuracy during autonomous cruising at close range (e.g., 2-5 meters) along the edge of the cage.

[0062] The stable output of attitude and heading benefits from the high-frequency measurement and filtering fusion of the IMU. The fused attitude information is not only used for hull stability monitoring, but also a key input to the trajectory tracking controller. The heading angle information is usually provided by a baseline vector direction composed of dual-antenna RTK to provide high-precision absolute heading, or by a high-dynamic relative heading provided by an IMU (combined with a magnetometer) after compass calibration and error compensation. The two are fused in a filter to provide the control system with stable, hysteresis-free, and high-precision heading data.

[0063] The Kalman filter that integrates RTK and IMU data is fundamentally linked to the construction of its state vector, which is crucial for ensuring the accuracy and robustness of navigation solutions. This filter typically employs an error-state Kalman filter or a direct-state Kalman filter architecture, and its complete state vector X is generally composed of the following types of state variables:

[0064] Basic navigation state variables include the unmanned surface vessel's three-dimensional position (latitude, longitude, altitude, or north, east, and celestial position) in the navigation coordinate system (usually the local northeast-sky coordinate system), three-dimensional velocity (northward velocity, eastward velocity, and celestial velocity), and three-dimensional attitude. Attitude can be represented using quaternions or Euler angles (roll angle, pitch angle, and yaw angle). These variables directly describe the vessel's spatial motion state.

[0065] Inertial sensor error state variables: To compensate for the inherent errors of the gyroscope and accelerometer in the IMU, their zero bias is incorporated into the state vector for real-time estimation, including:

[0066] Gyroscope bias: A three-dimensional vector that characterizes the constant or slowly varying output bias of the gyroscope on the three axes (unit: degrees / second or radians / second).

[0067] Accelerometer bias: a three-dimensional vector that characterizes the constant or slowly varying output deviation of the accelerometer on the three axes (unit: m / s² or g). These bias states are one of the main sources of error accumulation. Online estimation and compensation of these biases are key to suppressing divergence in pure inertial navigation solutions.

[0068] Other optional error state variables: Depending on the sensor configuration and model accuracy requirements, the state vector can be further extended, for example, to include:

[0069] IMU scaling factor error;

[0070] The carrier phase ambiguity parameter of the RTK receiver (especially important in deeply coupled models);

[0071] Clock bias in RTK receivers;

[0072] Sensor installation deviation angle (i.e., misalignment error between the IMU and the ship's coordinate system);

[0073] The time update (prediction) step of the filter is based on the angular velocity and specific force measured by the IMU. Through the inertial navigation mechanical arrangement equations, it recursively derives the basic navigation states such as position, velocity, and attitude. At the same time, based on the modeling of error states such as zero bias of the gyroscope and accelerometer (usually modeled as a first-order Gaussian-Markov process or random walk process), the covariance matrix of all state vectors is predicted.

[0074] The measurement update (correction) step of the filter is triggered when the RTK provides valid positioning and velocity solutions. At this time, the position and velocity information calculated by the RTK is used as the observation Z, which, together with the predicted corresponding navigation states (position and velocity) in the state vector, constitutes the measurement residual. The Kalman filter uses this residual and the predicted state covariance matrix to calculate the Kalman gain, and then performs optimal estimation updates on all state vectors (including the basic navigation state and all error states).

[0075] Through the aforementioned continuous "prediction-correction" cycle, the system not only outputs high-precision, high-frequency fused navigation results (position, velocity, attitude), but more importantly, it estimates and compensates for the core error (zero bias) of the IMU in real time. This online estimation of error states enables the navigation system to have adaptive correction capabilities. Even if there are individual differences in the performance of IMU devices or if they drift with temperature and time, the system can track and suppress them through periodic RTK observations, thereby maintaining high-precision navigation performance over a long period of time. This is an important algorithmic guarantee for achieving accurate tracking of complex tracks.

[0076] The main controller is an embedded core board based on a high-performance ARM Cortex-A series processor, such as an Allwinner A64 or Rockchip RK3399 chip, running a Linux operating system. This core board has multiple UART, SPI, I2C, USB, and Ethernet interfaces, responsible for running the autonomous navigation control program, sensor data acquisition and preprocessing program, communication management program, and feed command parsing program. Its navigation control program adopts a layered architecture: the upper layer generates a global path based on task planning, and the lower layer uses a controller combining feedforward and feedback. The feedback controller is preferably an adaptive fuzzy PID controller to cope with time-varying interference such as ocean currents and waves, achieving accurate tracking of the preset trajectory.

[0077] The adaptive fuzzy PID controller, as the core control algorithm for unmanned surface vessel (USV) trajectory tracking, is implemented as follows. This controller is designed to overcome the shortcomings of traditional fixed-parameter PID controllers in adapting to complex, time-varying marine environments. Its input variable is defined as the heading deviation angle. The lateral position deviation is ed. Wherein, the heading deviation angle is... The value represents the difference between the real-time heading of the unmanned vessel and the target heading of the current segment; the lateral position deviation ed is the distance between the real-time position of the unmanned vessel and the current target track line, calculated through coordinate transformation. This value has directionality (e.g., the left side of the track is set to positive and the right side to negative).

[0078] The controller performs fuzzification on the two precise input values ​​mentioned above. The fuzzification process is implemented using a membership function, which transforms the precise input values ​​into fuzzy ones. The ed value is mapped to linguistic variables such as "negative large (NB)", "negative medium (NM)", "negative small (NS)", "zero (ZO)", "positive small (PS)", "positive medium (PM)", and "positive large (PB)". The shape of the membership function (e.g., triangle, trapezoid) and its universe of discourse (i.e.) The actual range of values ​​for ed needs to be determined in advance based on the maneuverability of the unmanned vessel, typical sea conditions, and the desired control accuracy.

[0079] The core of fuzzy reasoning lies in the aforementioned fuzzy rule table. This rule table is a knowledge base summarized from expert experience and a large amount of maritime trial data, and its form is a collection of "IF-THEN" conditional statements. The rows and columns of the rule table typically correspond to... The table content shows the fuzzy linguistic values ​​of ed and ed, and the table content shows the fuzzy output of the adjustment amounts of the three parameters (Kp, Ki, Kd) of the PID. The "large amount of marine test data" specifically refers to: recording the unmanned surface vessel under various combinations of deviations (under different wind, wave, and current interference conditions). The ideal rudder angle response under ( , ed) is obtained, and the optimal PID parameter combination at this time is deduced in reverse through parameter identification or optimization algorithm, and then the fuzzy mapping relationship between the deviation state and the parameter adjustment amount is statistically summarized.

[0080] Fuzzy inference engine based on the current input The fuzzy sets of ed and the fuzzy rule table are used to calculate the fuzzy output sets of ΔKp, ΔKi, and ΔKd using a specific reasoning method (such as the Mamdani method). Subsequently, the defuzzification module (such as using the centroid method) transforms these fuzzy output sets into precise ΔKp, ΔKi, and ΔKd values.

[0081] The final online adaptive PID parameters are determined by the following formula:

[0082] Kp = Kp0 + ΔKp

[0083] Ki = Ki0 + ΔKi

[0084] Kd = Kd0 + ΔKd

[0085] Where Kp0, Ki0, and Kd0 are pre-tuned PID reference parameters. The final rudder angle command δ output by the controller is calculated using the following formula:

[0086] The feedforward compensation term can be pre-calculated based on the trajectory curvature to improve tracking performance. Through the above process, the adaptive fuzzy PID controller can dynamically and non-linearly adjust its control parameters according to the real-time heading and position deviations, thereby maintaining high-precision trajectory tracking even under wind, wave, and current interference.

[0087] The energy system includes a lithium battery pack, a battery management system, and a multi-channel isolated DC-DC power conversion module. The lithium battery pack has a nominal voltage of 48V or 72V, and its capacity is selected according to the mission's endurance requirements. The battery management system monitors the battery pack voltage, current, temperature, and remaining power in real time. The multi-channel DC-DC power conversion module converts the battery pack voltage to different levels such as +12V, +5V, and +3.3V to independently power various onboard electronic devices, ensuring power stability and safety.

[0088] The environmental sensing integrated module is specifically divided into a surface unit and an underwater unit. The surface unit is fixed to the ship's mast and integrates a small weather station, capable of measuring wind speed, wind direction, air temperature and humidity, and light intensity. The underwater unit is connected to the ship's hull via a waterproof cable. Its core is a sensor cabin that can be raised, lowered, or towed. The cabin integrates the multi-parameter water quality monitor (such as the YSI EXO series or similar products), a high-frequency acoustic Doppler profiler, and an underwater high-definition camera. The water quality monitor continuously monitors parameters such as water temperature, salinity, dissolved oxygen, pH, turbidity, chlorophyll, and blue-green algae. The high-frequency acoustic Doppler profiler is used to obtain the profile velocity and direction of the ocean currents around the net cages. The underwater high-definition camera is equipped with a supplementary light for capturing video of fish activity. All sensor data is uploaded to the main controller via a unified bus protocol based on RS-485 or Ethernet.

[0089] The mechanical structure design of the liftable sensor compartment is as follows:

[0090] The mechanism mainly consists of a lifting actuator, a support cable, a sensor cabin, and a depth detection unit. The lifting actuator is fixed to the deck or cabin of the unmanned vessel, and its core is an electric winch driven by a waterproof DC geared motor. This winch has an electromagnetic braking function to ensure that the cable position is locked in the event of a power outage. The support cable is wound around the output shaft of the winch. This cable is made of armored coaxial cable or multi-core waterproof cable, which can withstand mechanical tension and transmit power and data signals to the sensors inside the cabin.

[0091] The sensor housing is a streamlined, pressure-resistant, sealed shell, typically made of corrosion-resistant metal (such as 316L stainless steel) or engineering plastic (such as POM). The upper part of the housing is equipped with a lifting ring or quick-connect interface for reliable connection to the supporting cable. The multi-parameter water quality monitor, underwater camera, and other sensing equipment are fixed inside via mounting plates. A fairing or small stabilizing fin can be added to the exterior of the housing to reduce swaying caused by water flow impact and ensure measurement stability.

[0092] The core of the depth detection unit is a high-precision pressure sensor integrated into the sensor housing. This sensor employs a sputtered thin-film or silicon piezoresistive measurement principle, with a measurement range covering 0 to the maximum operating water depth of the unmanned vessel (e.g., 100 meters). The output signal is temperature-compensated and transmitted to the main controller on the vessel via a signal line within the support cable. The pressure value is converted into real-time water depth data based on the hydrostatic pressure principle, providing closed-loop position feedback for the main controller during lifting and lowering.

[0093] Its workflow is as follows: When the sensor cabin needs to be lowered for profile measurement, the main controller sends a command to the electric winch motor to release the cable. The sensor cabin then enters the water and sinks under its own weight. The pressure sensor provides real-time depth feedback until the preset target water depth is reached, at which point the winch brakes and stops releasing. During the constant depth measurement, the winch can be finely adjusted according to commands to compensate for the ship's heave and maintain a stable depth for the sensor cabin. After the measurement task is completed, the winch reverses to gather the cable, retrieving the sensor cabin to the fixed mounting on the ship's deck.

[0094] To ensure safety and reliability, the mechanical structure also includes the following auxiliary designs: physical markers or magnetic markers can be set on the cable as a redundancy depth reference; the winch system integrates an overload protection sensor that automatically stops and alarms when the cable tension increases abnormally (such as when it gets caught on a foreign object); a mechanical guide device can be installed between the sensor compartment and the hull to ensure accurate positioning during retrieval.

[0095] The intelligent feeding module includes a storage bin, a precision screw feeder, a spreading mechanism, and a local microcontroller. The storage bin is waterproof and has an adjustable capacity. The precision screw feeder is driven by a stepper motor, and its rotation speed and feeding rate are precisely correlated through calibration. The spreading mechanism uses a high-speed rotating centrifugal disc or an adjustable-angle pneumatic nozzle, and its throwing radius and direction can be adjusted by a servo motor. The local microcontroller receives feeding rate, feeding speed, and spreading mode commands from the main controller and performs closed-loop control of the feeding motor and the servo motor of the spreading mechanism.

[0096] The calibration of the precision screw feeder is a key pre-processing step to ensure that the intelligent feeding module can achieve accurate quantitative feeding.

[0097] The calibration system mainly consists of a calibration platform, a feeder unit to be calibrated, a high-precision weighing sensor, and a calibration control unit. The feeder unit is mounted on the calibration platform, and the high-precision weighing sensor (e.g., an electronic balance) is placed below its discharge port. The calibration control unit is communicatively connected to the feeder's stepper motor driver and the weighing sensor, and is used to send control commands and synchronously record data.

[0098] The calibration process is performed according to the following steps:

[0099] First, zero-point calibration is performed. With the feeder unloaded and stationary, the initial reading of the load cell is recorded as the zero mass point. Second, a stepped rate test is executed. The calibration control unit drives the feeder for a fixed period of time (e.g., 10 seconds) at a series of pre-set pulse frequencies covering the feeder's operating range (e.g., from a minimum speed of 100Hz to a maximum speed of 5000Hz, increasing at fixed intervals). Before each run, the load cell reading is reset to zero; after each run, a stable cumulative mass reading is recorded. This process is repeated multiple times, and the average value is taken to eliminate random errors, thus obtaining a set of raw data pairs of "pulse frequency - feed mass within time interval". Finally, data fitting and model building are performed. The obtained data pairs are processed to calculate the feed rate per unit time (mass / time) at each frequency point. Subsequently, curve fitting algorithms such as the least squares method are used, with "pulse frequency" as the independent variable and "feed rate" as the dependent variable, to fit a continuous characteristic curve or a mathematical formula (e.g., a piecewise linear function or a quadratic polynomial). This curve or formula constitutes the correspondence between "pulse frequency and feed quality" and is stored in the non-volatile memory of the local microcontroller of the intelligent feeding module or the main controller of the unmanned vessel.

[0100] In actual operation, the control system calculates the required target feed rate based on the target feed amount sent from the cloud and the allocated operation time for the current feeding task. Then, by querying or calculating the calibration model, it inversely solves for the corresponding drive pulse frequency, thereby achieving precise open-loop control of the feed amount. To further improve long-term accuracy, a periodic self-calibration mechanism can be introduced into the system. This involves using a small onboard weighing module to perform a simple check during task breaks and fine-tuning the parameters of the calibration model. The implementation of this system provides a reliable foundation for achieving precise and controllable metering of feed under dynamic sea conditions.

[0101] The wireless communication module employs a multi-mode redundancy design. The primary link is a 4G / 5G cellular communication module, used for high-bandwidth data transmission (such as images, video streams, and large amounts of sensor data) and efficient interaction with the cloud server. The backup link is a long-range data transmission radio operating at UHF or VHF, used for transmitting critical commands and status information, ensuring basic communication is maintained even when the primary link signal is weak. Both are connected to the main controller via a serial port or USB interface.

[0102] The cloud server is equipped with an aquaculture big data analysis model and an intelligent feeding decision algorithm, which are used to receive and analyze data from unmanned vessels, dynamically generate feeding strategy instructions, and issue them.

[0103] The cloud server, serving as the system's remote data center and intelligent decision-making hub, is deployed on a cloud computing platform with high availability. It continuously receives structured environmental data packets, compressed video feature data, and device status information uploaded by the unmanned vessel cluster via wireless communication modules through a public network API interface. This data is parsed and persistently stored in a time-series database and a relational database. The server core deploys a multi-model fusion intelligent feeding decision engine. This engine first performs data quality assessment and fusion, spatiotemporally aligning and normalizing real-time water quality parameters (such as dissolved oxygen and temperature), surface meteorological data, and fish activity characteristics (aggregation and activity levels) extracted from edge computing. Subsequently, the engine invokes a pre-set aquaculture knowledge base, which includes physiological models of specific aquaculture species at different growth stages, expert experience rules (such as prohibiting feeding when dissolved oxygen is below 5 mg / L), and a water temperature-feeding response surface trained based on historical big data. The decision-making algorithm integrates the current environmental state, knowledge base rules, and prediction models, with maximizing feed conversion rate as the objective function. Under the conditions of satisfying animal welfare and environmental constraints, it optimizes and calculates the key parameters for this feeding operation, including: feeding window (based on short-term weather and dissolved oxygen forecasts), feed type (such as bin numbers corresponding to different particle sizes), total feed amount, and a suggested distribution point matrix (a set of latitude and longitude coordinates used to guide the unmanned surface vessel's navigation and distribution points within the cage). The final generated feeding strategy instructions are encapsulated into a specific formatted JSON instruction package and sent to the designated unmanned surface vessel via downlink.

[0104] The shore-based control center operates on a dedicated monitoring workstation or high-performance tablet. Its software, developed based on a cross-platform framework (such as Qt), provides a graphical human-machine interface. The core interface includes: an electronic chart and situation display area, which overlays the unmanned vessel's position, track, cage boundaries, and sensor sampling points in real time; a multi-dimensional data dashboard, dynamically updating and displaying the vessel's attitude, power status, water quality parameter curves, and alarm information; a task planning area, allowing operators to set cruise path points, monitoring areas, and safety boundaries on the map by clicking or dragging; and a feeding management panel, used to visually view feeding strategy suggestions sent from the cloud, and providing interactive options for "one-click approval," "parameter fine-tuning," or "manual rejection." The software has full-duplex communication control capabilities. In automatic mode, it primarily monitors data flow and executes cloud commands. When necessary, the operator can switch to "manual remote control" mode at any time, directly taking over real-time control of the unmanned vessel's propulsion, servo motors, and feeding mechanism via a virtual joystick or command input box to achieve emergency obstacle avoidance or special operations. All interactive operations, status changes, and alarm events are logged in a local log file.

[0105] Upon receiving the feeding strategy command from the cloud, the main controller of the unmanned vessel immediately initiates the local task execution sequence. First, the navigation control submodule, based on the path points planned in the "dispensing point matrix" of the command and combined with real-time RTK / IMU data, plans the optimal operational trajectory to each dispensing point and controls the vessel to navigate autonomously. When the vessel arrives near the target dispensing point and its attitude meets the operational requirements (e.g., roll angle less than a threshold), the main controller sends a precise quantitative control command to the drive controller of the intelligent feeding module. This command includes calibrated operating parameters of the feeding motor, angle settings of the dispensing mechanism, and dispensing duration. Throughout the execution process, the controller continuously monitors the status of each mechanism to ensure that the feeding amount and location are consistent with the cloud strategy. To cope with potential temporary interruptions in maritime communication, the main controller stores a simplified emergency decision-making model (e.g., a linear feeding amount table based on the average water temperature of the last 24 hours) and a cache of the last valid cloud command in non-volatile memory. When communication is interrupted for more than the preset time limit and the scheduled feeding time window arrives, the controller automatically activates the emergency model, generates an emergency feeding task based on the cached strategy (such as feed type), and executes it. This ensures the basic continuity of aquaculture operations under extreme conditions. The execution results are then reported to the cloud after communication is restored. This mechanism significantly improves the system's robustness and operational autonomy in harsh communication environments.

[0106] The intelligent feeding decision-making algorithm, deployed on a cloud server, is a hybrid computing model that integrates deterministic rules, historical statistical patterns, and real-time biological behavior perception. Its specific implementation and collaborative working mechanism are as follows:

[0107] 1) The rule-based expert system serves as the primary decision-making layer for ensuring aquaculture safety. It incorporates a series of hard-sense rules defined by aquaculture experts' experience and industry standards. These rules are stored in a knowledge base in the form of "IF <condition> THEN <action>". Core rules include, but are not limited to: * If the current dissolved oxygen (DO) concentration is below a set threshold L_do_min (e.g., 4.0 mg / L for the target fish species), feeding is prohibited; if the water temperature exceeds the suitable growth range [T_min, T_max] for the fish species, feeding is prohibited or reduced; if harmful algal blooms (determined by the combination of chlorophyll a and specific phycocyanin concentrations) exceed the threshold, an alert is triggered and the feeding strategy is adjusted. * This system takes precedence over other models. Once a prohibitive rule is triggered, it directly rejects or suspends subsequent feeding recommendations calculated by other models, ensuring that the metabolic load on aquaculture organisms is not increased under environmental stress.

[0108] 2) Statistical regression models are responsible for calculating the theoretical feed rate to meet the basic growth requirements of fish under safe environmental conditions. This model is trained by analyzing historical aquaculture big data accumulated over a long period. Its input variables typically include: current water temperature (T), the average estimated weight of the cultured fish (W, which can be estimated based on stocking records, growth models, and acoustic size assessments), and fish age (D). The output is the percentage of body weight of the target fish species under specific T and W conditions (SFR). The model form may be a piecewise linear regression or a nonlinear function (such as a quadratic polynomial). For example, a common model expression is: SFR = a * T^2 + b * T + c * ln(W) + d, where the coefficients a, b, c, and d are determined through multiple regression analysis of historical feeding records and corresponding growth performance data. The SFR calculated by this model is the basic theoretical value for determining the amount of feed per feeding.

[0109] 3) A machine learning-based image analysis model is used to perceive the real-time feeding intentions of fish schools and dynamically fine-tune the theoretical feeding rate. This model receives pre-processed video feature data (such as the proportion of foreground target pixels and statistical features of the motion optical flow vector field) from the edge computing unit of the unmanned vessel. Using a pre-trained deep learning network (e.g., using a convolutional neural network (CNN) to classify feeding scenes on the water surface, or using an object detection network such as YOLO to count the number of fish feeding in real time), the model can quantify two key behavioral indicators: the fish school aggregation index (AI) and the feeding intensity index (FI). Subsequently, a lightweight gain coefficient prediction sub-model (such as support vector regression (SVR) or gradient boosting tree (GBT)) outputs a real-time feeding gain coefficient K (e.g., ranging from 0.5 to 1.5) based on AI and FI. When the fish school aggregation is high and feeding is aggressive, K > 1, and the feeding amount is adjusted upwards; conversely, it is adjusted downwards.

[0110] The hybrid model of the intelligent feeding decision algorithm includes the following key design details to ensure its accuracy, adaptability, and continuous optimization capabilities in actual aquaculture scenarios:

[0111] Training a regression model for feed differentiation: Different brands or formulations of feed vary in nutritional composition, particle size, water stability, and palatability, directly impacting fish feeding rates and growth. Therefore, the statistical regression model is not a single, universal model, but rather a sub-model established for each type of feed used by the system, or feed identification parameters are added to its feature inputs. During the training phase, historical aquaculture data needs to be categorized by feed type. The training data for each sub-model independently comes from feeding records during the use of that specific feed, corresponding environmental data (water temperature, dissolved oxygen), and growth performance data measured at harvest. In addition to water temperature (T) and average fish weight (W), the model's feature variables can include key nutritional parameters such as crude protein and fat content of the feed as inputs, thereby establishing a more refined "feed-environment-growth" response relationship. When switching feeds, the system calls the corresponding sub-model for calculation to ensure the accuracy of the baseline feeding rate prediction.

[0112] Robust Training and Data Augmentation of Image Analysis Models: The generalization ability of deep learning models used to identify fish gathering and feeding behavior relies on high-quality, diverse training datasets. Dataset construction requires underwater video capture from multiple fish cages, different seasons, and various weather and water quality conditions. Specifically, it needs to cover various scenes ranging from clear to highly turbid, and from strong light (midday surface) to weak light (overcast or deep water). The captured raw video frames need to be manually or semi-automatically labeled to indicate the feeding fish areas and intensity levels. To improve model robustness, systematic data augmentation operations are required on the dataset before training, including but not limited to: randomly adjusting image brightness, contrast, and saturation to simulate lighting changes; adding random Gaussian noise or simulating the blurring effect of turbid water; and randomly rotating, cropping, and scaling images. Models trained in this way can effectively resist interference from the complex optical environment of real underwater environments and stably extract behavioral features.

[0113] Online update mechanism for the gain coefficient prediction sub-model: To adapt to the unique behavioral habits of specific cage-cultured fish populations (such as feeding response speed and sensitivity to specific feeding signals), the gain coefficient prediction sub-model supports online incremental learning. The system continuously collects the input features (behavioral indices AI and FI) corresponding to each feeding event and the effect evaluation indirectly fed back through subsequent observations or growth data. Using new data within a sliding time window, the sub-model is fine-tuned periodically (e.g., weekly) or triggered (when the prediction deviation remains large). The update process can employ online learning algorithms (such as stochastic gradient descent) or periodic full retraining to ensure that the model parameters are progressively optimized as the aquaculture cycle progresses, thus better reflecting the actual behavioral patterns of the cage-cultured fish population.

[0114] Confidence assessment and manual review of hybrid model output: The system calculates a confidence score for the overall feeding amount of each decision output. This score can be based on several factors: the prediction variance of each sub-model, the consistency between the outputs of different sub-models (such as a water temperature-based regression model and an image-based model), and the similarity between the current input environment data and the distribution of historical training data. When the confidence score is lower than a preset safety threshold, it indicates that the uncertainty of the current decision is high. At this time, the system will automatically adopt a conservative strategy, such as limiting the feeding amount to the historical safety range or directly using the previous reliable feeding amount. At the same time, a high-priority alarm will be generated, and all input data, intermediate results, and low-confidence alerts for this decision will be pushed to the shore-based control center software interface for manual review and final decision by the operator, thereby minimizing the risk of the automated system.

[0115] A reinforcement learning-based long-term optimization framework: To achieve continuous self-optimization of the entire feeding decision-making system, the algorithm can integrate a reinforcement learning framework. In this framework, the agent is the hybrid feeding decision-making model, the environment is the net cage aquaculture system, the state (S) is the periodic comprehensive environmental and fish state observations, the action (A) is the feeding strategy recommended by the model, and the reward (R) is the core indicator reflecting long-term aquaculture benefits, such as the feed conversion ratio (FCR) or weight gain per unit of feed calculated at the end of a farming cycle as the main reward signal. The framework is trained offline in a simulation environment or on historical data, or through safe exploratory strategy deployment across multiple net cages, continuously trying different parameter combinations and decision rules to maximize cumulative rewards. It dynamically adjusts the rule thresholds of the expert system, the coefficients of the regression model, and the fusion weights of the image model and the gain model, enabling the entire system to autonomously evolve towards improving final economic benefits.

[0116] The hybrid decision-making process is as follows: The expert system first performs an environmental safety check; if successful, the regression model calculates the base feed amount (BaseFeed); simultaneously, the image analysis model calculates the real-time gain coefficient (K). The final recommended feed amount (FinalFeed = BaseFeed * K) is calculated. Furthermore, the algorithm considers recent feeding history (to avoid overfeeding) and weather trends (such as predictions of cooling and rainfall in the next few hours) to make minor adjustments to the FinalFeed. Finally, the decision-making algorithm generates a complete strategy instruction including the feed amount, feeding location (generated based on a heatmap of fish distribution), and recommended navigation speed.

[0117] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A deep-sea cage aquaculture environment monitoring and intelligent feeding unmanned vessel system, characterized in that, include: shore-based control center, cloud server, and unmanned vessel; The unmanned vessel body includes a hull, a power propulsion unit, a high-precision navigation and positioning module, a main controller, an energy unit, an environmental perception integration module, an intelligent feeding module, and a wireless communication module. The cloud server is used to receive and analyze the data uploaded by the environmental perception integration module, generate intelligent feeding decision instructions, and send them to the unmanned vessel body. The main controller of the unmanned vessel is used to control the propulsion unit to perform autonomous navigation according to the intelligent feeding decision command, and to control the intelligent feeding module to perform feeding operations at a designated location. The shore-based control center is used to provide system monitoring and human-machine interface.

2. The unmanned vessel system for deep-sea cage aquaculture environment monitoring and intelligent feeding according to claim 1, characterized in that, The high-precision navigation and positioning module includes: a dual-frequency multi-mode GNSS receiver, which receives satellite positioning signals; RTK differential signal receiving unit receives differential correction data; Inertial measurement unit (IMU) measures the angular velocity and linear acceleration of the ship's hull. In addition, the data fusion unit fuses the data from the GNSS receiver, the RTK differential signal receiving unit, and the inertial measurement unit through a Kalman filter, and outputs the fused position, velocity, and attitude information.

3. The unmanned vessel system for deep-sea cage aquaculture environment monitoring and intelligent feeding according to claim 1, characterized in that, The main controller operates an adaptive fuzzy PID controller, which is configured as follows: The PID control parameters are dynamically adjusted based on the deviation angle between the real-time heading and the target heading of the unmanned vessel, as well as the lateral position deviation between the real-time position and the target track. The rudder angle command is calculated based on the adjusted PID parameters, and the propulsion unit is controlled to achieve precise tracking of the flight path.

4. The unmanned vessel system for deep-sea cage aquaculture environment monitoring and intelligent feeding according to claim 1, characterized in that, The environmental perception integration module includes a surface unit and an underwater unit; The above-water unit integrates a small weather station for measuring wind speed, wind direction, air temperature and humidity, and light intensity. The underwater unit includes a liftable sensor compartment, which integrates a multi-parameter water quality monitor and / or an underwater camera.

5. The unmanned vessel system for deep-sea cage aquaculture environment monitoring and intelligent feeding according to claim 4, characterized in that, The lifting mechanism of the liftable sensor cabin includes: Waterproof electric winch installed on the hull; The carrying cable is wound up and released by the electric winch, and the carrying cable is an armored coaxial cable or a multi-core waterproof cable; The sensor housing is connected to the end of the carrying cable; Additionally, a pressure sensor is installed on the sensor cabin to detect the water depth at which the cabin is located.

6. The unmanned vessel system for deep-sea cage aquaculture environment monitoring and intelligent feeding according to claim 1, characterized in that, The intelligent feeding module includes: a storage bin; A precision screw feeder connected to the discharge port of the storage bin, the precision screw feeder being driven by a stepper motor; A feeding distribution mechanism; In addition, a local microcontroller is used to receive feeding instructions and control the actions of the stepper motor and the scattering mechanism; The precision screw feeder has a calibrated correspondence model between its drive pulse frequency and feed rate.

7. The unmanned vessel system for deep-sea cage aquaculture environment monitoring and intelligent feeding according to claim 1, characterized in that, The cloud server is equipped with an intelligent feeding decision algorithm, which includes: A rule-based expert system determines whether feeding is permitted based on thresholds for water quality and environmental parameters. Based on a statistical regression model, the basic theoretical feeding rate is calculated according to parameters such as water temperature and fish body weight. Based on a machine learning-based image analysis model, the fish school activity video is analyzed to determine the fish school aggregation degree and the intensity of feeding competition, and the real-time gain coefficient of the basic theoretical feeding rate is output. The intelligent feeding decision algorithm integrates the outputs of the expert system, regression model, and image analysis model to generate feeding strategy instructions that include feeding amount and feeding location.

8. The unmanned vessel system for deep-sea cage aquaculture environment monitoring and intelligent feeding according to claim 1, characterized in that, The wireless communication module adopts a multi-mode redundancy design, including: a 4G / 5G cellular communication module as the primary link; and a long-range data transmission radio as the backup link.

9. The unmanned vessel system for deep-sea cage aquaculture environment monitoring and intelligent feeding according to claim 1, characterized in that, The main controller is also equipped with an emergency decision-making mechanism. When the communication interruption with the cloud server exceeds the preset time limit and reaches the predetermined feeding time window, the locally stored emergency feeding model is activated to generate and execute the emergency feeding task.

10. The unmanned vessel system for deep-sea cage aquaculture environment monitoring and intelligent feeding according to claim 1, characterized in that, The software interface of the shore-based control center includes: The electronic chart and situation display area is used to display the real-time position, track, and net cage boundary of the unmanned vessel; The data dashboard is used to dynamically display the ship's attitude, power status, and water quality parameters; The mission planning area is used to receive the cruise route points and monitoring areas set by the operator. In addition, there is a feeding management panel, which displays the feeding strategies issued from the cloud and provides interactive options for approval, fine-tuning, or rejection.

Images