Intelligent feeding method and system for benthic single-looped spine eel based on multi-modal fusion

Abstract

The present application relates to the field of intelligent feeding technology for aquaculture, in particular to a benthic single-looped spoonworm intelligent feeding method and system based on multi-modal fusion, comprising: acquiring visual perception signals and bottom monitoring signals of the breeding environment; during the water static period when aeration is suspended, collecting a fluorescence image of chlorophyll-containing feed by blue light excitation combined with a red filter and calculating the feed consumption rate to obtain a basic feeding amount; fusing the bottom oxidation-reduction potential, the fluorescence method dissolved oxygen decay slope and the ammonia nitrogen ion concentration parameters to calculate the bottom carrying capacity coefficient; determining the bottom state grade through a multi-parameter weighted matrix containing safe state, pressure state and collapse state, matching the corresponding upper limit of the feeding amount, taking the smaller value of the basic feeding amount and the upper limit as the actual feeding amount and driving the feeding device to work. This method can accurately calculate the feed consumption rate and dynamically constrain the feeding amount in combination with the bottom state, which is suitable for the feeding control requirements of single-looped spoonworm benthic culture.

Description

Technical Field

[0001] This invention relates to the field of intelligent feeding technology in aquaculture, and in particular to an intelligent feeding method and system for benthic monocyclic urchin based on multimodal fusion. Background Technology

[0002] In the cultivation of benthic *Ulva unicinctus*, feeding control relies heavily on manual experience to determine the amount of feed. Feed consumption is monitored using conventional visual methods, and only single physicochemical parameters of the bottom sediment are measured. Feeding amounts are set based solely on feeding activity, without establishing a linkage mechanism between feeding status and the bottom sediment environment. Conventional image acquisition methods are easily affected by water flow, suspended impurities, and ambient light, making it difficult to accurately identify chlorophyll-containing feed. The calculated feed consumption rate exhibits significant bias, hindering the accurate determination of the basic feeding amount.

[0003] Substrate environmental monitoring only acquires parameters such as redox potential, dissolved oxygen, and ammonia nitrogen concentration individually, without integrating multiple parameters to form a quantitative coefficient characterizing the substrate health level. It also lacks a multi-state classification method for determining the substrate's safe, pressure-bearing, and collapse states. Furthermore, the feeding amount is not set with corresponding upper limits based on the substrate's carrying capacity, resulting in a mismatch between the feeding amount and the actual carrying capacity of the substrate.

[0004] This invention requires the use of targeted optical signal acquisition methods to obtain feed fluorescence images, to achieve accurate calculation of feed consumption rate, to integrate multiple types of bottom sediment monitoring parameters to calculate bottom sediment carrying capacity coefficient, to complete the bottom sediment state level determination through a multi-parameter weighted matrix, to determine the upper limit of feeding amount based on bottom sediment state, and to compare the basic feeding amount with the upper limit of feeding amount to select a value to form the final actual feeding amount adapted to aquaculture needs. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and to propose an intelligent feeding method and system for benthic single-ringed urticaria based on multimodal fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a smart feeding method for benthic *Urechis tomentosa* based on multimodal fusion, comprising: Acquire visual perception signals and bottom sediment monitoring signals of the aquaculture environment; Based on the visual perception signal, during the stillness of the water body when aeration is suspended, a fluorescent image of chlorophyll-containing feed is acquired by using a blue light excitation combined with a red filter, and the feed consumption rate per unit time is calculated based on the fluorescent image. Based on the aforementioned feed consumption rate, the basic feeding amount to meet the feeding needs of *Ulva unicornu* is calculated; Based on the sediment monitoring signals, the redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters characterizing the sediment environment are acquired simultaneously. By integrating the aforementioned redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters, the substrate bearing capacity coefficient, which reflects the health status of the substrate, is dynamically calculated. Construct a multi-parameter weighted matrix that includes safe state, pressure-bearing state and collapse state, input the bottom bearing capacity coefficient into the multi-parameter weighted matrix, and determine the state level corresponding to the current bottom state; Based on the status level, determine the maximum total feeding amount corresponding to the status level from the preset feeding amount limit strategy; The basic feeding amount is compared with the upper limit of the total feeding amount, and the smaller value is selected as the final actual feeding amount. The drive feeding device performs the feeding operation according to the final actual feeding amount.

[0007] As a further aspect of the present invention, the process of acquiring the visual perception signal and calculating the bait consumption rate includes: The aeration devices controlling the aquaculture water body are shut down to allow the water body to enter a quiescent period; During periods of water stillness, blue light sources are used to irradiate the aquaculture area to stimulate chlorophyll-containing feed to produce fluorescence. The fluorescent image is generated by capturing the fluorescence signal emitted by chlorophyll-containing bait after being excited by blue light using an image acquisition device equipped with a red filter. The fluorescence images acquired in two consecutive acquisitions are processed to extract the area of ​​the fluorescent region representing the bait or the fluorescence intensity value in the image. The change in the area or intensity of the fluorescent region per unit time is calculated, and the change is divided by the time interval to obtain the feed consumption rate.

[0008] As a further aspect of the present invention, the process of calculating the basic feeding amount to meet the feeding needs of *Urechis tomentosa* based on the feed consumption rate includes: Obtain the population estimate and basal metabolic parameters of *Ulva unicornu* corresponding to the current breeding stage; The consumption rate, the estimated number of single-ringed urticaria, and the basal metabolic parameters are input into a preset feeding demand model; The feeding requirement model outputs the total feed mass required in the next feeding cycle to compensate for the feed consumption rate and maintain the basal metabolism of *Ulva moniliforme*. This total feed mass is the basal feeding amount.

[0009] As a further aspect of the present invention, the process of acquiring the sediment monitoring signal and calculating the sediment bearing capacity coefficient includes: The real-time data stream of the redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters is continuously monitored and recorded by sensors deployed at the bottom of the aquaculture pond. The real-time data streams of the redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters are filtered and standardized to eliminate noise interference and unify the dimensions. From the processed data, feature values ​​characterizing the current instantaneous state of the sediment are extracted, including redox potential feature values, dissolved oxygen decay slope feature values, and ammonia nitrogen concentration feature values; The redox potential characteristic value, dissolved oxygen decay slope characteristic value, and ammonia nitrogen concentration characteristic value are input into the preset bearing capacity calculation model; The bearing capacity calculation model outputs a normalized value by weightedly fusing the redox potential characteristic value, dissolved oxygen decay slope characteristic value, and ammonia nitrogen concentration characteristic value. The normalized value is the substrate bearing capacity coefficient. The training and operation process of the bearing capacity calculation model includes: We collected a large amount of data covering redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters under different sediment conditions, as well as corresponding sediment carrying capacity labels determined by expert experience or water quality analysis. The redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters are used to perform supervised training on the preset machine learning model with corresponding substrate carrying capacity labels, so that the model learns to map the redox potential, dissolved oxygen decay slope, and ammonia nitrogen ion concentration parameters to the substrate carrying capacity coefficient. The trained machine learning model is deployed as the load-bearing capacity calculation model; During the model working phase, the real-time acquired redox potential characteristic values, dissolved oxygen decay slope characteristic values, and ammonia nitrogen concentration characteristic values ​​are used as inputs and fed into the bearing capacity calculation model. The bearing capacity calculation model automatically calculates and outputs the bottom bearing capacity coefficient based on the mapping relationship learned internally.

[0010] As a further aspect of the present invention, the process of constructing a multi-parameter weighted matrix and determining the state level includes: Three states corresponding to the health status of the substrate are predefined in the system: safe state, pressure-bearing state, and collapse state. The redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters are respectively set with numerical range thresholds corresponding to the safe state, the pressure state, and the collapse state. Establish a matrix structure in which rows correspond to the redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters, and columns correspond to the safe state, pressure-bearing state, and collapse state. In each cell of the matrix structure, the weight values ​​of the corresponding parameters in the corresponding states are stored, thereby forming the multi-parameter weighted matrix; The real-time acquired redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters are compared with the numerical range thresholds to determine the current safe state, pressure-bearing state, or collapse state of each parameter. Based on the current state of each parameter, the corresponding weight value is searched and extracted from the multi-parameter weighted matrix; By combining the bottom bearing capacity coefficient with the extracted parameter weight values, a comprehensive state score is obtained. Based on the interval in which the comprehensive state score falls, the state level corresponding to the current state of the sediment is determined.

[0011] As a further aspect of the present invention, the process of determining the upper limit of the total feeding amount corresponding to the state level from a preset feeding amount restriction strategy based on the state level includes: A feeding limit coefficient is pre-configured for each of the three state levels: safe state, pressure state, and collapse state. Based on the determined state level, obtain the corresponding feeding amount limit coefficient from the configuration; A baseline maximum feeding rate reference value is obtained, which is determined based on stocking density, water temperature and historical feeding data; The upper limit of the total feeding amount allowed under the current state level is calculated by multiplying the feeding amount limit coefficient by the reference value of the benchmark maximum feeding amount.

[0012] As a further aspect of the present invention, the process of comparing the basic feeding amount with the upper limit of the total feeding amount and selecting the smaller value as the final actual feeding amount includes: Receive the values ​​of the basic feeding amount and the upper limit of the total feeding amount from the calculation process; The value of the basic feeding amount is compared with the value of the upper limit of the total feeding amount in real time; If the value of the basic feeding amount is less than or equal to the value of the upper limit of the total feeding amount, then the value of the basic feeding amount is selected as the final actual feeding amount. If the value of the basic feeding amount is greater than the value of the upper limit of the total feeding amount, then the value of the upper limit of the total feeding amount is selected as the final actual feeding amount.

[0013] As a further aspect of the present invention, the process of driving the feeding device to perform the feeding operation according to the final actual feeding amount includes: The determined final actual feeding amount is converted into a control command, which includes the total mass or volume of feed to be fed. The control command is sent to the control unit of the feeding device; The control unit parses the control command and drives the actuator of the feeding device to release the corresponding mass of feed into the aquaculture water in a quantitative and targeted manner.

[0014] As a further aspect of the present invention, the process of constructing and updating the food intake requirement model includes: Data on feed consumption rate, corresponding growth data of *Ulva unicornu*, and environmental parameters were collected at different stages of the historical breeding cycle. Based on historical data, a functional relationship was established between the feed consumption rate, the estimated number of *Syntostomum unicornum*, basal metabolic parameters, and the theoretical feed requirement using regression analysis. The aforementioned functional relationship is solidified into an initial feeding requirement model; After each feeding cycle, actual feeding feedback data and growth status data of *Ulva unicornu* were obtained. The actual feeding feedback data and growth status data of *Ulva unicornu* were compared with the theoretical values ​​predicted by the feeding demand model. Based on the comparison differences, the parameters of the feeding demand model are fine-tuned using an incremental learning approach to achieve continuous optimization of the model.

[0015] As a further aspect of the present invention, the present invention also includes a smart feeding system for benthic monocyclic urchin based on multimodal fusion. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the smart feeding method for benthic monocyclic urchin based on multimodal fusion as described above.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By suspending aeration to keep the water still, a fluorescent image of chlorophyll-containing feed is acquired using blue light excitation combined with a red filter. This reduces interference from water flow and suspended impurities, minimizes the impact of ambient light on feed identification, and clearly shows the distribution and remaining status of chlorophyll-containing feed. The feed consumption rate calculated based on this is highly accurate. The basal feeding amount calculated based on the feed consumption rate closely matches the real-time feeding status of benthic *Ulva spp.*, ensuring that the determination of the basal feeding amount is consistent with the actual feeding needs.

[0017] Simultaneously collecting parameters such as redox potential, dissolved oxygen decay slope using fluorescence method, and ammonia nitrogen ion concentration in the substrate environment, and calculating the substrate carrying capacity coefficient after fusing multiple parameters, can comprehensively reflect the actual carrying capacity level of the substrate environment. Constructing a multi-parameter weighted matrix including safe state, pressure state, and collapse state, the substrate carrying capacity coefficient can be converted into the corresponding substrate state level, realizing the quantitative differentiation of substrate environment state. According to different substrate state levels, corresponding feeding amount restriction strategies can be matched to determine the upper limit of total feeding amount adapted to the current substrate conditions. The basic feeding amount and the upper limit of total feeding amount are compared numerically, and the smaller value is selected as the final actual feeding amount. This allows the determination of feeding amount to take into account both the feeding needs of benthic *Ulva spp.* and the carrying capacity of the substrate environment, keeping the feeding operation and the state of the breeding environment coordinated, avoiding the situation where the feeding amount is mismatched with the substrate carrying capacity, and making the setting of feeding amount adaptable to both biological feeding and substrate environment conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intelligent feeding method for benthic single-ringed urchin based on multimodal fusion as described in this invention; Figure 2 A flowchart for acquiring visual perception signals and calculating the consuming rate; Figure 3 This is a flowchart for acquiring bottom sediment monitoring signals and calculating the bottom sediment bearing capacity coefficient. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1 This invention provides an intelligent feeding method for benthic single-ringed urchins based on multimodal fusion, the specific method including: Visual perception signals and bottom sediment monitoring signals of the aquaculture environment were acquired. Based on the visual perception signals, during the water stasis period when aeration was suspended, fluorescent images of chlorophyll-containing feed were acquired using blue light excitation combined with a red filter, and the feed consumption rate per unit time was calculated based on the fluorescent images. The basic feeding amount to meet the feeding requirements of *Ulva unicornu* was calculated based on the feed consumption rate. Based on the bottom sediment monitoring signals, parameters characterizing the bottom sediment environment, such as redox potential, dissolved oxygen decay slope (fluorescence method), and ammonia nitrogen ion concentration, were simultaneously acquired. The redox potential, dissolved oxygen decay slope (fluorescence method), and ammonia nitrogen ion concentration parameters were integrated to dynamically calculate the bottom sediment carrying capacity coefficient, reflecting the bottom sediment health status. A multi-parameter weighted matrix including safe, pressure-bearing, and collapse states was constructed. The bottom sediment carrying capacity coefficient was input into the multi-parameter weighted matrix to determine the state level corresponding to the current bottom sediment state. Based on the state level, the upper limit of the total feeding amount corresponding to the state level was determined from a preset feeding amount restriction strategy. The basic feeding amount and the upper limit of the total feeding amount were compared, and the smaller value was selected as the final actual feeding amount. The drive feeding device performs the feeding operation according to the final actual feeding amount.

[0022] In one embodiment of the present invention, see [reference] Figure 2 The process of acquiring visual perception signals and calculating the feeding rate involves suspending the aeration device in the aquaculture water body to induce a quiescent period. During this quiescent period, a blue light source is used to irradiate the aquaculture area, stimulating chlorophyll-containing feed to fluoresce. An image acquisition device equipped with a red filter captures the fluorescence signal emitted by the chlorophyll-containing feed after blue light excitation, generating a fluorescence image. Two consecutive fluorescence images are processed to extract the area or intensity of the fluorescent region representing the feed. The change in the area or intensity of the fluorescent region per unit time is calculated, and this change is divided by the time interval to obtain the feeding rate.

[0023] In the specific implementation, visual perception signals were acquired and feed consumption rates were calculated for a single-ringed Ulva protozoan rearing pond measuring 10 meters long, 5 meters wide, and 1.2 meters deep. The feed used in the rearing pond was a formulated feed containing spirulina powder, whose chlorophyll can fluoresce under specific wavelength light excitation. The aeration device in the rearing water was kept off, and the water flow gradually weakened after the aeration device stopped. After a preset 180-second water stillness period, suspended particles in the water had mostly settled, and background interference was significantly reduced. During the water stillness period, a blue light-emitting diode array with a peak wavelength of 450 nanometers was turned on to uniformly irradiate the rearing area. The irradiation intensity of the blue light source was set to 50 lumens per square meter, and the irradiation duration was 30 seconds to fully excite the residual chlorophyll-containing feed in the water to fluoresce. Fluorescence signals are captured by an industrial camera installed above the aquaculture pond. A red bandpass filter with a center wavelength of 680 nm and a bandwidth of 20 nm is installed in front of the camera lens to filter out blue light reflection and mainly transmit the red fluorescence emitted by the feed after stimulation, generating a fluorescence image characterizing the spatial distribution of the feed. Two fluorescence images are collected and stored at two adjacent feeding monitoring times, such as 10:00 AM and 11:00 AM.

[0024] In some embodiments, the fluorescence images acquired between two consecutive acquisitions are processed. The image processing includes grayscale conversion, threshold segmentation, and connected component analysis. From the fluorescence image acquired at 10:00 AM, all pixel regions with fluorescence intensity values ​​greater than a preset threshold of 200 are extracted, and the total area of ​​these pixel regions is calculated. The unit is square pixels. The total area of ​​pixel regions satisfying the same fluorescence intensity threshold is extracted from the fluorescence image acquired at 11:00 AM. Calculate the change in the area of ​​the fluorescent region per unit time, divide the change by the time interval, and obtain the feed consumption rate. The calculation formula is as follows: in: This represents the feed consumption rate, which physically represents the decrease in the area of ​​the fluorescent region per unit time. This represents the total area of ​​the fluorescent region in the first acquired fluorescence image. This represents the total area of ​​the fluorescent region in the next adjacent fluorescence image acquisition. This represents the time interval between two image acquisitions. When... 150,000 square pixels, 120,000 square pixels, time interval The calculated bait consumption rate at 3600 seconds It is 8.33 square pixels per second.

[0025] In some embodiments, the feed rate can also be calculated using the total fluorescence intensity. The sum of the grayscale values ​​of all pixels in the fluorescence image at 10:00 AM is extracted and denoted as... Extract the sum of the gray values ​​of all pixels in the fluorescence image at 11:00 AM, and denot it as . Calculate the change in total fluorescence intensity per unit time, divide the change by the time interval, and obtain the feed rate based on fluorescence intensity. , , The unit is grayscale values ​​per second. Optionally, the image acquisition device may employ a scientific-grade complementary metal-oxide-semiconductor camera equipped with cooling elements to reduce image thermal noise during long exposures and improve the ability to detect weak fluorescence signals.

[0026] In one embodiment of the present invention, the process of calculating the basic feeding amount to meet the feeding requirements of *Urechis tomentosa* based on the feeding rate includes obtaining estimated *Urechis tomentosa* population information and basal metabolic parameters corresponding to the current breeding stage. The feeding rate, estimated *Urechis tomentosa* population information, and basal metabolic parameters are input into a preset feeding requirement model. The feeding requirement model outputs the total feed mass required to compensate for the feeding rate and maintain the basal metabolism of *Urechis tomentosa* in the next feeding cycle; this total feed mass is the basic feeding amount. The construction and updating process of the feeding requirement model includes collecting feeding rate data, corresponding *Urechis tomentosa* growth data, and environmental parameter data at different stages of historical breeding cycles. Based on historical data, a functional relationship is established between the feeding rate, estimated *Urechis tomentosa* population information, basal metabolic parameters, and the theoretically required feed amount using regression analysis. This functional relationship is solidified into the initial feeding requirement model. After each feeding cycle, actual *Urechis tomentosa* feeding feedback data and growth status data are obtained. Actual feeding feedback data and growth status data of *Ulva unicornu* were compared with theoretical values ​​predicted by the feeding demand model. Based on the differences, incremental learning was used to fine-tune the parameters of the feeding demand model to achieve continuous optimization.

[0027] In practice, the basic feeding amount to meet the feeding needs of *Urechis tomentosa* is determined based on the calculated feeding rate. Assuming a 50-square-meter rearing unit, the current feeding rate is measured to be 8.33 square pixels per second. The estimated number of *Urechis tomentosa* corresponding to the current rearing stage is obtained from initial stocking records and periodic sampling estimates; the estimated number of *Urechis tomentosa* in the current rearing unit is 5000. Basal metabolic parameters are obtained, determined by consulting the *Urechis tomentosa* physiology manual or local historical rearing data. These parameters include the daily feed mass required for basic life activities of an individual *Urechis tomentosa* at standard body weight; the currently selected basal metabolic parameter is 0.1 grams per *Urechis tomentosa* per day. The feeding rate, estimated number of *Urechis tomentosa*, and basal metabolic parameters are then input into a pre-defined feeding requirement model. The feeding requirement model outputs the total feed mass required in the next feeding cycle to compensate for the feed consumption rate and maintain the basal metabolism of *Ulva moniliformis*. This total feed mass is the basal feed amount. In a specific calculation, the basal feed amount output by the feeding requirement model... It is 1200 grams, and the calculation formula is: in: This represents the basic feeding amount, in grams; This represents the feed consumption rate compensation coefficient, used to convert the feed consumption rate unit obtained from image analysis into the actual feed mass unit; Represents the rate of bait consumption; This represents the basal metabolic weighting coefficient, used to adjust basal metabolic requirements based on factors such as water temperature. This represents an estimated population of Ulva unicinctus; This represents basal metabolic parameters.

[0028] The construction and updating of the feeding requirement model includes collecting feed consumption rate data, corresponding *Ulva pertusa* growth data, and environmental parameter data at different stages of the historical breeding cycle. Historical data comes from complete records of the past three breeding quarters, including daily feed consumption rate images, periodically measured average body weight and length data of *Ulva pertusa*, and water temperature and dissolved oxygen records. Based on historical data, a functional relationship is established between feed consumption rate, *Ulva pertusa* population prediction information, basal metabolic parameters, and theoretical feed requirements using regression analysis. In some embodiments, the regression analysis employs a multiple linear regression model, using historical feed consumption rate, historical *Ulva pertusa* population, and historical basal metabolic parameters as independent variables, and the actual amount of feed fed and confirmed as the dependent variable. The fitted functional relationship, i.e., the mathematical expression containing specific coefficients, is solidified into the initial feeding requirement model. After each feeding cycle, actual feeding feedback data and growth status data of *Ulva pertusa* were obtained. Feeding feedback data was obtained by detecting uneaten feed again the following day using fluorescence imaging and then extrapolating the feeding amount. Growth status data was obtained by periodically sampling and measuring body weight. The actual feeding feedback data and growth status data were then compared with the theoretical values ​​predicted by the feeding requirement model. It is understandable that discrepancies will arise during the comparison, such as the model-predicted baseline feeding amount. The recommended intake was 1200 grams, while the actual appropriate intake, calculated based on feeding feedback, was 1250 grams. Based on the differences, incremental learning was used to fine-tune the parameters of the feeding requirement model to achieve continuous optimization.

[0029] In some embodiments, the population prediction information of *Urechis unicinctus* can be periodically estimated through underwater video analysis combined with machine learning target detection algorithms, rather than relying solely on feeding records. The basal metabolic parameter can be set as a dynamic variable related to water temperature, for example, expressed by an empirical function related to water temperature. It is understood that the feeding demand model is not limited to a linear model; when sufficient data is available, a nonlinear feeding demand model based on neural networks can also be constructed. In specific implementations, the trigger condition for model updates can be set to automatically initiate a round of parameter fine-tuning when the absolute value of the prediction error for three consecutive feeding cycles exceeds a preset threshold. Optionally, the initial construction of the feeding demand model can also integrate different growth stages of *Urechis unicinctus* to establish staged sub-models, each with an independent parameter set.

[0030] In one embodiment of the present invention, see [reference] Figure 3The process of acquiring bottom sediment monitoring signals and calculating the bottom sediment carrying capacity coefficient involves continuously monitoring and recording real-time data streams of parameters such as oxidation-reduction potential (ORP), dissolved oxygen decay slope (fluorescence method), and ammonia nitrogen ion concentration using sensors deployed at the bottom of the aquaculture pond. The real-time data streams of these parameters are filtered and standardized to eliminate noise interference and unify dimensions. From the processed data, feature values ​​characterizing the current instantaneous state of the bottom sediment are extracted, including ORP, dissolved oxygen decay slope, and ammonia nitrogen concentration. These feature values ​​are then input into a pre-defined carrying capacity calculation model. The model weighted and fused these feature values ​​to output a normalized value, which is the bottom sediment carrying capacity coefficient.

[0031] The training and operation of the carrying capacity calculation model involves collecting a large amount of data on redox potential, dissolved oxygen decay slope (fluorescence method), and ammonia nitrogen ion concentration parameters under different sediment conditions, as well as corresponding sediment carrying capacity labels determined by expert experience or water quality analysis. The pre-set machine learning model is then supervisedly trained using these data and the corresponding sediment carrying capacity labels. This allows the model to learn the relationship between redox potential, dissolved oxygen decay slope, and ammonia nitrogen ion concentration parameters and the sediment carrying capacity coefficient. The trained machine learning model is then deployed as the carrying capacity calculation model. During the model's operation, real-time acquired redox potential, dissolved oxygen decay slope, and ammonia nitrogen concentration feature values ​​are fed into the carrying capacity calculation model. Based on the mapping relationship learned internally, the carrying capacity calculation model automatically calculates and outputs the sediment carrying capacity coefficient.

[0032] In practice, a sensor network deployed at the bottom of the aquaculture pond acquires bottom sediment monitoring signals and calculates the bottom sediment bearing capacity coefficient. The sensor network consists of redox potential electrodes, fluorescence dissolved oxygen sensors, and ammonia nitrogen ion selective electrodes. These sensors are arranged in a grid pattern on the pond bottom, spaced 2 meters apart, and installed 5 centimeters below the bottom sediment. The sensors continuously monitor and record real-time data streams of redox potential, fluorescence dissolved oxygen decay slope, and ammonia nitrogen ion concentration parameters, with a data acquisition frequency of once per minute. The real-time data streams of redox potential, fluorescence dissolved oxygen decay slope, and ammonia nitrogen ion concentration parameters are filtered and standardized. One-dimensional median filtering is used for filtering, with a sliding window width of 10 data points to eliminate impulse noise interference. Z-score standardization is used to convert the real-time data streams of redox potential, fluorescence dissolved oxygen decay slope, and ammonia nitrogen ion concentration parameters into a data sequence with a mean of 0 and a standard deviation of 1, eliminating dimensional differences. From the processed data, feature values ​​characterizing the current instantaneous state of the sediment are extracted, including redox potential feature values, dissolved oxygen decay slope feature values, and ammonia nitrogen concentration feature values. The feature value extraction operation is to take the median of the processed data sequence in the most recent hour.

[0033] The redox potential (RPP) characteristic values, dissolved oxygen decay slope characteristic values, and ammonia nitrogen concentration characteristic values ​​are input into a pre-defined carrying capacity calculation model. The model then weights and fuses these values ​​to output a normalized value, which is the sediment carrying capacity coefficient. The training and operation of the carrying capacity calculation model involves collecting a large amount of data covering RPP, dissolved oxygen decay slope (via fluorescence method), and ammonia nitrogen ion concentration parameters under different sediment conditions, as well as corresponding sediment carrying capacity labels determined by expert experience or water quality analysis. An exemplary subset of training data is shown in Table 1. Table 1: Training Data Table for Substrate Bearing Capacity Model Supervised training of a pre-defined machine learning model was conducted using data on redox potential, dissolved oxygen decay slope (via fluorescence method), and ammonia nitrogen ion concentration, along with corresponding substrate carrying capacity labels. The pre-defined machine learning model employed a three-layer feedforward neural network: three neurons in the input layer corresponding to the three features, five neurons in the hidden layer, and one neuron in the output layer corresponding to the carrying capacity coefficient. Mean squared error was used as the loss function during training. The trained machine learning model, i.e., the trained three-layer feedforward neural network and its weight parameters, was then deployed as a carrying capacity calculation model. During the model's operation, real-time acquired redox potential, dissolved oxygen decay slope, and ammonia nitrogen concentration feature values ​​were fed into the carrying capacity calculation model. Based on the internally learned mapping relationships, the carrying capacity calculation model automatically calculated and output the substrate carrying capacity coefficient. The calculation formula is as follows: in: This represents the bearing capacity coefficient of the substrate, which is a scalar value between 0 and 1. The activation function of the output layer; This represents the connection weight matrix from the hidden layer to the output layer; The activation function representing the hidden layer; This represents the connection weight matrix from the input layer to the hidden layer; This represents the input vector composed of redox potential characteristic values, dissolved oxygen decay slope characteristic values, and ammonia nitrogen concentration characteristic values; Represents the bias vector of the hidden layer; This represents the bias scalar of the output layer.

[0034] In some embodiments, feature extraction may also use the average value of data from the most recent half hour. Optionally, the standardization process may also use the Min-Max scaling method to scale the redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters to the [0,1] interval. The training data label for the carrying capacity calculation model, i.e., the substrate carrying capacity label, can be obtained by simultaneously measuring more complex indicators such as organic matter content and sulfide concentration in the substrate, and then quantifying and assigning values ​​after comprehensive evaluation by experts. It is understood that the carrying capacity calculation model is not limited to neural networks; in some alternative implementations, support vector regression or random forest regression models may also be used. In some embodiments, when preprocessing the real-time data stream of redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters collected by sensors, an outlier detection step based on time series is also added, such as using the isolated forest algorithm to identify and remove abnormal readings that deviate significantly from normal fluctuations. During the model operation phase, when the input feature value combination exceeds the coverage of the training data, the carrying capacity calculation model can output a flag to indicate high prediction uncertainty. It is understood that the substrate carrying capacity coefficient The closer the value is to 1, the better the health of the substrate and the higher its carrying capacity for feeding; the closer the value is to 0, the worse the substrate condition and the closer it is to collapse.

[0035] In one embodiment of the present invention, the process of constructing a multi-parameter weighted matrix and determining the state level includes pre-defining three states corresponding to the health status of the substrate in the system: safe state, pressure-bearing state, and collapse state. Value range thresholds corresponding to the safe state, pressure-bearing state, and collapse state are set for the parameters of redox potential, dissolved oxygen decay slope (fluorescence method), and ammonia nitrogen ion concentration, respectively. A matrix structure is established, with rows corresponding to the redox potential, dissolved oxygen decay slope (fluorescence method), and ammonia nitrogen ion concentration parameters, and columns corresponding to the safe state, pressure-bearing state, and collapse state. In each cell of the matrix structure, the pre-set weight value of the corresponding parameter in the corresponding state is stored, thus forming a multi-parameter weighted matrix. The real-time acquired redox potential, dissolved oxygen decay slope (fluorescence method), and ammonia nitrogen ion concentration parameters are compared with the value range thresholds to determine the current safe state, pressure-bearing state, or collapse state of each parameter. Based on the current state of each parameter, the corresponding weight value is searched and extracted from the multi-parameter weighted matrix. Combined with the substrate bearing capacity coefficient, the extracted parameter weight values ​​are comprehensively calculated to obtain a comprehensive state score. Based on the interval of the comprehensive state score, determine the state level corresponding to the current state of the sediment.

[0036] The process of determining the maximum total feed amount corresponding to a state level from a preset feeding limit strategy includes: pre-configuring a feeding limit coefficient for each of the three state levels: safe state, stressed state, and collapse state. Based on the determined state level, the corresponding feeding limit coefficient is retrieved from the configuration. A baseline maximum feeding reference value is obtained, determined based on stocking density, water temperature, and historical feeding data. The feeding limit coefficient is multiplied by the baseline maximum feeding reference value to calculate the maximum allowable total feed amount for the current state level.

[0037] In practical implementation, the process of constructing a multi-parameter weighted matrix and determining the state level includes pre-defining three discrete states corresponding to the health status of the substrate in the control system: safe state, pressure-bearing state, and collapse state. Threshold values ​​corresponding to the safe state, pressure-bearing state, and collapse state are set for the parameters of redox potential, dissolved oxygen decay slope (fluorescence method), and ammonia nitrogen ion concentration, respectively. For example, the threshold range for redox potential is set as follows: greater than 100 mV for safe state, between 0 and 100 mV for pressure-bearing state, and less than 0 mV for collapse state; the threshold range for dissolved oxygen decay slope (fluorescence method) is set as follows: less than 0.1 m / min for safe state, between 0.1 and 0.25 m / min for pressure-bearing state, and greater than 0.25 m / min for collapse state; the threshold range for ammonia nitrogen ion concentration is set as follows: less than 0.5 mg / L for safe state, between 0.5 and 1.2 mg / L for pressure-bearing state, and greater than 1.2 mg / L for collapse state. A matrix structure is established where rows correspond to redox potential, dissolved oxygen decay slope (based on fluorescence method), and ammonia nitrogen ion concentration parameters, and columns correspond to safe state, pressure-bearing state, and collapse state. Each cell in the matrix structure stores a pre-set weight value for the corresponding parameter in the corresponding state. For any parameter, the pre-set weight value in the safe state is higher than that in the pressure-bearing state, and the weight value in the pressure-bearing state is higher than that in the collapse state, thus forming a multi-parameter weighted matrix. A specific multi-parameter weighted matrix is ​​shown in Table 2. Table 2: Multi-parameter weighted matrix table In the specific implementation, the real-time redox potential characteristic value acquired from the sensor is 120 mV, the dissolved oxygen decay slope characteristic value is 0.08 m / min, and the ammonia nitrogen concentration characteristic value is 0.3 mg / L. The real-time acquired redox potential, dissolved oxygen decay slope (fluorescence method), and ammonia nitrogen ion concentration parameters are compared with their respective numerical range thresholds to determine the current safe, pressure-bearing, or collapse state for each parameter. In this example, the redox potential of 120 mV is greater than 100 mV, belonging to the safe state; the dissolved oxygen decay slope of 0.08 m / min is less than 0.1 m / min, belonging to the safe state; and the ammonia nitrogen concentration of 0.3 mg / L is less than 0.5 mg / L, belonging to the safe state. Based on the current state of each parameter, the corresponding weight value is retrieved and extracted from the multi-parameter weighted matrix. The redox potential belongs to the safe state, so a weight value of 0.50 is extracted; the dissolved oxygen decay slope belongs to the safe state, so a weight value of 0.30 is extracted; and the ammonia nitrogen concentration belongs to the safe state, so a weight value of 0.20 is extracted. This is combined with the substrate bearing capacity coefficient calculated from Example 3. (Assuming the weight is 0.8), the extracted parameter weights are combined to obtain a comprehensive state score. The calculation formula is: in: Represents the overall status score; Represents the bearing capacity coefficient of the substrate; This represents the weight value of the redox potential parameter in the current state; This represents the weight of the dissolved oxygen decay slope parameter in the current state. This represents the weight of the ammonia nitrogen concentration parameter in the current state. Based on the overall state score. The scoring interval determines the state level corresponding to the current state of the sediment. The scoring interval is preset as follows: Corresponding to the safe state, Corresponding to the pressure state, Corresponding to the crash state. , , , Substituting into the formula, we get Overall status score The value is 0.8, which is greater than or equal to 0.7, therefore the current sediment condition is determined to be a safe state. If the situation changes, and the ammonia nitrogen concentration increases to 1.5 mg / L (greater than 1.2 mg / L), it belongs to a collapse state, while other parameters remain unchanged. Based on the state after each parameter change, the corresponding weight value is found and extracted from the multi-parameter weighted matrix: redox potential belongs to a safe state, so a weight value of 0.50 is extracted; dissolved oxygen decay slope belongs to a safe state, so a weight value of 0.30 is extracted; ammonia nitrogen concentration belongs to a collapse state, so a weight value of 0.05 is extracted. , , , Substituting into the formula, we get Overall status score The value is 0.68, which is greater than or equal to 0.4 and less than 0.7. Therefore, the current sediment condition is determined to be a pressure state.

[0038] In some embodiments, comprehensive status score The calculation can be performed by introducing a normalized weighted sum of parameter weight values, and then combining it with the bottom bearing capacity coefficient. Multiplication. The process of determining the maximum total feed amount corresponding to the determined state level from the preset feeding limit strategy includes pre-configuring a feeding limit coefficient for each of the three state levels: safe state, stressed state, and collapse state. For example, the coefficient for safe state is 1.0, for stressed state it is 0.7, and for collapse state it is 0.3. Based on the determined state level, the corresponding feeding limit coefficient is obtained from the configuration. Since the current state level is safe, the obtained feeding limit coefficient is 1.0. A baseline maximum feeding reference value is obtained. This baseline maximum feeding reference value is determined based on stocking density, water temperature, and historical feeding data. For example, by querying a preset table, based on the current stocking density of 100 fish per square meter and a water temperature of 22 degrees Celsius, the baseline maximum feeding reference value is determined. The allowable total feeding amount is 2000 grams. Multiply the feeding limit factor by the baseline maximum feeding reference value to calculate the maximum allowable total feeding amount under the current state level. ,Right now gram.

[0039] In some embodiments, the reference value for the maximum feed rate is... The feed quantity limit can be dynamically calculated using an empirical function related to stocking density and water temperature, rather than by querying a static table. Optionally, the coefficients in the feed quantity restriction strategy can be differentiated according to different stocking stages; for example, the coefficient might be set to 0.6 during the growth stage under stress, and 0.8 during the seedling stage. It can be understood that the weight values ​​in the multi-parameter weighted matrix reflect the relative importance of different substrate parameters under different conditions. For example, under collapse conditions, the dissolved oxygen decay slope and ammonia nitrogen concentration have higher weights, reflecting that substrate deterioration is more sensitive to these two indicators at this time. In practice, the determination of the state level can be periodic, for example, performed hourly, thereby dynamically updating the total feed quantity limit.

[0040] In one embodiment of the present invention, the process of comparing the basic feeding amount with the upper limit of the total feeding amount and selecting the smaller value as the final actual feeding amount includes receiving the basic feeding amount and the upper limit of the total feeding amount from the calculation process. The basic feeding amount and the upper limit of the total feeding amount are compared in real time. If the basic feeding amount is less than or equal to the upper limit of the total feeding amount, the basic feeding amount is selected as the final actual feeding amount. If the basic feeding amount is greater than the upper limit of the total feeding amount, the upper limit of the total feeding amount is selected as the final actual feeding amount.

[0041] The process of driving the feeding device to perform the feeding operation according to the final actual feeding amount includes converting the determined final actual feeding amount into a control command. The control command includes the total mass or volume of feed to be fed. The control command is sent to the control unit of the feeding device. The control unit parses the control command and drives the actuator of the feeding device to release the corresponding mass of feed into the aquaculture water in a quantitative and targeted manner.

[0042] In practice, the system receives the baseline feeding amount and the maximum total feeding amount from the calculation process. The baseline feeding amount is derived from the output of the feeding requirement model, while the maximum total feeding amount is derived from the calculation results based on the substrate condition level. The baseline feeding amount and the maximum total feeding amount are compared in real time; this comparison is performed in the decision logic unit of the central controller. If the baseline feeding amount is less than or equal to the maximum total feeding amount, the baseline feeding amount is selected as the final actual feeding amount. If the baseline feeding amount is greater than the maximum total feeding amount, the maximum total feeding amount is selected as the final actual feeding amount. This selection logic can be expressed by the following formula: in: This represents the final actual amount of feed given. The value representing the basic feeding amount. This value represents the maximum total amount of food that can be fed. It is a function that takes the smaller of the two values.

[0043] In some embodiments, the comparison and selection process is completed within a single decision cycle. Assume the base feed amount obtained from the process described in Embodiment 2. The total feeding amount is 1200 grams, which is the upper limit of the total feeding amount obtained from the process described in Example 4. It is 2000 grams. Since 1200 is less than 2000, according to... The final actual amount of feed It was selected as 1200 grams. In another scenario, if the base feeding amount... The maximum total feeding amount is 1800 grams, but due to the substrate entering a pressurized state, the total feeding amount is limited. It was calculated as 1200 grams. Since 1800 is greater than 1200, according to... The final actual amount of feed The selected amount is 1200 grams. This comparison mechanism ensures that the final actual feeding amount will not exceed the current carrying capacity of the substrate environment, even if the feeding requirements of *Urechis unicinctus* may be higher. The process of driving the feeding device to execute the feeding operation according to the final actual feeding amount includes converting the determined final actual feeding amount into a control command. In specific implementation, the control command is generated in the form of a digital signal, containing the total mass or volume of feed to be fed, for example, the command content is "feeding mass: 1200 grams". The control command is sent to the control unit of the feeding device via an industrial fieldbus or wireless transmission network. The control unit of the feeding device parses the control command and drives the actuator of the feeding device to release the corresponding mass of feed into the aquaculture water in a quantitative and targeted manner. The actuator can be a screw conveyor; after the control unit parses the "1200 grams" command, it will drive a stepper motor to rotate the screw a predetermined number of times, thereby accurately pushing 1200 grams of feed to the feeding point.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A smart feeding method for benthic *Ulva spp.* based on multimodal fusion, characterized in that, include: Acquire visual perception signals and bottom sediment monitoring signals of the aquaculture environment; Based on the visual perception signal, during the stillness of the water body when aeration is suspended, a fluorescent image of chlorophyll-containing feed is acquired by using a blue light excitation combined with a red filter, and the feed consumption rate per unit time is calculated based on the fluorescent image. Based on the aforementioned feed consumption rate, the basic feeding amount to meet the feeding needs of *Ulva unicornu* is calculated; Based on the sediment monitoring signals, the redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters characterizing the sediment environment are acquired simultaneously. By integrating the aforementioned redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters, the substrate bearing capacity coefficient, which reflects the health status of the substrate, is dynamically calculated. Construct a multi-parameter weighted matrix that includes safe state, pressure-bearing state and collapse state, input the bottom bearing capacity coefficient into the multi-parameter weighted matrix, and determine the state level corresponding to the current bottom state; Based on the status level, determine the maximum total feeding amount corresponding to the status level from the preset feeding amount limit strategy; The basic feeding amount is compared with the upper limit of the total feeding amount, and the smaller value is selected as the final actual feeding amount. The drive feeding device performs the feeding operation according to the final actual feeding amount.

2. The intelligent feeding method for benthic single-ringed urchins based on multimodal fusion according to claim 1, characterized in that, The process of acquiring the visual perception signal and calculating the consuming rate includes: The aeration devices controlling the aquaculture water body are shut down to allow the water body to enter a quiescent period; During periods of water stillness, blue light sources are used to irradiate the aquaculture area to stimulate chlorophyll-containing feed to produce fluorescence. The fluorescent image is generated by capturing the fluorescence signal emitted by chlorophyll-containing bait after being excited by blue light using an image acquisition device equipped with a red filter. The fluorescence images acquired in two consecutive acquisitions are processed to extract the area of ​​the fluorescent region representing the bait or the fluorescence intensity value in the image. The change in the area or intensity of the fluorescent region per unit time is calculated, and the change is divided by the time interval to obtain the feed consumption rate.

3. The intelligent feeding method for benthic single-ringed urchins based on multimodal fusion according to claim 2, characterized in that, The process of calculating the basic feeding amount to meet the feeding requirements of *Urechis tomentosa* based on the aforementioned feed consumption rate includes: Obtain the population estimate and basal metabolic parameters of *Ulva unicornu* corresponding to the current breeding stage; The consumption rate, the estimated number of single-ringed urticaria, and the basal metabolic parameters are input into a preset feeding demand model; The feeding requirement model outputs the total feed mass required in the next feeding cycle to compensate for the feed consumption rate and maintain the basal metabolism of *Ulva moniliforme*. This total feed mass is the basal feeding amount.

4. The intelligent feeding method for benthic single-ringed urchins based on multimodal fusion according to claim 1, characterized in that, The process of acquiring sediment monitoring signals and calculating the sediment bearing capacity coefficient includes: The real-time data stream of the redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters is continuously monitored and recorded by sensors deployed at the bottom of the aquaculture pond. The real-time data streams of the redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters are filtered and standardized to eliminate noise interference and unify the dimensions. From the processed data, feature values ​​characterizing the current instantaneous state of the sediment are extracted, including redox potential feature values, dissolved oxygen decay slope feature values, and ammonia nitrogen concentration feature values; The redox potential characteristic value, dissolved oxygen decay slope characteristic value, and ammonia nitrogen concentration characteristic value are input into the preset bearing capacity calculation model; The bearing capacity calculation model outputs a normalized value by weightedly fusing the redox potential characteristic value, dissolved oxygen decay slope characteristic value, and ammonia nitrogen concentration characteristic value. The normalized value is the substrate bearing capacity coefficient. The training and operation process of the bearing capacity calculation model includes: We collected a large amount of data covering redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters under different sediment conditions, as well as corresponding sediment carrying capacity labels determined by expert experience or water quality analysis. The redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters are used to perform supervised training on the preset machine learning model with corresponding substrate carrying capacity labels, so that the model learns to map the redox potential, dissolved oxygen decay slope, and ammonia nitrogen ion concentration parameters to the substrate carrying capacity coefficient. The trained machine learning model is deployed as the load-bearing capacity calculation model; During the model working phase, the real-time acquired redox potential characteristic values, dissolved oxygen decay slope characteristic values, and ammonia nitrogen concentration characteristic values ​​are used as inputs and fed into the bearing capacity calculation model. The bearing capacity calculation model automatically calculates and outputs the bottom bearing capacity coefficient based on the mapping relationship learned internally.

5. The intelligent feeding method for benthic single-ringed urchins based on multimodal fusion according to claim 1, characterized in that, The process of constructing a multi-parameter weighted matrix and determining the state level includes: Three states corresponding to the health status of the substrate are predefined in the system: safe state, pressure-bearing state, and collapse state. The redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters are respectively set with numerical range thresholds corresponding to the safe state, the pressure state, and the collapse state. Establish a matrix structure in which rows correspond to the redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters, and columns correspond to the safe state, pressure-bearing state, and collapse state. In each cell of the matrix structure, the weight values ​​of the corresponding parameters in the corresponding states are stored, thereby forming the multi-parameter weighted matrix; The real-time acquired redox potential, dissolved oxygen decay slope by fluorescence method, and ammonia nitrogen ion concentration parameters are compared with the numerical range thresholds to determine the current safe state, pressure-bearing state, or collapse state of each parameter. Based on the current state of each parameter, the corresponding weight value is searched and extracted from the multi-parameter weighted matrix; By combining the bottom bearing capacity coefficient with the extracted parameter weight values, a comprehensive state score is obtained. Based on the interval in which the comprehensive state score falls, the state level corresponding to the current state of the sediment is determined.

6. The intelligent feeding method for benthic single-ringed urchins based on multimodal fusion according to claim 5, characterized in that, The process of determining the total feeding limit corresponding to the state level from the preset feeding limit strategy, based on the state level, includes: A feeding limit coefficient is pre-configured for each of the three state levels: safe state, pressure state, and collapse state. Based on the determined state level, obtain the corresponding feeding amount limit coefficient from the configuration; A baseline maximum feeding rate reference value is obtained, which is determined based on stocking density, water temperature and historical feeding data; The upper limit of the total feeding amount allowed under the current state level is calculated by multiplying the feeding amount limit coefficient by the reference value of the benchmark maximum feeding amount.

7. The intelligent feeding method for benthic single-ringed urchins based on multimodal fusion according to claim 1, characterized in that, The process of comparing the basic feeding amount with the upper limit of the total feeding amount and selecting the smaller value as the final actual feeding amount includes: Receive the values ​​of the basic feeding amount and the upper limit of the total feeding amount from the calculation process; The value of the basic feeding amount is compared with the value of the upper limit of the total feeding amount in real time; If the value of the basic feeding amount is less than or equal to the value of the upper limit of the total feeding amount, then the value of the basic feeding amount is selected as the final actual feeding amount. If the value of the basic feeding amount is greater than the value of the upper limit of the total feeding amount, then the value of the upper limit of the total feeding amount is selected as the final actual feeding amount.

8. The intelligent feeding method for benthic single-ringed urchins based on multimodal fusion according to claim 1, characterized in that, The process of driving the feeding device to perform the feeding operation according to the final actual feeding amount includes: The determined final actual feeding amount is converted into a control command, which includes the total mass or volume of feed to be fed. The control command is sent to the control unit of the feeding device; The control unit parses the control command and drives the actuator of the feeding device to release the corresponding mass of feed into the aquaculture water in a quantitative and targeted manner.

9. The intelligent feeding method for benthic single-ringed urchins based on multimodal fusion according to claim 3, characterized in that, The process of constructing and updating the feeding requirement model includes: Data on feed consumption rate, corresponding growth data of *Ulva unicornu*, and environmental parameters were collected at different stages of the historical breeding cycle. Based on historical data, a functional relationship was established between the feed consumption rate, the estimated number of *Syntostomum unicornum*, basal metabolic parameters, and the theoretical feed requirement using regression analysis. The aforementioned functional relationship is solidified into an initial feeding requirement model; After each feeding cycle, actual feeding feedback data and growth status data of *Ulva unicornu* were obtained. The actual feeding feedback data and growth status data of *Ulva unicornu* were compared with the theoretical values ​​predicted by the feeding demand model. Based on the comparison differences, the parameters of the feeding demand model are fine-tuned using an incremental learning approach to achieve continuous optimization of the model.

10. A multimodal fusion-based intelligent feeding system for benthic *Urechis unicinctus*, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent feeding method for benthic monocyclic urchin based on multimodal fusion as described in any one of claims 1 to 9.

Images