Intelligent feeding method based on fish school feeding analysis for cargo ship cabin breeding

Abstract

The present application relates to the technical field of deep-sea aquaculture, and specifically discloses an intelligent feeding method for a factory ship cabin based on fish feeding analysis, which aims to solve the problem of precise feeding in a high-density and complex underwater environment in a factory ship cabin. The present application can construct a five-layer closed-loop architecture of perception, analysis, decision-making, execution and feedback. Images are collected by a multi-water-layer underwater camera device and processed through a cascade of defogging, contrast enhancement and noise suppression. A dual-branch spatiotemporal coupling network is used to extract and fuse feeding features. A fuzzy inference output continuous feeding intensity index is obtained. An improved Dueling DQN deep reinforcement learning model is used to learn the optimal feeding strategy autonomously. Model predictive control is used to smooth the instructions. A weighing device is used to correct the feeding amount through a closed-loop PID, and adaptive and precise feeding is achieved. The present application reduces the feeding misjudgment rate, improves the feeding accuracy, reduces feed waste and improves the water quality in the cabin, and is suitable for deep-sea aquaculture factory ship cabin scenarios, thereby improving the overall aquaculture benefits.

Description

Technical Field

[0001] This invention relates to the field of deep-sea aquaculture technology, and in particular to an intelligent feeding method for shipboard aquaculture based on fish feeding analysis. Background Technology

[0002] With the rapid development of the deep-sea aquaculture industry, the containerized aquaculture vessel model has been applied on a large scale and commercially. This model is characterized by high stocking density, enclosed water bodies, three-dimensional fish distribution, strong interference from underwater suspended particles and bubbles, and complex lighting conditions, which places stringent requirements on the accuracy, real-time nature, and adaptability of feeding.

[0003] Existing feeding technologies are mostly directly adopted from cage and pond aquaculture schemes, lacking dedicated systems designed for the closed, high-density, and complex underwater environments of vessel-based aquaculture. This results in significant shortcomings in practical applications. Firstly, feeding heavily relies on manual on-site or remote observation and judgment. The ambiguity of the underwater environment easily leads to judgment delays and significant subjective biases, resulting in high labor costs and difficulty in responding to feeding dynamics in real time. Secondly, existing methods only extract spatial features from single-frame images, ignoring the temporal evolution of fish feeding patterns, making it prone to misjudging feeding intervals and terminations. Furthermore, they lack targeted enhancement for underwater image degradation, leading to distorted feature extraction. Simultaneously, feeding decisions often rely on static threshold rules, failing to adapt to dynamic changes in water temperature and fish growth, easily causing feeding imbalances. The lack of real-time weighing feedback and closed-loop correction in the feeding process results in large actual feeding deviations, further exacerbating feed waste and water quality deterioration. In summary, existing technologies are ill-suited to the complex scenarios of vessel-based aquaculture, making it difficult to achieve precise, on-demand feeding, thus becoming a key bottleneck restricting the efficient and sustainable development of deep-sea smart fisheries.

[0004] Therefore, there is an urgent need for an intelligent feeding method for shipboard aquaculture based on fish feeding analysis to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent feeding method for shipboard tankers based on fish feeding analysis, comprising the following steps: Images of fish activity and synchronized water quality parameters were collected from each aquaculture compartment of the aquaculture vessel. The images were then subjected to cascade enhancement processing, including defogging, contrast enhancement, and noise suppression, to obtain an enhanced image sequence. A dual-branch spatiotemporal coupled neural network is constructed. The enhanced image sequence is input into the spatial feature extraction branch and the temporal feature extraction branch respectively to extract the spatial distribution features and temporal evolution patterns of fish feeding behavior. The features are then fused through an adaptive gating mechanism to obtain fused spatiotemporal features. The fused spatiotemporal features are input into the fuzzy inference system, and inference is performed in combination with the cumulative food intake to output a continuous food intake intensity index; Using the feeding intensity index as the core state variable, a feeding decision model based on deep reinforcement learning is constructed, which learns the optimal feeding action autonomously through interaction with the environment. After the optimal feeding action is smoothed by the control command, a specific feeding command is generated and sent to the feeding device for execution. The actual feeding amount fed in real time by the weighing device is used for closed-loop feedback adjustment until the feeding end condition of this meal is met.

[0006] Furthermore, this invention also discloses an intelligent feeding system for shipboard aquaculture based on fish feeding analysis, comprising: The acquisition module is used to acquire images of fish activity and synchronized water quality parameters in each aquaculture cabin of the aquaculture vessel, and to perform cascade enhancement processing on the images, including defogging, contrast enhancement and noise suppression, to obtain an enhanced image sequence. The construction module is used to construct a dual-branch spatiotemporal coupled neural network. The enhanced image sequence is input into the spatial feature extraction branch and the temporal feature extraction branch respectively to extract the spatial distribution features and temporal evolution patterns of fish feeding behavior. The features are then fused through an adaptive gating mechanism to obtain fused spatiotemporal features. The output module is used to input the fused spatiotemporal features into the fuzzy inference system, combine them with the cumulative food intake for inference, and output a continuous food intake intensity index. The autonomous learning module is used to construct a feeding decision model based on deep reinforcement learning, with the feeding intensity index as the core state variable, and autonomously learn the optimal feeding action through interaction with the environment. The generation module is used to generate specific feeding instructions after the optimal feeding action is smoothed by the control instructions and then sent to the feeding device for execution. The actual feeding amount fed in real time by the weighing device is used for closed-loop feedback adjustment until the feeding end condition of the meal is met.

[0007] Furthermore, the output module includes: The building unit is used to construct the fuzzy inference system. Its input variables include the feeding activity index, the rate of change of feeding activity, and the cumulative amount of food intake obtained by fusing spatiotemporal feature mapping. The output variable is the feeding intensity index. Establish a unit to define the fuzzy membership function of each input variable and establish a fuzzy rule base; The processing unit is used to defuzzify the fuzzy inference results using the centroid method to obtain a continuous feeding intensity index.

[0008] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described intelligent feeding method for shipboard rearing based on fish feeding analysis.

[0009] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described intelligent feeding method for shipboard rearing based on fish feeding analysis.

[0010] The beneficial effects of this application are as follows: Firstly, this invention enables fully autonomous closed-loop feeding, completely replacing manual intervention. Through automatic image acquisition, intelligent feature analysis, autonomous decision-making, and closed-loop execution, it eliminates the reliance on manual observation and remote judgment, adapts to the remote and complex underwater operating environment of deep-sea engineering vessels, significantly improves the real-time performance of decision-making, significantly reduces labor costs, and avoids feeding errors caused by human subjective errors.

[0011] Secondly, this invention accurately characterizes feeding behavior through spatiotemporal coupling feature extraction, significantly reducing the misjudgment rate. A dual-branch network simultaneously extracts spatial distribution and temporal evolution features, and through adaptive gating dynamic fusion, accurately distinguishes between feeding intervals and the end of feeding, reducing the misjudgment rate of feeding intensity. The feeding rhythm is highly matched to the actual feeding needs of the fish population.

[0012] Third, this invention effectively eliminates underwater interference through cascaded image enhancement, ensuring the accuracy of feature extraction. It specifically addresses image degradation caused by suspended particles, bubbles, and uneven lighting, outputting high-definition, high signal-to-noise ratio images that fully preserve the fish's outline and feeding posture details, providing reliable data for subsequent feature extraction and avoiding decision-making biases caused by data distortion from the source.

[0013] Fourth, this invention improves feeding accuracy and water quality in the tank through closed-loop weighing correction. Real-time weighing feedback and PID dynamic correction reduce feeding deviations, avoiding inaccurate feeding caused by mechanical errors and feed characteristics, reducing uneaten feed accumulation, lowering the risk of water quality deterioration, adapting to the closed water environment of tank aquaculture, and facilitating the sustainable development of high-density aquaculture. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of a method flow proposed in an embodiment of this application.

[0015] Figure 2 This is a schematic diagram of the system structure proposed in an embodiment of the present invention.

[0016] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0018] like Figure 1As shown, this application provides an intelligent feeding method for shipboard tankers based on fish feeding analysis, including the following steps: Step S1: Underwater cameras deployed at different water levels in each aquaculture tank of the aquaculture vessel collect images of fish activity and synchronize water quality parameters, and perform cascade enhancement processing on the original images, including defogging, contrast enhancement and noise suppression, to obtain an enhanced image sequence; Step S2: Construct a dual-branch spatiotemporal coupled neural network, inputting the enhanced image sequence into the spatial feature extraction branch and the temporal feature extraction branch of the bidirectional long short-term memory network based on the attention mechanism, respectively, to extract the spatial distribution features and temporal evolution patterns of the fish feeding behavior, and perform feature fusion through an adaptive gating mechanism to obtain fused spatiotemporal features; Step S3: Input the fused spatiotemporal features into the fuzzy inference system. The input variables of the system include the feeding activity index, the rate of change of feeding activity, and the cumulative feeding amount fed back by the weighing device. After fuzzy logic inference and defuzzification, the system outputs a continuous feeding intensity index. Step S4: Using the feeding intensity index as the core state variable, construct a deep reinforcement learning feeding decision model based on the improved Dueling DQN, design a multi-objective reward function that includes feeding response reward, economic reward, growth efficiency reward and negative reward, and autonomously learn the optimal feeding action through continuous interaction with the breeding environment; Step S5: After smoothing the optimal feeding action output by the deep reinforcement learning model through model predictive control, a specific feeding instruction containing the feeding amount, feeding duration and feeding interval is generated and sent to the feeding device equipped with a weighing device for execution. The actual feeding amount fed in real time by the weighing device is used for closed-loop PID adjustment and correction until the feeding end condition of this meal is met.

[0019] As described in steps S1-S5 above, aquaculture vessel rearing is a high-density, enclosed, deep-sea aquaculture model with a stocking density of 20-40 kg / m³. Water exchange is limited, suspended particulate matter and air bubbles cause significant interference, and lighting conditions are complex. Fish activity is three-dimensionally distributed, and feeding behavior is dynamically changing. In this scenario, traditional feeding methods rely on manual on-site observation or remote video judgment, which is easily affected by underwater environmental interference, leading to inaccurate judgments. Furthermore, they lack real-time data support and dynamic adjustment mechanisms. Simple automated feeding technologies suffer from poor image quality, limited feeding feature extraction, static and fixed decision rules, and lack of closed-loop feedback correction. This directly leads to a series of technical pain points, such as excessive feed waste, water quality deterioration due to uneaten feed accumulation, slow or uneven fish growth due to insufficient feeding, and untimely feeding response due to delayed feeding. Therefore, it is essential to construct a closed-loop intelligent feeding system from data collection to execution feedback to adapt to the complex dynamic scenarios of vessel rearing and achieve precise, on-demand feeding.

[0020] Traditional canopy feeding technology for fish in water tanks suffers from multiple systemic defects. Image acquisition is often limited to single-layer, fixed-point acquisition with simplistic processing methods, failing to address underwater fuzziness and noise interference. Feature extraction relies solely on single-frame spatial features, neglecting the evolutionary patterns of feeding time and easily misinterpreting feeding intervals as the end of feeding. Decision-making uses static thresholds or simple models, unable to adapt to dynamic changes in fish behavior, water temperature, and dissolved oxygen, resulting in delayed and biased decisions. Execution lacks real-time weighing feedback, leading to large and uncorrectable deviations in feeding amounts. The overall technology chain is fragmented and lacks intelligence, making it difficult to meet the precise feeding requirements of high-density canopy rearing. This technical solution, centered on closed-loop control, integrates five core modules: image enhancement, spatiotemporal feature fusion, fuzzy quantization, reinforcement learning decision-making, and closed-loop execution correction. It overcomes traditional technical bottlenecks across all dimensions—data source, feature dimensions, decision logic, and execution accuracy—forming a complete intelligent feeding technology closed loop and achieving full-process adaptive optimization.

[0021] The overall goal of this invention is to construct a five-layer closed-loop control architecture encompassing perception, analysis, decision-making, execution, and feedback. Through underwater image cascade enhancement, dual-branch spatiotemporal feature coupling extraction, fuzzy inference quantification of feeding intensity, improved deep reinforcement learning autonomous decision-making, and full-link collaboration of model predictive control and weighing closed-loop correction, this system systematically solves the problem of precise feeding in high-density, enclosed, turbid aquaculture scenarios with three-dimensional fish distribution on aquaculture vessels. It replaces traditional manual experience-based and static rule-based feeding, achieving intelligent, adaptive, closed-loop, and precise feeding processes. Ultimately, this aims to reduce feed waste, improve aquaculture water quality, enhance fish growth uniformity, and increase aquaculture economic benefits.

[0022] This invention relies on a five-layer closed-loop control logic of perception, analysis, decision-making, execution, and feedback to form a complete intelligent feeding workflow: First, images of fish schools and water quality parameters such as water temperature and dissolved oxygen are simultaneously acquired through a multi-layer underwater camera device in the aquaculture tank. After cascaded processing including dark channel prior defogging, adaptive histogram equalization contrast enhancement, and non-local mean filtering noise suppression, underwater image degradation interference is eliminated, outputting a high-quality enhanced image sequence. Subsequently, the image sequence is input into a dual-branch spatiotemporal coupled neural network to extract feeding spatial distribution features and temporal evolution patterns in parallel. Through an adaptive gating mechanism, the two types of features are dynamically fused to obtain a fused spatiotemporal feature comprehensively representing feeding behavior. Then, the fused spatiotemporal feature is input into a three-input fuzzy inference system, which, combined with the cumulative feeding amount, completes fuzzy inference and defuzzification, outputting a continuous feeding intensity index in the range of 0 to 100. Finally, an improved Dueling system is constructed using the feeding intensity index as the core. The DQN reinforcement learning model, relying on a multi-objective reward function, guides the model to continuously interact with the aquaculture environment, autonomously learns the optimal feeding action, and then uses model prediction control to smooth feeding instructions, which are then sent to the feeding device for execution. The actual feeding amount is fed in real time using a weighing device, and feeding deviations are corrected through closed-loop PID adjustment. At the same time, the cumulative feeding amount is fed back to the fuzzy inference system to form a data closed loop. The above process is executed cyclically until multiple consecutive decisions are made not to feed and the feeding intensity index is lower than the threshold, at which point the feeding session is considered over. The entire process adapts in real time to the feeding dynamics of the fish and changes in the aquaculture environment, realizing closed-loop, adaptive, precise and intelligent feeding in the context of shipboard aquaculture.

[0023] In one embodiment, step S1, which involves acquiring images of fish activity and synchronizing water quality parameters, and performing cascaded enhancement processing on the original images, further includes: S11: Deploy underwater high-definition cameras at 0.5 to 1 meter below the surface, 3 to 5 meters below the middle layer, and 1 to 2 meters below the bottom of each aquaculture tank to continuously collect images of fish activity at a sampling frequency of 10 to 15 frames per second, and simultaneously collect water temperature and dissolved oxygen water quality parameters of the aquaculture tank. S12: The dark channel prior algorithm is used to dehaze the acquired raw underwater image to eliminate the image scattering and blurring effects caused by suspended particles in the water and obtain a dehazed image. S13: Adaptive histogram equalization is applied to the dehazed image to enhance local contrast. The sliding window size of the histogram equalization is adaptively adjusted according to the fish density in the current image. The specific adjustment method is as follows: First, a minimum window size and a maximum window size are preset. Then, an empirical scaling factor is calibrated according to the turbidity of the aquaculture water. The value of this factor ranges from 0.5 to 2.0. Next, the fish density in the current image is calculated, which is the ratio of the total number of fish pixels to the total number of pixels in the image. A small constant greater than zero is added to obtain the adjustment denominator. The empirical scaling factor is divided by the adjustment denominator to obtain an intermediate value. Then, the larger of the minimum window size and the intermediate value is taken. Then, the smaller of the maximum window size and the aforementioned larger value is taken. Finally, the smaller value is determined as the sliding window size. S14: After contrast enhancement, a non-local mean filtering algorithm is used to suppress noise in order to eliminate salt and pepper noise and Gaussian noise caused by bubbles and suspended particles generated by feeding, and finally output the enhanced image sequence.

[0024] As described in steps S11-S14 above, this invention addresses the complex underwater environment of high-density aquaculture incubators, turbid water, and uneven lighting. It enables the simultaneous and accurate acquisition of multi-layered fish activity images and key water quality parameters. Through cascaded enhancement processing of defogging, contrast enhancement, and noise suppression, it systematically eliminates degradation problems such as underwater image scattering blur, insufficient contrast, and noise interference, outputting enhanced image sequences with high signal-to-noise ratio and high definition. This provides a reliable image data foundation for subsequent dual-branch spatiotemporal feature extraction, ensuring the accuracy of fish feeding behavior analysis from the source.

[0025] In aquaculture vessel rearing scenarios, stocking densities can reach 20-40 kg / m³, with fish distributed in a three-dimensional manner. The presence of suspended particles and feeding bubbles in the water, coupled with complex lighting conditions due to uneven reflections from the vessel walls and artificial light sources, leads to problems such as scattering blurring, low local contrast, superposition of salt-and-pepper noise and Gaussian noise, and obscuring of fish details in raw underwater images. Directly using raw images for fish feeding feature extraction can cause deviations in feeding posture recognition and distortion of spatial distribution feature extraction, resulting in incorrect judgments of feeding intensity and an inability to support precise feeding decisions. Therefore, it is essential to first achieve accurate underwater image acquisition and targeted enhancement processing to avoid subsequent analysis errors caused by image quality defects at the data source.

[0026] Traditional image acquisition and processing techniques for fish swarms in ship cabins often employ fixed cameras in a single water layer, lacking a standardized sampling frequency and only performing simple brightness adjustments. They fail to address the unique interference from underwater suspended particles and bubbles through cascaded enhancement schemes. Traditional contrast enhancement methods typically use fixed-window histogram equalization, and noise suppression relies on simple mean filtering. These approaches are ill-suited to the dynamic changes in fish density and complex noise types encountered in ship cabin fish farming. Consequently, processed images often suffer from blurriness, loss of detail, and incomplete noise elimination, hindering high-precision feeding behavior analysis. This new technical solution specifically designs a multi-water-layer synchronous acquisition and three-stage cascaded enhancement approach. Layered acquisition covers the three-dimensional fish distribution area, adaptive parameter adjustments adapt to the dynamic underwater environment, and step-by-step precise processing systematically addresses image degradation. From acquisition to enhancement, the entire process ensures image quality meets the requirements of subsequent spatiotemporal feature extraction.

[0027] Step S11 involves deploying underwater high-definition cameras at three locations: 0.5 to 1 meter below the surface of the aquaculture tank, 3 to 5 meters below the middle layer, and 1 to 2 meters below the bottom of the tank. These cameras continuously capture images of fish activity at a sampling frequency of 10 to 15 frames per second, simultaneously collecting water temperature and dissolved oxygen parameters. This layered deployment of cameras ensures complete coverage of the entire water column's fish activity area, preventing the omission of feeding behavior in the middle and bottom layers during single-layer sampling. The 10 to 15 frame / second sampling frequency matches the dynamic characteristics of rapidly changing feeding intensity, continuously capturing the complete temporal process of initial aggregation, mid-term feeding, and later dispersal. Simultaneously collected water temperature and dissolved oxygen levels serve as key environmental parameters, providing necessary data support for subsequent deep reinforcement learning state space construction. This step achieves synchronous, continuous, and accurate acquisition of fish images and core water quality parameters across the entire water column, avoiding biased feeding behavior analysis due to data gaps.

[0028] Step S12 uses the dark channel prior algorithm to dehaze the original underwater image, eliminating the image scattering and blurring effects caused by suspended particulate matter in the water. The aquaculture vessel's hold contains a large amount of suspended particulate matter, which scatters light as it penetrates the water, resulting in a foggy and blurry overall image. The dark channel prior algorithm is adapted to the physical characteristics of underwater scattering imaging, accurately identifying and removing scattering interference caused by suspended particulate matter. It effectively restores the clear texture of the underwater image, eliminates the overall fog, and provides a scatter-free, blur-free base image for subsequent contrast enhancement, ensuring the effectiveness of subsequent enhancement processing.

[0029] Step S13 applies adaptive histogram equalization to enhance local contrast in the dehazed image. The sliding window size is adaptively adjusted based on preset minimum window size, maximum window size, empirical scaling factor, fish density, and zero-prevention constant. The fish density in the aquaculture vessel's hold dynamically changes with the aquaculture cycle and feeding stage. At high densities, fish occlusion is severe, requiring a small window size to preserve details of fish edges and feeding postures. At low densities, fish distribution is sparse, requiring a large window size to improve global contrast. Histogram equalization with a fixed window size would result in loss of high-density details and insufficient contrast at low densities. Therefore, the minimum window size is set to 15 pixels, the maximum window size to 60 pixels, the empirical scaling factor ranges from 0.5 to 2.0, and the zero-prevention constant is set to 0.01. The sliding window size is determined by logical calculation using the following formula: ; in, The sliding window size (in pixels) for adaptive histogram equalization. To determine the minimum window size, 15 pixels is used in this embodiment. The maximum window size is set to 60 pixels in this embodiment. K is an empirical scaling factor, calibrated based on the turbidity of the aquaculture water, with a value ranging from 0.5 to 2.0. This represents the fish density in the current image (number of fish pixels / total number of pixels). A small constant (0.01) is used to prevent division by zero. Through the above calculation method, the contrast enhancement requirements under different fish densities can be adapted in real time, balancing detail preservation in high-density scenes with contrast enhancement in low-density scenes. It precisely enhances key features such as fish outlines and feeding angles, avoiding over- or under-enhancement.

[0030] Step S14 applies a nonlocal mean filtering algorithm to suppress noise in the contrast-enhanced image, eliminating salt-and-pepper noise and Gaussian noise caused by bubbles and suspended particles generated during feeding. The feeding process on the aquaculture boat generates a large number of bubbles, and suspended particles in the water persist. These two types of interference introduce salt-and-pepper noise and Gaussian noise, directly interfering with fish contour recognition and feature extraction. The nonlocal mean filtering algorithm utilizes the similarity of local image regions to remove noise while preserving fish texture details, avoiding image blurring caused by simple mean filtering. By thoroughly eliminating noise interference from bubbles and suspended particles, it avoids misjudgment of fish features caused by noise, ultimately outputting a high signal-to-noise ratio and high-definition enhanced image sequence. This provides qualified input data for subsequent dual-branch spatiotemporal feature extraction, ensuring the accuracy and reliability of subsequent feeding behavior analysis.

[0031] In one embodiment, the step S2 of constructing a dual-branch spatiotemporal coupled neural network to extract fused spatiotemporal features further includes: S21: Construct a spatial feature extraction branch and use an improved YOLOv8 network to perform fish target detection and instance segmentation on each frame of the enhanced image sequence. The extracted spatial feature parameters include: the proportion of fish pixels in feeding posture to the total number of fish pixels in the image, the fish aggregation entropy reflecting the uniformity of the spatial distribution of the fish, and the feeding angle features obtained by classification statistics of the angle between the long axis of the fish and the horizontal plane. The feeding activity index is obtained by weighted calculation and used as the spatial feature vector. The dimension of the feature vector is denoted as D. S22: Construct a temporal feature extraction branch, and use a bidirectional long short-term memory network with an attention mechanism to process the enhanced image sequence frames within a continuous time window, extract the temporal evolution pattern of the feeding intensity of the fish school over time, and output a temporal feature vector, the dimension of which is the same as the dimension of the spatial feature vector, both being D. The attention mechanism is used to identify the time window that has a key impact on feeding decisions. S23: An adaptive gating fusion mechanism is used to fuse spatial feature vectors and temporal feature vectors. First, the two are mapped to the same dimension D through a fully connected layer. Then, gating fusion calculation is performed: the spatial feature vector and the temporal feature vector are concatenated to obtain a concatenated vector. Then, the concatenated vector is multiplied by a gating weight matrix and a gating bias term is added. After processing by the Sigmoid activation function, a gating vector is obtained. The values ​​of each element of the gating vector are between 0 and 1. Next, the gating vector and the spatial feature vector are multiplied element-wise to obtain the first product. The constant 1 is subtracted from the gating vector and then multiplied element-wise with the temporal feature vector to obtain the second product. Finally, the first product and the second product are added to obtain the fused spatiotemporal features.

[0032] As described in steps S21-S23 above, considering the characteristics of fish feeding behavior in aquaculture vessel rearing environments, which exhibit both spatial distribution differences and temporal evolution dynamics, a dual-branch parallel spatiotemporal coupled neural network is constructed. This network accurately extracts the spatial distribution features and temporal evolution patterns of fish feeding behavior, and then dynamically fuses the two types of features through an adaptive gating mechanism. This yields a fused spatiotemporal feature that can comprehensively and accurately characterize fish feeding behavior, thus overcoming the technical deficiency that single-dimensional features cannot fully depict feeding behavior. This provides core feature support for the subsequent fuzzy inference system to output an accurate feeding intensity index.

[0033] Fish farming in the hold of aquaculture vessels is a high-density, enclosed aquaculture environment, where the feeding behavior of fish exhibits typical spatiotemporal coupling characteristics: Spatially, fish exhibit differences in aggregation and dispersion, and their feeding postures are clearly distinguishable. Under high density, mutual occlusion of fish bodies further exacerbates the complexity of spatial distribution. Temporally, fish rapidly aggregate and increase feeding intensity in the early stages of feeding, fluctuate feeding intensity in the middle stages, and decrease feeding intensity in the later stages. The temporal evolution patterns at different stages are significantly different. Relying solely on a single spatial feature can easily misjudge feeding intervals as the end of feeding, while relying solely on a single temporal feature cannot distinguish between actual feeding and random swimming, directly leading to errors in judging feeding intensity. Therefore, it is necessary to extract both spatial and temporal features simultaneously and achieve effective fusion to fully reconstruct the essence of fish feeding behavior.

[0034] Traditional fish feeding feature extraction techniques often employ single-branch networks, extracting simple spatial features such as fish aggregation based on a single frame image without constructing a temporal feature extraction branch. Alternatively, they may use ordinary recurrent neural networks to extract temporal features without introducing an attention mechanism to filter key temporal information. Feature fusion often uses fixed-weight concatenation or summation, failing to dynamically adjust the contribution ratio of spatial and temporal features according to different feeding stages. In high-density captive rearing scenarios with dynamic feeding changes, these techniques suffer from incomplete feature extraction, loss of key information, and weak feature representation capabilities after fusion, resulting in a high misjudgment rate of feeding behavior. This technical solution adopts a dual-branch parallel architecture, extracting spatial and temporal features separately. An attention mechanism is introduced to enhance the extraction of key temporal information. Adaptive gating fusion dynamically allocates the weights of the two types of features, achieving deep coupling of spatiotemporal features and comprehensively improving the accuracy and reliability of fish feeding behavior representation.

[0035] Step S21 constructs a spatial feature extraction branch, employing an improved YOLOv8 network to perform fish target detection and instance segmentation frame-by-frame on the enhanced image sequence. Three core spatial parameters are extracted: the proportion of the active feeding area, the fish aggregation entropy, and the feeding angle feature. These are weighted to obtain a feeding activity index of dimension D, which serves as the spatial feature vector. The improved YOLOv8 network adapts to underwater scenes with fish occlusion and blurred edges, accurately distinguishing fish from the background and recognizing different feeding postures. The proportion of the active feeding area quantifies the size of the feeding fish school, the fish aggregation entropy reflects the uniformity of spatial distribution, and the feeding angle feature distinguishes the feeding intensity level. The weighted calculation of these three parameters comprehensively characterizes the spatial state of fish feeding in a single frame image, avoiding the limitations of a single spatial parameter. This allows for the accurate extraction of key information about the spatial dimensions of fish feeding, providing reliable spatial feature input for subsequent feature fusion.

[0036] Step S22 constructs a temporal feature extraction branch, employing a bidirectional long short-term memory network with an attention mechanism to process enhanced image sequences within continuous time windows, outputting a temporal feature vector of dimension D. The bidirectional long short-term memory network can simultaneously capture the positive evolution and negative correlation patterns of feeding intensity, adapting to the temporal characteristics of fish feeding that first increases and then decreases. The attention mechanism can automatically identify key time windows such as initial aggregation, mid-term fluctuations, and late-term decline during feeding and strengthen the corresponding temporal information weights, filtering out irrelevant swimming noise. By accurately capturing the dynamic evolution of fish feeding intensity over time, it solves the problem that traditional temporal models cannot focus on key temporal nodes, outputting temporal features with the same dimension as the temporal feature vector, ensuring compatibility for subsequent feature fusion.

[0037] Step S23 employs an adaptive gating fusion mechanism. First, spatial and temporal feature vectors are mapped to dimension D. Then, through vector concatenation, weight matrix operations, and Sigmoid activation, a gating vector with values ​​from 0 to 1 is generated. Finally, element-wise multiplication and summation operations yield the fused spatiotemporal feature vector. Vector concatenation establishes a preliminary association between the two types of features. Sigmoid activation ensures the gating vector weights are within a reasonable range. Element-wise multiplication dynamically allocates the contribution ratios of spatial and temporal features. In the early stages of feeding, temporal features change drastically, so the gating vector increases the weight of temporal features. In the middle stages of feeding, spatial features can distinguish between feeding intervals and the end of feeding, so the gating vector increases the weight of spatial features. The formula for calculating the gating vector is: ; The formula for calculating the fused spatiotemporal feature vector is as follows: ; Among them, the This represents the gate vector, with each element taking values ​​from 0 to 1. This represents the Sigmoid activation function, a standard activation function with an output range of (0,1). This represents the gate weight matrix, with dimensions D×2D (D being the feature vector dimension). The model trainable parameters are updated through a combination of offline pre-training and online fine-tuning. This represents a spatial feature vector, with dimension D, whose data originates from the feeding activity index vector output above. This represents the temporal feature vector, with dimension D, whose data originates from the temporal evolution pattern vector output above. This represents a vector concatenation operation, which is a standard tensor operation. This represents the gated bias term, with dimension D. This represents element-wise multiplication (Hadamard product). The fused spatiotemporal feature vector, with dimension D, dynamically adapts to the differences in feature importance at different feeding stages by abandoning the fixed weight fusion method mentioned above. This avoids the representation bias caused by a single feature dominating the process. The fused spatiotemporal features simultaneously contain spatial distribution details and temporal evolution patterns, comprehensively and accurately representing the feeding behavior of fish schools. This effectively reduces the misjudgment rate of feeding intensity and provides high-quality feature input for fuzzy inference systems.

[0038] In one embodiment, step S3, which involves fusing spatiotemporal features into the fuzzy inference system to output a continuous feeding intensity index, further includes: S31: Construct a three-input single-output fuzzy inference system, wherein the input variables include: the feeding activity index obtained by fusing spatiotemporal feature mapping, the feeding activity change rate calculated by the feeding activity index at continuous time intervals, and the cumulative feeding amount fed in real time by the feeding device and the weighing device, and the output variable is the feeding intensity index; S32: Define the fuzzy membership function for each input variable: divide the feeding activity index into five fuzzy levels: very low, low, medium, high, and very high; divide the rate of change of feeding activity into five fuzzy levels: rapid decrease, slow decrease, stable, slow increase, and rapid increase; and divide the cumulative food intake into three fuzzy levels: insufficient, moderate, and sufficient. S33: Establish a fuzzy rule base, which contains 27 fuzzy inference rules in the form of "if-then", each rule is mapped to the corresponding fuzzy set of output feeding intensity index according to the fuzziness level of the input variable; S34: The centroid method is used to defuzzify the fuzzy inference results. Specifically, the activation intensity of each activated fuzzy rule is calculated as the product of the activation intensity and the center value of the membership function output by that rule. All products are summed to obtain the numerator, and the activation intensities of all activated rules are summed to obtain the denominator. The numerator is divided by the denominator to obtain the continuous feeding intensity index with a value range of 0 to 100.

[0039] As described in steps S31-S34 above, a three-input single-output fuzzy inference system is constructed. The fusion spatiotemporal feature mapping parameters, feeding dynamic change parameters, and feeding cumulative parameters fed back by the weighing device are fused and inferred to transform the fuzzy and gradual feeding state of the fish into a continuous quantitative feeding intensity index in the range of 0 to 100. This provides accurate and continuous core state input for the subsequent deep reinforcement learning feeding decision model, solves the technical problems of inaccurate qualitative judgment of feeding state and poor adaptability of discretized output, and ensures the accuracy and smoothness of feeding decisions.

[0040] In aquaculture vessel rearing scenarios, the feeding status of fish exhibits fuzzy, gradual, and multi-correlated characteristics, making it impossible to define using absolute strong and weak binary states. Furthermore, the feeding status is simultaneously influenced by three factors: current feeding activity level, short-term trends, and cumulative feed intake. A single-dimensional parameter cannot fully reflect the true feeding demand. For example, a high feeding activity index but a rapid decline in activity indicates that the fish are about to be satiated; a moderate feeding activity index with stable changes and a moderate cumulative feed intake indicates that the fish still have feeding needs; and a low feeding activity index with a continuous decline and sufficient cumulative feed intake indicates that the fish are satiated. Relying solely on a single parameter or discrete thresholds for judgment can easily lead to overfeeding or underfeeding. Therefore, it is essential to use fuzzy reasoning to integrate multi-dimensional parameters to achieve accurate quantification of the feeding status.

[0041] Traditional feeding intensity assessment techniques often employ single-parameter discrete thresholds, relying solely on a single feeding activity index to classify a limited number of levels. They fail to incorporate feeding activity change rate and cumulative feed intake as auxiliary parameters. This results in a limited number of rules, incomplete scenario coverage, and an inability to adapt to the complex changes in feeding status under captive conditions. The output is a discrete level rather than a continuous numerical value, failing to accurately reflect subtle changes in feeding intensity. This leads to coarse and frequent feeding decisions, potentially causing feed waste, water quality deterioration, or fish starvation. This technical solution constructs a three-input coupled fuzzy inference system. Through multi-dimensional parameter fusion, full-scenario rule coverage, and centroid-based fuzzification, it achieves continuous and accurate quantification of feeding intensity, systematically addressing the shortcomings of traditional techniques in terms of judgment dimensions, rule completeness, and output accuracy.

[0042] Step S31 constructs a three-input, single-output fuzzy inference system. The input variables are the feeding activity index obtained by fusing spatiotemporal feature mapping, the feeding activity change rate calculated from the feeding activity index at continuous time intervals, and the cumulative feeding amount fed in real time by the feeding device and weighing device. The output variable is the feeding intensity index. By clarifying the input and output dimensions of the fuzzy inference system, the feeding activity index directly inherits the spatial feature quantification results output above, the feeding activity change rate is calculated based on the dynamic change law of temporal features, and the cumulative feeding amount is statistically analyzed in real time by the weighing data of the feeding device. The three types of inputs characterize the feeding demand from the three dimensions of current state, dynamic trend, and total feeding amount, respectively, realizing multi-dimensional information coupling, avoiding the one-sidedness of judgment caused by a single input dimension, and laying the input foundation for accurate fuzzy inference.

[0043] Step S32 defines the fuzzy membership function for each input variable, dividing the feeding activity index into five fuzzy levels: very low, low, medium, high, and very high. The feeding activity change rate is divided into five fuzzy levels: rapidly decreasing, slowly decreasing, stable, slowly increasing, and rapidly increasing. The cumulative feed intake is divided into three fuzzy levels: insufficient, moderate, and sufficient. Through reasonable grading, continuous numerical inputs are transformed into fuzzy linguistic variables that conform to artificial breeding experience, adapting to the logical characteristics of fuzzy inference. The five levels cover the full range of feeding activity index changes, the five change rate levels accurately match the temporal dynamic characteristics of the entire feeding cycle, and the three cumulative amount levels fit the stage division of total feed intake, reducing fuzzy ambiguity and improving the accuracy of subsequent rule matching.

[0044] Step S33 establishes a fuzzy rule base containing 27 "if-then" rules. Each rule maps to a corresponding fuzzy set of feeding intensity indices based on the fuzziness level of the input variables. This rule base is generated by a full combination of feeding activity index (level 5), feeding activity change rate (level 5), and cumulative feed intake (level 3), completely covering all possible feeding state combinations in the captive rearing scenario, with no rule blind spots. For example, when the feeding activity index is very high, the feeding activity change rate is rapidly increasing, and the cumulative feed intake is insufficient, it is mapped to a high feeding intensity index. When the feeding activity index is moderate, the feeding activity change rate is stable, and the cumulative feed intake is moderate, it is mapped to a moderate feeding intensity index. When the feeding activity index is very low, the feeding activity change rate is rapidly decreasing, and the cumulative feed intake is sufficient, it is mapped to a low feeding intensity index. Each rule is consistent with actual breeding experience, ensuring the rationality of the reasoning results.

[0045] Step S34 uses the centroid method to defuzzify the fuzzy inference results. By calculating the weighted average of the activation intensity of the activation rule and the center value of the output membership function, a continuous feeding intensity index in the interval of 0 to 100 is obtained. This transforms the fuzzy set obtained from fuzzy inference into a precise continuous numerical value. Compared with discrete defuzzification methods such as the maximum membership method, the centroid method can smoothly reflect subtle changes in feeding intensity and output numerical accuracy. The formula for calculating the feeding intensity index is as follows: ; in, The output is the feeding intensity index, with a value range of [0, 100], where n is the total number of fuzzy rules. In this embodiment, n = 27. For the first The activation strength of a fuzzy rule (the degree of matching of the preconditions). For the first The rule outputs the center value of the membership function, such as 58.2, 76.5, etc., to accurately match the dynamic changes in the feeding intensity of the fish school. This provides high-precision and continuous core state variables for the constructed deep reinforcement learning model, avoids feeding decision jumps caused by discrete values, and ensures the smoothness and accuracy of feeding actions. From the perspective of quantitative accuracy, it supports the on-demand feeding needs in the high-density environment of shipboard aquaculture.

[0046] In one embodiment, step S4, which involves constructing a deep reinforcement learning feeding decision model using the feeding intensity index as the core state variable, further includes: S41: Define the state space of the deep reinforcement learning agent. The system state at time t is represented as: the feeding intensity index at the current time, the historical sequence of the feeding intensity indices of the most recent N consecutive frames, the current water temperature, the current dissolved oxygen level, the cumulative feeding amount in this feeding session, the time interval since the last feeding action, and the fused spatiotemporal features. S42: Define the discrete action space of the deep reinforcement learning agent. The action values ​​are 0, 1, 2, 3 and 4, a total of five discrete values. Action 0 corresponds to no feeding, and actions 1 to 4 correspond to small amount of feeding, medium amount of feeding, large amount of feeding and large amount of feeding respectively, which are incremented by the preset calibration value. S43: Construct a multi-objective weighted reward function to evaluate the immediate reward obtained by the agent after performing an action in the current state. This reward function is a weighted sum of the rewards of each sub-item. The first sub-item is the feeding response reward, which is equal to the response sensitivity coefficient multiplied by the change in the feeding intensity index. The second sub-item is the economic reward, which is equal to the negative feed cost coefficient multiplied by the single feeding amount. The third sub-item is the growth efficiency reward, which is equal to the growth reward scaling factor multiplied by the ratio of the fish's body weight gain in this feeding cycle to the single feeding amount. The fourth sub-item is the negative reward item, including feed sinking penalty, water quality deterioration penalty, and starvation penalty. Multiply the above four sub-items by their respective weight coefficients and then sum them to obtain the comprehensive reward value. S44: Construct a deep Q-network based on the Dueling DQN architecture. This network decomposes the action value function into the sum of the state value function and the advantage function minus the mean of the advantage function. Specifically, for a given state and action, first calculate the state value of the state, then calculate the advantage value of the action relative to the state, then subtract the average advantage value of all possible actions in the state from the advantage value, and finally add the state value to the aforementioned difference to obtain the Q value of the state-action pair. S45: A strategy combining offline pre-training and online continuous fine-tuning is adopted to train the deep reinforcement learning agent: In the pre-training stage, the Dueling DQN network is pre-trained using historical aquaculture data. In the online fine-tuning stage, the network parameters are updated every 24 hours using the interaction data collected on the same day during the actual aquaculture process to continuously optimize the feeding strategy.

[0047] As described in steps S41-S45 above, the continuous feeding intensity index output above is used as the core state variable. By integrating multi-dimensional aquaculture environment and feeding-related parameters, a deep reinforcement learning feeding decision model based on an improved Dueling DQN architecture is constructed. A multi-objective weighted reward function that takes into account feeding response, economy, growth efficiency and risk constraints is designed. A training strategy that combines offline pre-training and online continuous fine-tuning is adopted, enabling the model to autonomously learn the optimal feeding action through continuous interaction with the aquaculture environment. This replaces the traditional static decision rules, adapts to the complex scenario of high-density and dynamic changes in aquaculture vessel holding, and achieves adaptive, precise and dynamic optimization of feeding decisions, providing the optimal action output for the generation of subsequent feeding instructions.

[0048] In aquaculture vessel rearing scenarios, fish feeding habits vary depending on species and growth stage. Environmental factors such as water temperature and dissolved oxygen levels fluctuate in real time, resulting in continuously changing feeding needs. Feeding decisions must simultaneously meet multiple objectives, including timely response to feeding demands, control of feed costs, ensuring fish growth, and avoiding feed settling and water quality deterioration. Single decision-making criteria or static rules cannot balance these multiple objectives. Traditional models rely solely on offline training based on historical data, which cannot adapt to the long-term dynamic changes in the aquaculture process. This can easily lead to decision rigidity and decreased adaptability, directly resulting in overfeeding causing feed waste and water quality deterioration, or underfeeding causing slow fish growth and reduced economic benefits. Therefore, it is necessary to construct a reinforcement learning decision-making model with multi-dimensional state input, multi-objective reward guidance, and dynamic updates to achieve autonomous optimization and dynamic adaptation of feeding strategies.

[0049] Traditional feeding decision-making techniques often employ static thresholds or simple classification models, with limited state input dimensions. They rely solely on feeding intensity, neglecting key parameters such as water temperature, dissolved oxygen, and historical feeding sequences, failing to comprehensively depict the dynamics of the aquaculture environment. Their reward function designs are simplistic, focusing only on feeding response while ignoring feed costs, growth efficiency, and water quality risks, easily leading to biased decisions towards a single objective. Using ordinary DQN networks results in large variance in action value estimation, poor training stability, and difficulty converging to the optimal strategy. Furthermore, the models undergo only one-time offline training without online fine-tuning mechanisms, failing to adapt to long-term dynamics such as fish growth and environmental changes throughout the aquaculture cycle, leading to a continuous decline in decision accuracy over extended aquaculture time. This technical solution systematically addresses the core shortcomings of traditional techniques—limited state dimensions, single objectives, unstable training, and poor adaptability—through multi-dimensional state space construction, improved Dueling DQN network design, multi-objective weighted reward function settings, and a training strategy combining offline pre-training and online fine-tuning. This achieves intelligent and adaptive optimization of feeding decisions.

[0050] Step S41 defines the state space of the deep reinforcement learning agent. The system state at time t includes the current feeding intensity index, the historical sequence of the feeding intensity index of the most recent N consecutive frames, the current water temperature, the current dissolved oxygen level, the cumulative feeding amount in this feeding session, the time interval since the last feeding action, and fused spatiotemporal features. This step constructs a seven-dimensional state space, comprehensively depicting the state of the aquaculture system from the dimensions of real-time feeding status of the fish, historical feeding trends, core environmental parameters, feeding execution parameters, and behavioral characteristics, avoiding the one-sidedness of decision-making caused by a single state variable. Among them, the feeding intensity index and historical sequence are the quantitative results output above, the water temperature and dissolved oxygen level are environmental parameters collected synchronously in step S1, the cumulative feeding amount is fed back in real time by the weighing device in step S5, and the fused spatiotemporal features are behavioral feature parameters. The multi-dimensional state input provides comprehensive data support for the model to accurately perceive the aquaculture dynamics and make reasonable decisions.

[0051] Step S42 defines the discrete action space of the deep reinforcement learning agent, with action values ​​of 0, 1, 2, 3, and 4, corresponding to no feeding, small amount feeding, medium amount feeding, large amount feeding, and large amount feeding, respectively. This step discretizes the feeding action into five levels, and the feeding amount for each level is pre-calibrated according to the feed characteristics and the throwing diameter of the feeding device. Discretizing the action space can reduce the model training complexity and avoid the high-dimensional search problem of continuous action space. The five action levels fully cover the entire feeding amount range in the shipboard feeding scenario, accurately match the feeding needs under different feeding intensities, facilitate the model to quickly learn the mapping relationship between actions and states, and improve decision-making efficiency.

[0052] Step S43 constructs a multi-objective weighted reward function. The comprehensive reward value at time t is obtained by weighted summation of feeding response reward, economic reward, growth efficiency reward, and negative reward items. The feeding response reward equals the response sensitivity coefficient multiplied by the change in feeding intensity index, with a response sensitivity coefficient of 5.0. The economic reward equals the negative feed cost coefficient multiplied by the single feeding amount, with a feed cost coefficient of 0.2. The growth efficiency reward equals the growth reward scaling factor multiplied by the ratio of fish weight gain to single feeding amount, with a growth reward scaling factor of 10.0. The negative reward items include feed sinking penalty, water quality deterioration penalty, and starvation penalty, corresponding to penalty values ​​of -2.0, -1.0, and -1.5 respectively. The weight coefficients for each sub-item are 0.4, 0.3, 0.2, and 0.1, respectively, and the sum of the weights is 1. The formula for calculating the comprehensive reward value at time t is: R t = R feed + R eco + R growth + R penalty in, R feed ; R eco ; R growth ; Among them, the This represents the total reward value at time t. , , and Denotes the weight coefficients, and satisfies + + + =1, Indicates a reward for the feeding response. This represents the response sensitivity coefficient. This indicates the change in the food intensity index. Indicates economic reward, This represents the feed cost coefficient. Indicates the amount fed per feeding. This indicates a reward for growth efficiency. This represents the growth reward scaling factor. This indicates the weight gain of the fish during the feeding cycle. The negative reward items include feed settling to the bottom (2.0), water quality deterioration (1), and starvation (1.5). Through multi-objective reward design, the four core requirements of feeding response, cost control, growth efficiency, and risk avoidance are balanced. This avoids decision-making bias under the guidance of a single reward. Feeding response rewards encourage the model to respond to changes in feeding intensity in a timely manner, economic rewards inhibit overfeeding, growth efficiency rewards guide the improvement of feed conversion rate, negative reward items constrain risky behavior, and weighted design ensures that all objectives are balanced and takes into account, guiding the model to learn the globally optimal feeding strategy.

[0053] Step S44 constructs a deep Q-network based on the Dueling DQN architecture, decomposing the action value function into the sum of the state value function and the advantage function, minus the mean of the advantage function. This architecture can separate the inherent value of the state from the relative advantage of the action, reduce the variance of value estimation between similar actions, and improve the model training stability and convergence speed. Compared with ordinary DQN, Dueling DQN can more accurately evaluate the merits of different feeding actions, reduce the decision instability caused by value estimation fluctuations, ensure that the model outputs stable and reasonable optimal feeding actions under complex breeding conditions, and enhance the reliability of decision-making.

[0054] Step S45 employs a strategy combining offline pre-training and online continuous fine-tuning to train the deep reinforcement learning agent. During the pre-training phase, historical aquaculture data containing 2 million labeled image frames across three aquaculture cycles is used to train the Dueling DQN network, quickly establishing an initial feeding strategy. During the online fine-tuning phase, network parameters are updated every 24 hours using the interactive data collected that day, continuously optimizing the feeding strategy. This strategy can shorten the initial deployment cycle of the model and avoid the inefficiency of training from scratch. Online fine-tuning can adapt to long-term dynamic changes such as fish growth, seasonal changes in water temperature, and wear and tear of feeding devices, enabling the model to continuously adapt to the aquaculture scenario and ensuring that the accuracy of feeding decisions remains stable as the aquaculture cycle progresses. This avoids the adaptation decay problem caused by model solidification and supports the long-term stable operation of the model from the training mechanism level.

[0055] In one embodiment, step S5, which involves generating a feeding command, executing the feeding, and using weighing feedback for closed-loop PID control and correction, further includes: S51: Map the optimal feeding action output by the deep reinforcement learning model to the target feeding amount, and use model predictive control to perform rolling time-domain optimization and smoothing on the target feeding amount sequence to generate a smooth feeding instruction that satisfies the single maximum feeding amount constraint, feeding interval constraint and total feeding amount constraint. S52: Convert the smoothed target feed amount into the target rotation speed and target running time of the rotary feeder in the feeding device, and at the same time determine the target rotation speed of the feeder based on the throwing characteristic parameters of the electric impeller feeder, and generate the underlying execution command; S53: The feeding device executes the underlying execution command to feed the material, and at the same time, the weighing device at the bottom of the feeding device collects the weight change of the hopper in real time at a sampling frequency of not less than 10 Hz and calculates the actual amount of material fed. S54: The control system compares the target feeding amount with the actual feeding amount in real time, and uses a closed-loop PID control algorithm to dynamically fine-tune the rotation speed of the rotary feeder so that the steady-state deviation between the actual feeding amount and the target feeding amount is controlled within ±3%, and the actual feeding amount accumulated in this meal is used as the feedback input of the accumulated feed intake in step S3. S55: Repeat steps S1 to S54. When the optimal feeding action output by the decision is no feeding for N consecutive times and the current feeding intensity index is continuously lower than the preset hunger threshold, it is determined that the feeding of this meal is over and the system enters standby state until the next feeding cycle is triggered.

[0056] As described in steps S51-S55 above, the discrete optimal feeding action output by the deep reinforcement learning model is transformed into a smooth and executable feeding instruction. The feeding operation is precisely executed through the feeding device. A closed-loop PID correction mechanism is constructed based on real-time feedback from the weighing device to dynamically calibrate the feeding amount deviation. At the same time, a scientific feeding end judgment logic is set to form a complete closed-loop execution link from decision output to execution feedback and then to end judgment. This solves the technical problems of feeding instruction jumps, large execution deviations, and inaccurate end judgments, ensuring the smoothness, accuracy, and rationality of the feeding process, and adapting to the refined feeding needs of high-density, dynamic feeding scenarios in aquaculture boats.

[0057] In high-density aquaculture scenarios on boats, fish feeding intensity exhibits a continuous, gradual variation. The optimal feeding action output by deep reinforcement learning is discrete in level, and direct mapping execution can easily lead to frequent jumps in feeding amount, duration, and interval. Fish struggle to adapt to these abrupt feeding rhythms, resulting in feeding disorder. Feeding devices also suffer from mechanical transmission errors, differences in feed flowability, and feed hopper arching. The theoretical feeding amount can easily deviate from the actual feeding amount, and this deviation accumulates over time. The end of each feeding session often relies on fixed durations or manual observation, which cannot accurately match the actual satiety state of the fish. This can easily lead to premature cessation of feeding, resulting in insufficient feeding, or delayed cessation of feeding, causing feed waste and water pollution. Therefore, it is essential to construct a complete execution system that features smooth instructions, precise execution, closed-loop correction, and dynamic termination to ensure a high degree of matching between feeding actions and the dynamics of fish feeding.

[0058] Traditional feeding techniques directly use discrete actions corresponding to fixed feeding parameters, lacking smoothing processing. This leads to abrupt changes in feeding rhythm, disrupting the fish's normal feeding. The lack of real-time weighing feedback makes it impossible to detect actual feeding deviations, and there is no closed-loop correction mechanism, resulting in persistent and accumulated discrepancies between theoretical and actual feeding. Meal end determination relies on preset fixed durations, failing to consider real-time feeding intensity dynamics, leading to low accuracy and potential underfeeding or overfeeding, making it unsuitable for dynamic feeding scenarios in captive rearing. This technical solution introduces model predictive control to achieve smooth command execution, ensures accurate execution through precise mapping of mechanical parameters, achieves closed-loop deviation correction through high-frequency weighing feedback and PID control, and establishes a multi-condition linked end-determination logic. It systematically addresses the shortcomings of traditional execution processes from four aspects: command output, mechanical execution, deviation correction, and end-determination, constructing a precise closed-loop execution system across the entire chain.

[0059] Step S51 maps the optimal feeding action output by the deep reinforcement learning model to the target feeding amount. Model predictive control is used for rolling time-domain optimization and smoothing to generate a smooth feeding instruction that meets the constraints of maximum single feeding amount, feeding interval, and total feeding amount. Direct mapping of discrete actions can easily cause sudden changes in feeding parameters. Model predictive control, through rolling time-domain optimization, integrates historical feeding data, current feeding intensity, and constraints to generate a continuous and smooth sequence of feeding amount, duration, and interval, eliminating discrete action jumps, adapting to the gradual feeding needs of the fish, avoiding feeding disorder caused by sudden changes in feeding rhythm, and strictly following various constraints to prevent excessive feeding in a single instance, excessively short intervals, or excessive total amount, ensuring that the feeding process is compliant and controllable.

[0060] Step S52 converts the smoothed target feeding amount into the target rotation speed and target running time of the rotary feeder. Combining the spreading characteristics of the electric impeller-type feed spreader, the target rotation speed is determined, and the underlying execution command is generated. The feeding amount of the feeding device is precisely controlled by the rotation speed and running time of the rotary feeder. The rotation speed of the feed spreader determines the uniformity of feed spreading. By establishing a precise mapping relationship between the target feeding amount and the mechanical operating parameters, and adapting to the mechanical structural characteristics of the conical hopper, rotary feeder, and electric impeller-type feed spreader, it is ensured that the feed is quantitatively and evenly spread to the feeding area of ​​the aquaculture tank, avoiding local feed accumulation or uneven spreading, and ensuring that the fish feed evenly.

[0061] In step S53, the feeding device executes the underlying command to feed the food. The weighing device collects the weight change of the hopper in real time at a sampling frequency of 10 Hz and calculates the actual feeding amount. High-frequency sampling can capture the small weight fluctuations in the feeding process in real time, accurately reflect the instantaneous feeding amount and the cumulative feeding amount, eliminate the random error of a single sampling, and provide high data acquisition accuracy and strong real-time performance. This provides high-frequency and reliable feedback data for subsequent PID correction and timely captures feeding deviations caused by mechanical errors, changes in feed flowability, etc.

[0062] Step S54: The control system compares the target feeding amount with the actual feeding amount in real time. It uses closed-loop PID regulation to dynamically fine-tune the rotation speed of the rotary feeder, keeping the steady-state deviation within ±3%. At the same time, the cumulative actual feeding amount is fed back to the fuzzy inference system of claim 4. The closed-loop PID regulation can quickly respond to deviation changes, dynamically correct the feeding speed, and continuously narrow the gap between the theoretical and actual feeding amounts. The ±3% deviation control accuracy is suitable for the precise feeding requirements of high-density canopy rearing. The cumulative feeding amount feedback realizes the closed-loop return of execution data to the analysis stage, providing real-time feeding data for subsequent quantification of feeding intensity and improving the closed-loop data of the entire chain.

[0063] Step S55 executes the aforementioned process repeatedly. When the decision output is "no feeding" for N consecutive times and the feeding intensity index is consistently lower than the preset hunger threshold, the feeding session is determined to be over and the system enters a standby state. This end-of-session determination logic combines the decision output and real-time feeding intensity as dual indicators to avoid misjudgment based on a single condition, accurately match the actual satiated state of the fish, prevent premature cessation of feeding leading to insufficient feeding, or delay in cessation of feeding leading to feed waste. The standby state can reduce ineffective energy consumption until the next feeding cycle is triggered, adapting to the periodic feeding management needs of the ship's hold and ensuring reasonable connection of feeding cycles.

[0064] like Figure 2 As shown, this invention also discloses an intelligent feeding system for shipboard aquaculture based on fish feeding analysis, comprising: The acquisition module is used to acquire images of fish activity and synchronized water quality parameters in each aquaculture cabin of the aquaculture vessel, and to perform cascade enhancement processing on the images, including defogging, contrast enhancement and noise suppression, to obtain an enhanced image sequence. The construction module is used to construct a dual-branch spatiotemporal coupled neural network. The enhanced image sequence is input into the spatial feature extraction branch and the temporal feature extraction branch respectively to extract the spatial distribution features and temporal evolution patterns of fish feeding behavior. The features are then fused through an adaptive gating mechanism to obtain fused spatiotemporal features. The output module is used to input the fused spatiotemporal features into the fuzzy inference system, combine them with the cumulative food intake for inference, and output a continuous food intake intensity index. The autonomous learning module is used to construct a feeding decision model based on deep reinforcement learning, with the feeding intensity index as the core state variable, and autonomously learn the optimal feeding action through interaction with the environment. The generation module is used to generate specific feeding instructions after the optimal feeding action is smoothed by the control instructions and then sent to the feeding device for execution. The actual feeding amount fed in real time by the weighing device is used for closed-loop feedback adjustment until the feeding end condition of the meal is met.

[0065] In one embodiment, the output module includes: The building unit is used to construct the fuzzy inference system. Its input variables include the feeding activity index, the rate of change of feeding activity, and the cumulative amount of food intake obtained by fusing spatiotemporal feature mapping. The output variable is the feeding intensity index. Establish a unit to define the fuzzy membership function of each input variable and establish a fuzzy rule base; The processing unit is used to defuzzify the fuzzy inference results using the centroid method to obtain a continuous feeding intensity index.

[0066] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described intelligent feeding method for shipboard rearing based on fish feeding analysis.

[0067] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described intelligent feeding method for shipboard rearing based on fish feeding analysis.

[0068] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0069] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0070] The above description is merely a preferred embodiment of the present invention and does not limit the scope of this application. Any equivalent results or equivalent process transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of protection of this application.

Claims

1. A method for intelligent feeding in shipboard cabins based on fish feeding analysis, characterized in that, Includes the following steps: Images of fish activity and synchronized water quality parameters were collected from each aquaculture compartment of the aquaculture vessel. The images were then subjected to cascade enhancement processing, including defogging, contrast enhancement, and noise suppression, to obtain an enhanced image sequence. A dual-branch spatiotemporal coupled neural network is constructed. The enhanced image sequence is input into the spatial feature extraction branch and the temporal feature extraction branch respectively to extract the spatial distribution features and temporal evolution patterns of fish feeding behavior. The features are then fused through an adaptive gating mechanism to obtain fused spatiotemporal features. The fused spatiotemporal features are input into the fuzzy inference system, and inference is performed in combination with the cumulative food intake to output a continuous food intake intensity index; Using the feeding intensity index as the core state variable, a feeding decision model based on deep reinforcement learning is constructed, which learns the optimal feeding action autonomously through interaction with the environment. After the optimal feeding action is smoothed by the control command, a specific feeding command is generated and sent to the feeding device for execution. The actual feeding amount fed in real time by the weighing device is used for closed-loop feedback adjustment until the feeding end condition of this meal is met.

2. The intelligent feeding method for shipboard rearing based on fish feeding analysis according to claim 1, characterized in that, Perform cascaded enhancement processing, including dehazing, contrast enhancement, and noise suppression, on the image to obtain an enhanced image sequence, further including: Underwater cameras deployed at different water levels in each aquaculture tank collect images of fish activity and simultaneously collect water quality parameters; Perform dehazing on the acquired raw images; Adaptive histogram equalization is used to enhance local contrast in the dehazed image, where the sliding window size of equalization is adaptively adjusted according to the fish density in the image. The contrast-enhanced image is subjected to noise suppression processing, and the enhanced image sequence is output.

3. The intelligent feeding method for shipboard rearing based on fish feeding analysis according to claim 1, characterized in that, The step of performing feature fusion through an adaptive gating mechanism to obtain fused spatiotemporal features further includes: A spatial feature extraction branch is constructed to detect and segment fish swarms in each frame of the enhanced image sequence, extract spatial features, and obtain spatial feature vectors. A temporal feature extraction branch is constructed, and a bidirectional long short-term memory network with an attention mechanism is used to process image sequences within a continuous time window to extract temporal evolution patterns and obtain temporal feature vectors. An adaptive gating fusion mechanism is adopted to fuse the spatial feature vector and the temporal feature vector. A gating vector is calculated through a gating network, and the spatial feature vector and the temporal feature vector are dynamically weighted and fused using the gating vector to obtain fused spatiotemporal features.

4. The intelligent feeding method for shipboard rearing based on fish feeding analysis according to claim 1, characterized in that, The fused spatiotemporal features are input into the fuzzy inference system, and inference is performed in conjunction with the cumulative food intake to output a continuous food intake intensity index, which further includes: Construct a fuzzy inference system whose input variables include the feeding activity index, the rate of change of feeding activity, and the cumulative amount of food consumed, obtained by fusing spatiotemporal feature mapping, and whose output variable is the feeding intensity index. Define the fuzzy membership function for each input variable and establish a fuzzy rule base; The centroid method is used to defuzzify the fuzzy inference results to obtain a continuous feeding intensity index.

5. The intelligent feeding method for shipboard rearing based on fish feeding analysis according to claim 1, characterized in that, Using the feeding intensity index as the core state variable, a feeding decision model based on deep reinforcement learning is constructed. This model autonomously learns the optimal feeding action through interaction with the environment, and further includes: Define the state space of the deep reinforcement learning model, wherein the state includes at least the current feeding intensity index, the historical feeding intensity index sequence, the current water quality parameters, the cumulative feeding amount in this feeding session, the time interval since the last feeding action, and the fused spatiotemporal features; Define the discrete action space of the model, which includes no feeding and multiple feeding actions with different amounts; Construct a multi-objective reward function, which includes at least a feeding response reward for encouraging responses to feeding needs, an economic reward for inhibiting overfeeding, a growth efficiency reward for improving feed conversion rate, and a negative reward for constraining adverse feeding consequences. Construct a deep Q-network based on the Dueling DQN architecture to decompose the action value function into a state value function and an advantage function; The deep reinforcement learning model is trained using a strategy that combines offline pre-training with online continuous fine-tuning.

6. The intelligent feeding method for shipboard rearing based on fish feeding analysis according to claim 1, characterized in that, The method of using the actual feeding amount fed in real time by a weighing device for closed-loop feedback adjustment until the feeding end condition for that meal is met further includes: The optimal feeding action is mapped to the target feeding amount, and a predictive control method is used to smooth the target feeding amount to generate a smooth feeding instruction that meets preset constraints. The smooth feeding instruction is converted into an execution instruction for the feeding device; The execution command is used to feed the material, and the actual amount of material fed is collected in real time by a weighing device. The target feeding amount is compared with the actual feeding amount in real time. A closed-loop control algorithm is used to dynamically correct the feeding process so that the actual feeding amount approaches the target feeding amount, and the cumulative actual feeding amount is fed back. Repeat the above steps of data collection and closed-loop feedback adjustment. When the optimal feeding action output by the decision is no feeding for multiple consecutive times and the current feeding intensity index is continuously lower than the preset threshold, it is determined that the feeding for this meal is over.

7. A smart feeding system for shipboard aquaculture based on fish feeding analysis, characterized in that, include: The acquisition module is used to acquire images of fish activity and synchronized water quality parameters in each aquaculture cabin of the aquaculture vessel, and to perform cascade enhancement processing on the images, including defogging, contrast enhancement and noise suppression, to obtain an enhanced image sequence. The construction module is used to construct a dual-branch spatiotemporal coupled neural network. The enhanced image sequence is input into the spatial feature extraction branch and the temporal feature extraction branch respectively to extract the spatial distribution features and temporal evolution patterns of fish feeding behavior. The features are then fused through an adaptive gating mechanism to obtain fused spatiotemporal features. The output module is used to input the fused spatiotemporal features into the fuzzy inference system, combine them with the cumulative food intake for inference, and output a continuous food intake intensity index. The autonomous learning module is used to construct a feeding decision model based on deep reinforcement learning with the feeding intensity index as the core state variable, and autonomously learn the optimal feeding action through interaction with the environment. The generation module is used to generate specific feeding instructions after the optimal feeding action is smoothed by the control instructions and then sent to the feeding device for execution. The actual feeding amount fed in real time by the weighing device is used for closed-loop feedback adjustment until the feeding end condition of the meal is met.

8. The intelligent feeding system for shipboard rearing based on fish feeding analysis according to claim 7, characterized in that, The output module includes: The building unit is used to construct the fuzzy inference system. Its input variables include the feeding activity index, the rate of change of feeding activity, and the cumulative amount of food intake obtained by fusing spatiotemporal feature mapping. The output variable is the feeding intensity index. Establish a unit to define the fuzzy membership function of each input variable and establish a fuzzy rule base; The processing unit is used to defuzzify the fuzzy inference results using the centroid method to obtain a continuous feeding intensity index.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images