Red fin lutjanus culture feeding optimization method and system based on dynamic perception

Abstract

The invention discloses a lutjanus erythropterus culture feeding optimization method and system based on dynamic perception, and the method comprises the steps: obtaining the perception information of a culture region, carrying out the preprocessing of the perception information of the culture region, and generating a culture perception data set; extracting single-mode features based on the breeding perception data set, performing feature fusion according to the extracted single-mode features to generate multi-mode breeding fusion features, and performing appetite state modeling by using the multi-mode breeding fusion features to obtain appetite state indexes; according to the appetite state index, whether feeding needs to be carried out or not is judged, and a feeding strategy is formulated by combining the multi-mode breeding fusion characteristics to carry out feeding strategy optimization; and after bait casting is finished each time, residual bait collection and environment monitoring are carried out, the incidence relation between residual bait and environment parameters is analyzed, and a residual bait punishment mechanism is established to optimize the breeding bait casting degree. Precise sensing of feeding is achieved, the breeding risk caused by excessive feeding and other conditions is avoided, and therefore the breeding efficiency and intelligence of the lutjanus erythropterus are improved.

Description

Technical Field

[0001] This invention relates to the field of aquaculture feeding technology, and in particular to a method and system for optimizing feeding in redfin snapper aquaculture based on dynamic perception. Background Technology

[0002] In the aquaculture industry, the red snapper (Lutjanus erythropterus), as a high-economic-value marine fish, faces the core technological challenge of precise feeding in its large-scale farming. Traditional artificial feeding relies on fixed schedules and empirical quantitative models, ignoring the dynamic changes in fish feeding behavior and the coupling effect with the environment, resulting in a feed waste rate as high as 30%-40%. Although existing technologies attempt to apply computer vision analysis of fish aggregation behavior, static threshold algorithms are difficult to adapt to the unique intermittent feeding rhythm of the red snapper (70% of its daily feeding occurs during the morning and evening peak) and ignore dynamic environmental disturbances such as sudden changes in dissolved oxygen and the formation of thermoclines. Controlling feeding by using a preset water temperature-feeding amount lookup table does not solve the problem of regional uneaten feed accumulation caused by fish migration, and the rate of ammonia nitrogen exceedances in the water due to uneaten feed decay still exists. Therefore, this paper proposes a dynamic perception-based method and system for optimizing feeding in red snapper farming, achieving precise feeding while avoiding farming risks caused by overfeeding, thereby improving the efficiency and intelligence of red snapper farming. Summary of the Invention

[0003] This invention overcomes the shortcomings of the prior art and provides a method and system for optimizing feeding in red snapper farming based on dynamic perception. Its main purpose is to improve the efficiency and intelligence of red snapper farming.

[0004] To achieve the above objectives, the first aspect of this invention provides a method for optimizing feeding in redfin snapper farming based on dynamic perception, comprising:

[0005] Fish condition monitoring and water condition monitoring are carried out in the target aquaculture area to obtain aquaculture area perception information. The aquaculture area perception information is preprocessed to generate an aquaculture perception dataset.

[0006] Based on the aquaculture perception dataset, unimodal features are extracted, and feature fusion is performed on the extracted unimodal features to generate multimodal aquaculture fusion features. The multimodal aquaculture fusion features are then used to model appetite state to obtain an appetite state index.

[0007] The feeding strategy is optimized by determining whether feeding is necessary based on the appetite state index and combining the multimodal aquaculture integration characteristics.

[0008] After each feeding, uneaten feed is collected and the environment is monitored. The correlation between uneaten feed and environmental parameters is analyzed, and a penalty mechanism for uneaten feed is established to optimize the feeding level in aquaculture.

[0009] In this solution, the monitoring of fish condition and water condition in the target aquaculture area to obtain aquaculture area perception information, and the preprocessing of the aquaculture area perception information to generate an aquaculture perception dataset, specifically includes:

[0010] A heterogeneous sensor network is deployed in the target aquaculture area to monitor the aquaculture ponds in the target aquaculture area and obtain aquaculture area perception information, which includes fish body perception information and water body perception information.

[0011] Adaptive gamma correction technology is used to preprocess the image data in the fish's sensory information. The pixel values ​​are dynamically adjusted according to the brightness distribution of each frame. Then, the structure of similar image blocks is analyzed by non-local mean denoising algorithm to filter out the interference of suspended particles and output an enhanced image sequence.

[0012] Doppler frequency shift correction is performed on the sonar data in the fish body perception information, coordinate offset is compensated according to the sensor motion posture, and an outlier elimination algorithm based on kurtosis detection is used to eliminate occasional equipment vibration interference, generating a three-dimensional motion trajectory sequence.

[0013] Based on the enhanced image sequence and the three-dimensional motion trajectory sequence, preprocessed fish body perception information is generated, the standard deviation of the water body perception information is calculated, and outlier detection and missing value supplementation are performed in combination with the preset standard deviation range to generate preprocessed water body perception information.

[0014] The pre-processed fish body perception information and the pre-processed water body perception information are combined to form an aquaculture perception dataset.

[0015] In this scheme, the step of extracting single-modal features based on the aquaculture perception dataset and generating multimodal aquaculture fusion features by fusing the extracted single-modal features specifically includes:

[0016] We acquired an aquaculture perception dataset and input the enhanced image sequence into a pre-trained four-channel YOLOv5 detection network. We set an extended near-infrared spectral channel in the input layer and extracted visible light and infrared band features simultaneously through a four-dimensional convolution kernel.

[0017] An embedded channel attention module is used in the backbone network to enhance the thermal signal response region of the organism based on spectral energy, while a spatial attention mechanism is introduced to focus on the contour change features of the moving target.

[0018] The pooling layer is used to obtain the fish body bounding box coordinates, confidence scores, and spatial mask matrix of fish distribution. A dynamic fish density index is generated. The inverse distance weighted interpolation method is used to map the position coordinates of all detected targets into a two-dimensional planar heat map to generate fish aggregation distribution features and obtain the first modality features.

[0019] The three-dimensional motion trajectory sequence is extracted from the aquaculture perception dataset. The instantaneous velocity of the individual is calculated based on the Euclidean distance between adjacent timestamps. The median value of all target velocities is taken as the average swimming speed index of the fish group. The discrete angular standard deviation of all fish movement directions within a preset period is calculated using the direction vector analysis method to generate the second modal feature.

[0020] The pre-processed water body sensing information is obtained through the aquaculture sensing dataset and an environmental index matrix is ​​generated. The environmental anomaly index is calculated based on the dissolved oxygen change rate, temperature gradient abrupt change point and turbidity cumulative fluctuation, and the third modal feature is obtained.

[0021] The first modal feature, the second modal feature, and the third modal feature are input into a gated recurrent unit network and combined with a multi-head attention mechanism to calculate the feature fusion weights, thereby generating multimodal aquaculture fusion features.

[0022] In this scheme, the step of using the multimodal aquaculture fusion features to model appetite state and obtain an appetite state index specifically includes:

[0023] Obtain multimodal aquaculture fusion features, calculate the original value of cluster density of the target aquaculture area based on the multimodal aquaculture fusion features, normalize the cluster density and linearly scale it based on the historical maximum density of the day to obtain the cluster density index;

[0024] Based on the multimodal aquaculture fusion features, the original value of the vertical activity entropy of the target aquaculture area is calculated using the Shannon entropy principle, and then divided by the preset theoretical maximum entropy value to obtain the vertical distribution index.

[0025] The appetite status of redfin snapper in the target aquaculture area is modeled using the aggregation density index, vertical distribution index, and multimodal aquaculture fusion features based on a deep neural network.

[0026] A bidirectional GRU unit is set up at the network front end to perform time-series dependency analysis. Two hidden layers are used to capture the precursor patterns of appetite changes. An attention mechanism is introduced at the network back end to build a triple parallel analysis channel, namely the spatial channel, the behavioral channel and the environmental channel.

[0027] The output values ​​of the triple parallel analysis channels are obtained. The input value aggregation unit performs weighted calculations according to preset weights and uses the Sigmoid function to compress the weighted results to generate an appetite state index.

[0028] In this scheme, the step of determining whether feeding is necessary based on the appetite index and optimizing the feeding strategy by combining the multimodal aquaculture integration characteristics specifically includes:

[0029] The appetite state index and multimodal aquaculture fusion features are obtained. The appetite state index is compared with the preset feeding judgment rules. If the appetite state index is greater than a certain threshold, it means that feed needs to be fed and the feeding command is triggered.

[0030] If the appetite index is less than a Class I threshold and greater than a Class II threshold, a delayed feeding instruction is triggered; if the appetite index is less than a Class II threshold, it means that no feeding is needed, and a feeding freeze instruction is triggered.

[0031] When it is necessary to feed the redfin snapper in the target aquaculture area, the feeding strategy is optimized by using the multimodal aquaculture fusion feature. The feeding area is gridded to generate several grid cells, and the multimodal aquaculture fusion feature is aligned with each grid cell.

[0032] Acquire the cluster density index and vertical distribution index, define candidate feeding grid cells using the cluster density index and vertical distribution index, and generate the first dataset based on the multimodal aquaculture fusion features corresponding to each candidate feeding grid cell;

[0033] A Bayesian inference network is introduced. The first dataset is input into the Bayesian inference network for training to obtain the transient posterior distribution. The transient posterior distribution is used to infer the movement state of the fish in each candidate feeding grid cell at the next moment in the Bayesian network.

[0034] The next moment motion state is obtained by inference and secondary feeding grid cells are defined based on the inferred probability value. The secondary feeding grid cells are merged with the candidate feeding grid cells to generate recommended feeding areas for aquaculture feeding optimization.

[0035] In this plan, the collection of uneaten feed and environmental monitoring after each feeding, analysis of the correlation between uneaten feed and environmental parameters, and establishment of an uneaten feed penalty mechanism to optimize the feeding level in aquaculture specifically include:

[0036] A grid-type residual feed collection device is laid in the aquaculture pond. The residual feed is collected and weighed within a preset time after each feeding by suction and filtration. At the same time, heterogeneous sensors are deployed to monitor water parameters before and after feeding to obtain aquaculture feeding perception information.

[0037] Based on the aquaculture feeding perception information, a heat map of residual feed distribution is generated. When the amount of residual feed corresponding to a certain grid unit exceeds a preset threshold, it is marked as an abnormal residual feed area. The corresponding water parameters are extracted through the aquaculture feeding perception information to generate abnormal residual feed area information.

[0038] Based on the information on abnormal residual bait areas, a covariance matrix is ​​constructed with the amount of residual bait exceeding the standard as the dependent variable and the fluctuation of water body environmental parameters as the independent variable. The singular value decomposition method is used to perform eigenvalue decomposition on the covariance matrix to analyze the correlation strength between the amount of residual bait exceeding the standard and each water body environmental parameter.

[0039] The correlation strength between the amount of uneaten bait exceeding the standard and various water environmental parameters is judged against the preset threshold. The main water environmental factors that lead to abnormal uneaten bait are analyzed, and information on the cause of abnormal uneaten bait is generated.

[0040] A residual bait penalty mechanism is established based on the residual bait anomaly attribution information. The residual bait penalty mechanism is established based on the water environment parameters when residual bait anomalies occur, and a preset amount of bait is penalized during the next feeding.

[0041] A second aspect of this invention provides a dynamic perception-based feeding optimization system for red snapper farming. The system includes a memory and a processor. The memory contains a dynamic perception-based feeding optimization method program for red snapper farming. When executed by the processor, the dynamic perception-based feeding optimization method program for red snapper farming implements the following steps:

[0042] Fish condition monitoring and water condition monitoring are carried out in the target aquaculture area to obtain aquaculture area perception information. The aquaculture area perception information is preprocessed to generate an aquaculture perception dataset.

[0043] Based on the aquaculture perception dataset, unimodal features are extracted, and feature fusion is performed on the extracted unimodal features to generate multimodal aquaculture fusion features. The multimodal aquaculture fusion features are then used to model appetite state to obtain an appetite state index.

[0044] The feeding strategy is optimized by determining whether feeding is necessary based on the appetite state index and combining the multimodal aquaculture integration characteristics.

[0045] After each feeding, uneaten feed is collected and the environment is monitored. The correlation between uneaten feed and environmental parameters is analyzed, and a penalty mechanism for uneaten feed is established to optimize the feeding level in aquaculture.

[0046] This invention discloses a method and system for optimizing feeding in red snapper farming based on dynamic perception. The method includes: acquiring perception information of the farming area; preprocessing the perception information to generate a farming perception dataset; extracting single-modal features from the dataset; fusing the extracted single-modal features to generate multi-modal farming fusion features; using the multi-modal farming fusion features to model appetite state and obtain an appetite state index; determining whether feeding is necessary based on the appetite state index; and optimizing the feeding strategy by combining the multi-modal farming fusion features; collecting uneaten feed and monitoring the environment after each feeding; analyzing the correlation between uneaten feed and environmental parameters; and establishing a uneaten feed penalty mechanism to optimize the feeding level. This method achieves precise perception of feeding and avoids farming risks caused by overfeeding, thereby improving the efficiency and intelligence of red snapper farming. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0048] Figure 1 This is a flowchart of a method for optimizing feeding in redfin snapper farming based on dynamic perception, provided in an embodiment of the present invention.

[0049] Figure 2 This is a flowchart of a method for optimizing feed placement strategy in redfin snapper farming, provided by an embodiment of the present invention.

[0050] Figure 3 This is a block diagram of a dynamic perception-based feeding optimization system for redfin snapper farming, provided in one embodiment of the present invention.

[0051] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0052] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0054] Figure 1 This is a flowchart of a method for optimizing feeding in redfin snapper farming based on dynamic perception, provided in an embodiment of the present invention.

[0055] like Figure 1 As shown, this invention provides a flowchart of a method for optimizing feeding in redfin snapper farming based on dynamic perception, including:

[0056] S102, monitor the fish condition and water condition of the target aquaculture area to obtain aquaculture area perception information, preprocess the aquaculture area perception information to generate aquaculture perception dataset;

[0057] S104, extract single-modal features based on the aquaculture perception dataset, perform feature fusion based on the extracted single-modal features to generate multimodal aquaculture fusion features, and use the multimodal aquaculture fusion features to model appetite state and obtain appetite state index;

[0058] S106, determine whether feeding is necessary based on the appetite state index, and optimize the feeding strategy by combining the multimodal aquaculture integration characteristics;

[0059] S108 involves collecting uneaten feed and monitoring the environment after each feeding, analyzing the correlation between uneaten feed and environmental parameters, and establishing a penalty mechanism for uneaten feed to optimize the feeding level in aquaculture.

[0060] Furthermore, in a preferred embodiment of the present invention, the step of monitoring the fish condition and water condition of the target aquaculture area to obtain aquaculture area perception information, and preprocessing the aquaculture area perception information to generate an aquaculture perception dataset, specifically includes:

[0061] A heterogeneous sensor network is deployed in the target aquaculture area to monitor the aquaculture ponds in the target aquaculture area and obtain aquaculture area perception information, which includes fish body perception information and water body perception information.

[0062] Adaptive gamma correction technology is used to preprocess the image data in the fish's sensory information. The pixel values ​​are dynamically adjusted according to the brightness distribution of each frame. Then, the structure of similar image blocks is analyzed by non-local mean denoising algorithm to filter out the interference of suspended particles and output an enhanced image sequence.

[0063] Doppler frequency shift correction is performed on the sonar data in the fish body perception information, coordinate offset is compensated according to the sensor motion posture, and an outlier elimination algorithm based on kurtosis detection is used to eliminate occasional equipment vibration interference, generating a three-dimensional motion trajectory sequence.

[0064] Based on the enhanced image sequence and the three-dimensional motion trajectory sequence, preprocessed fish body perception information is generated, the standard deviation of the water body perception information is calculated, and outlier detection and missing value supplementation are performed in combination with the preset standard deviation range to generate preprocessed water body perception information.

[0065] The pre-processed fish body perception information and the pre-processed water body perception information are combined to form an aquaculture perception dataset.

[0066] It should be noted that a heterogeneous sensor network was deployed in the redfin snapper rearing pond. This network included a visible-infrared dual-mode camera group, a multi-band sonar array, and distributed water quality probes. The camera group acquired images of fish activity from the water surface to the bottom of the pond at a sampling rate of 25 frames per second. The sonar array scanned every 2 seconds to generate three-dimensional biological movement trajectories. The water quality probes collected water parameters such as dissolved oxygen, turbidity, and temperature in real time. This formed a perception information of the aquaculture area, including the distribution of fish positions, swimming postures, and water physicochemical indicators. Subsequently, the image data stream was input into an adaptive gamma correction module, which dynamically calculated the optimal gamma coefficient based on the histogram of each frame (γ = 1.8 for dark scenes and γ = 0.6 for strong reflective scenes) to compensate for the uneven light intensity caused by light refraction in the water. The corrected images were analyzed by a non-local mean denoising processor, which searched for structurally similar blocks within a 10×10 pixel area and performed joint filtering to effectively filter out point noise caused by plankton, outputting an enhanced image sequence. Sonar data is fed into a motion compensation channel, where it is combined with equipment tilt data recorded by sensors. A three-dimensional rotation matrix is ​​used to correct coordinate offsets. Subsequently, a kurtosis detection algorithm (32 samples per window) is employed to remove abnormal spikes caused by equipment vibration, generating a three-dimensional motion trajectory sequence. For aquatic environmental parameters, the standard deviation of hourly data for each parameter is calculated; periods exceeding a preset standard deviation are automatically marked as abnormal. Missing data is extracted vertically from neighboring sensor data and horizontally loaded with historical averages from the same period to generate a complete and continuous time series of aquaculture parameters, ultimately yielding the aquaculture sensing dataset.

[0067] Furthermore, in a preferred embodiment of the present invention, the step of extracting single-modal features based on the aquaculture perception dataset and generating multimodal aquaculture fusion features by fusing the extracted single-modal features specifically includes:

[0068] We acquired an aquaculture perception dataset and input the enhanced image sequence into a pre-trained four-channel YOLOv5 detection network. We set an extended near-infrared spectral channel in the input layer and extracted visible light and infrared band features simultaneously through a four-dimensional convolution kernel.

[0069] An embedded channel attention module is used in the backbone network to enhance the thermal signal response region of the organism based on spectral energy, while a spatial attention mechanism is introduced to focus on the contour change features of the moving target.

[0070] The pooling layer is used to obtain the fish body bounding box coordinates, confidence scores, and spatial mask matrix of fish distribution. A dynamic fish density index is generated. The inverse distance weighted interpolation method is used to map the position coordinates of all detected targets into a two-dimensional planar heat map to generate fish aggregation distribution features and obtain the first modality features.

[0071] The three-dimensional motion trajectory sequence is extracted from the aquaculture perception dataset. The instantaneous velocity of the individual is calculated based on the Euclidean distance between adjacent timestamps. The median value of all target velocities is taken as the average swimming speed index of the fish group. The discrete angular standard deviation of all fish movement directions within a preset period is calculated using the direction vector analysis method to generate the second modal feature.

[0072] The pre-processed water body sensing information is obtained through the aquaculture sensing dataset and an environmental index matrix is ​​generated. The environmental anomaly index is calculated based on the dissolved oxygen change rate, temperature gradient abrupt change point and turbidity cumulative fluctuation, and the third modal feature is obtained.

[0073] The first modal feature, the second modal feature, and the third modal feature are input into a gated recurrent unit network and combined with a multi-head attention mechanism to calculate the feature fusion weights, thereby generating multimodal aquaculture fusion features.

[0074] It should be noted that the enhanced image sequences from the aquaculture perception dataset are input into a four-channel YOLOv5 detection network. The network input layer expands the near-infrared spectral dimension, enabling the four-dimensional convolutional kernel to simultaneously resolve visible light textures and infrared thermal radiation features, effectively penetrating the interference of high-turbidity water. The backbone network embeds a dual-path attention mechanism, which enhances the response of fish thermal signal regions based on spectral energy distribution, while the spatial attention mechanism focuses on abrupt change zones in the contours of moving targets (such as the tail fin swinging area). The forward propagation to the pooling layer outputs two core data paths: a localization vector containing the target detection box coordinates and confidence scores, and a spatial mask matrix representing the fish population density. Based on the localization vector, the effective number of fish per unit water volume is calculated in real time to generate a dynamic fish population density index; simultaneously, an inverse distance weighted interpolation algorithm is applied to map the detection location coordinates into a two-dimensional heatmap, thereby extracting the spatial features of fish population aggregation and distribution, forming the first modality feature vector. For continuous timestamp coordinate data acquired by sonar, the instantaneous velocity of each target is calculated using the Euclidean distance between adjacent frames. After median filtering, the median velocity of all targets is taken as the quantified value of the average swimming speed of the fish swarm. The direction analysis module calculates the angular standard deviation of the fish movement direction vector with a period of 2 seconds, generating a directional discrete entropy index. The velocity and entropy values ​​are compressed to the 0-1 interval using a linear normalization model and fused to form the second modal feature. Preprocessed water body sensing information is obtained from the aquaculture sensing dataset to generate an environmental indicator matrix. An environmental anomaly index is generated based on the dissolved oxygen change rate, temperature gradient abrupt change point, and turbidity cumulative fluctuation, and the environmental anomaly index is calculated. Combined with the environmental indicator matrix, the third modal feature is generated. The first, second, and third modal features are integrated into a gate control recurrent unit (GRU) fusion network. The network is configured with a three-stage processing phase: First, the GRU unit learns the temporal dependencies of the three-modal features and outputs the cross-period associated hidden states; then, a multi-head attention mechanism is connected—the spatial head focuses on the changing trends of high-density areas in the heatmap, the behavioral head tracks the frequency of swimming speed surge events, and the environmental head monitors the slope of abnormal index fluctuations; finally, the feature contribution rate is dynamically allocated according to the attention weight to generate a multi-modal aquaculture fusion feature vector containing the coupling effect of biological behavioral response and environment, providing a high-order perceptual foundation for appetite state modeling.

[0075] Furthermore, in a preferred embodiment of the present invention, the step of using the multimodal aquaculture fusion features to model appetite state and obtain an appetite state index specifically includes:

[0076] Obtain multimodal aquaculture fusion features, calculate the original value of cluster density of the target aquaculture area based on the multimodal aquaculture fusion features, normalize the cluster density and linearly scale it based on the historical maximum density of the day to obtain the cluster density index;

[0077] Based on the multimodal aquaculture fusion features, the original value of the vertical activity entropy of the target aquaculture area is calculated using the Shannon entropy principle, and then divided by the preset theoretical maximum entropy value to obtain the vertical distribution index.

[0078] The appetite status of redfin snapper in the target aquaculture area is modeled using the aggregation density index, vertical distribution index, and multimodal aquaculture fusion features based on a deep neural network.

[0079] A bidirectional GRU unit is set up at the network front end to perform time-series dependency analysis. Two hidden layers are used to capture the precursor patterns of appetite changes. An attention mechanism is introduced at the network back end to build a triple parallel analysis channel, namely the spatial channel, the behavioral channel and the environmental channel.

[0080] The output values ​​of the triple parallel analysis channels are obtained. The input value aggregation unit performs weighted calculations according to preset weights and uses the Sigmoid function to compress the weighted results to generate an appetite state index.

[0081] It should be noted that the aggregation density calculation of the target aquaculture area is performed based on the acquired multimodal aquaculture fusion features. Raw spatial location data of fish schools is extracted from the fusion features, and the number of fish per unit water volume is used as the baseline value for aggregation density. The dynamic normalization module is input: the historical peak density of the day (e.g., the maximum monitored value of 350 fish / m³) is used as the normalization baseline, and the current density is mapped to a 0-1 standardized interval using a linear scaling formula, generating an aggregation density index (CDI) that updates over time. This index is refreshed every 5 seconds and written to a cache queue. Simultaneously, the vertical activity distribution characteristics of the water body are calculated. Sonar vertical motion trajectory data is decomposed from the multimodal features, and the activity disorder is quantified using the Shannon entropy principle: water layers are divided based on preset depth units, the frequency distribution of fish residence in each layer is statistically analyzed, and the raw entropy value is calculated. This raw value is divided by the preset theoretical maximum entropy value to generate an index (VDI) characterizing the vertical distribution balance of the fish school; a lower value indicates that the fish school is concentrating and migrating to a specific water layer. Subsequently, the Cluster Density Index (CDI), Vertical Distribution Index (VDI), and multimodal aquaculture fusion features are input into a deep neural network (DNN) to model appetite status. The network front-end is configured with a bidirectional gated recurrent unit (GRU) and two 32-node hidden layers to process continuous feature sequences and capture precursor patterns of appetite changes (e.g., the exponential growth trend of CDI and the synchronous decline of VDI 20 minutes before feeding). The back-end introduces a triple parallel attention channel: a spatial channel focuses on the minute-level change rate of CDI within a 0.5 m² feeding area; a behavioral channel analyzes the proportion of sonar high-frequency motion pulse energy in the fusion features; and an environmental channel monitors the coupling fluctuations of dissolved oxygen hourly change rate and temperature gradient. Next, the output values ​​of each channel are fed into a value aggregation unit. Preset weights are dynamically allocated according to the aquaculture water state (40% for the spatial channel, 35% for the behavioral channel, and 25% for the environmental channel), and a weighted summation operation is performed. The aggregation result is nonlinearly compressed using a sigmoid function and mapped to a continuous 0-10 scale, ultimately outputting the Appetite Status Index (ASI).

[0082] Furthermore, in a preferred embodiment of the present invention, the step of determining whether feeding is necessary based on the appetite state index and optimizing the feeding strategy by combining the multimodal aquaculture integration characteristics specifically includes:

[0083] The appetite state index and multimodal aquaculture fusion features are obtained. The appetite state index is compared with the preset feeding judgment rules. If the appetite state index is greater than a certain threshold, it means that feed needs to be fed and the feeding command is triggered.

[0084] If the appetite index is less than a Class I threshold and greater than a Class II threshold, a delayed feeding instruction is triggered; if the appetite index is less than a Class II threshold, it means that no feeding is needed, and a feeding freeze instruction is triggered.

[0085] When it is necessary to feed the redfin snapper in the target aquaculture area, the feeding strategy is optimized by using the multimodal aquaculture fusion feature. The feeding area is gridded to generate several grid cells, and the multimodal aquaculture fusion feature is aligned with each grid cell.

[0086] Acquire the cluster density index and vertical distribution index, define candidate feeding grid cells using the cluster density index and vertical distribution index, and generate the first dataset based on the multimodal aquaculture fusion features corresponding to each candidate feeding grid cell;

[0087] A Bayesian inference network is introduced. The first dataset is input into the Bayesian inference network for training to obtain the transient posterior distribution. The transient posterior distribution is used to infer the movement state of the fish in each candidate feeding grid cell at the next moment in the Bayesian network.

[0088] The next moment motion state is obtained by inference and secondary feeding grid cells are defined based on the inferred probability value. The secondary feeding grid cells are merged with the candidate feeding grid cells to generate recommended feeding areas for aquaculture feeding optimization.

[0089] It should be noted that the feeding decision triggering mechanism is executed first: the current ASI is compared with a preset threshold. When the ASI remains above the first-class threshold (set ≥7.2) for 10 consecutive minutes, an immediate feeding command is triggered; if the ASI is between the first-class and second-class thresholds (5.0≤ASI<7.2), a 30-minute delayed feeding command is initiated; when the ASI is below the second-class threshold (<5.0), a feeding freeze signal is generated and sent to the control system. When an immediate or delayed feeding command is triggered, the intelligent area optimization system is activated. The target aquaculture pond is divided into three-dimensional meshes to generate rasterized units with depth stratification. The multimodal aquaculture fusion features are aligned with the raster units in real time, and the fish distribution heat values ​​and environmental parameters within each unit are extracted. Focusing on high-potential areas: raster units with CDI > 0.6 and VDI < 0.3 (indicating high-density aggregation and concentrated vertical activity) are selected and marked as candidate feeding units, and the corresponding multimodal aquaculture fusion features within the units are extracted simultaneously to form the first dataset. The dataset is input into a Bayesian inference network, which trains a transient posterior distribution model: predicting the probability distribution of future fish movements given the current state of the fish population. Secondary feeding units are defined based on the probability output: neighboring units with predicted probabilities greater than a preset threshold are included in the expanded region. Candidate units and secondary units are merged to generate the final recommended feeding area. The region output is a spatial bounding box, which is overlaid on a digital twin of the aquaculture pond using a 3D visualization engine. This data is then converted into pitch and azimuth control signals for the feeder to drive the actuator, thereby optimizing feed utilization.

[0090] Figure 2This is a flowchart of a method for optimizing feed placement strategy in redfin snapper farming, provided by an embodiment of the present invention.

[0091] like Figure 2 As shown, this invention provides a flowchart of a method for optimizing feed placement strategies in redfin snapper farming, including:

[0092] Furthermore, in a preferred embodiment of the present invention, the step of collecting uneaten feed and monitoring the environment after each feeding, analyzing the correlation between uneaten feed and environmental parameters, and establishing a uneaten feed penalty mechanism to optimize the feeding level in aquaculture specifically includes:

[0093] S202, a grid-type residual feed collection device is laid in the aquaculture pond. The residual feed within a preset time after each feeding is obtained by suction filtration and weighed. At the same time, the water parameters before and after feeding are monitored by the deployed heterogeneous sensors to obtain aquaculture feeding perception information.

[0094] S204. Based on the aquaculture feeding perception information, generate a heat map of residual feed distribution. When the amount of residual feed corresponding to a certain grid unit exceeds a preset threshold, it is marked as an abnormal residual feed area. The corresponding water parameters are extracted through the aquaculture feeding perception information to generate abnormal residual feed area information.

[0095] S206. Based on the information on abnormal residual bait areas, a covariance matrix is ​​constructed with the amount of residual bait exceeding the standard as the dependent variable and the fluctuation of water environmental parameters as the independent variable. The singular value decomposition method is used to perform eigenvalue decomposition on the covariance matrix to analyze the correlation strength between the amount of residual bait exceeding the standard and each water environmental parameter.

[0096] S208 determines the correlation strength between the excessive amount of uneaten food and various water environmental parameters and preset thresholds, analyzes the main water environmental factors that lead to abnormal uneaten food, and generates information on the cause of abnormal uneaten food.

[0097] S210, establish a residual bait penalty mechanism based on the residual bait abnormality attribution information, establish a residual bait penalty mechanism based on the water environment parameters when residual bait abnormality occurs, and penalize the preset amount of bait when feeding next time.

[0098] It should be noted that a high-precision grid-type uneaten feed collection device is installed in the aquaculture pond, with each unit equipped with an independent suction and filtration system. After each feeding operation, a timed suction program is activated to perform negative pressure adsorption on all units. The net weight of uneaten feed in each unit is measured using a precision electronic scale, and the deployed dissolved oxygen sensors, water temperature detectors, and turbidity monitors are simultaneously activated to collect the water parameter fluctuation sequence of the target area before and after feeding. The weighing data and sensor records are integrated to generate structured aquaculture feeding perception information. Based on the aquaculture feeding perception information, a heat map of uneaten feed distribution is generated: the weight of uneaten feed in each unit is mapped to a thermal gradient. When the amount of uneaten feed in a specific unit exceeds a preset dynamic threshold twice consecutively (threshold = historical average of the unit × 130%), it is automatically marked as an abnormal uneaten feed area. The time series of water parameters associated with this area (including dissolved oxygen concentration decrease curve, water temperature jump fluctuation value, turbidity increase rate, etc.) are extracted to construct an abnormal uneaten feed area information matrix. This matrix includes spatial coordinates, uneaten feed exceeding the limit, and abnormal water environment indicators. A covariance matrix is ​​constructed with the amount of uneaten bait exceeding the standard as the dependent variable and the fluctuation of water parameters as the independent variables. The matrix is ​​decomposed using singular value decomposition (SVD), and the loadings of each eigenvector are calculated. When the loading of a parameter in the dominant eigenvector is >0.7 (e.g., the weight of water temperature fluctuation variance is 0.82), this parameter is determined to be a key cause of uneaten bait anomalies. If the loadings are evenly distributed (maximum value <0.5), it is marked as a complex environmental stress effect. The analysis results are output as the quantitative value of the correlation strength between uneaten bait exceeding the standard and water parameters, and the code of the dominant factor. When the correlation strength of a single parameter is >0.65, a clearly directional attribution information for uneaten bait anomalies is generated; if the correlation strengths of multiple parameters are all in the range of 0.4-0.6, a composite attribution label is generated. The attribution information for uneaten bait anomalies is written into a correction knowledge base for adjusting the feeding strategy. Based on the attribution conclusions, an uneaten bait penalty mechanism is established: using the key water parameters at the time of the anomaly as a benchmark, a parameter-penalty coefficient mapping rule is set. During the next feeding instruction generation phase, the water parameters of the area to be fed are compared with the penalty benchmark value in real time. When the similarity is >80% (e.g., current water temperature 28.3℃, dissolved oxygen 5.3mg / L), a dynamic penalty algorithm is triggered to reduce the preset feeding amount by 20-35% (the penalty range is calculated based on the historical residual feed exceedance rate), and the optimized feeding amount is locked to drive the execution unit. By continuously updating the penalty weight coefficient until the residual feed amount in the corresponding area returns to the threshold range, a closed-loop control from anomaly monitoring to decision optimization is completed.

[0099] Figure 3 A dynamic perception-based feeding optimization system 3 for red snapper farming is provided in one embodiment of the present invention. The system includes: a memory 31 and a processor 32. The memory 31 contains a dynamic perception-based feeding optimization method program for red snapper farming. When the dynamic perception-based feeding optimization method program for red snapper farming is executed by the processor 32, it performs the following steps:

[0100] Fish condition monitoring and water condition monitoring are carried out in the target aquaculture area to obtain aquaculture area perception information. The aquaculture area perception information is preprocessed to generate an aquaculture perception dataset.

[0101] Based on the aquaculture perception dataset, unimodal features are extracted, and feature fusion is performed on the extracted unimodal features to generate multimodal aquaculture fusion features. The multimodal aquaculture fusion features are then used to model appetite state to obtain an appetite state index.

[0102] The feeding strategy is optimized by determining whether feeding is necessary based on the appetite state index and combining the multimodal aquaculture integration characteristics.

[0103] After each feeding, uneaten feed is collected and the environment is monitored. The correlation between uneaten feed and environmental parameters is analyzed, and a penalty mechanism for uneaten feed is established to optimize the feeding level in aquaculture.

[0104] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0105] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0106] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0107] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0109] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for optimizing feeding in redfin snapper farming based on dynamic perception, characterized in that, include: Fish condition monitoring and water condition monitoring are carried out in the target aquaculture area to obtain aquaculture area perception information. The aquaculture area perception information is preprocessed to generate an aquaculture perception dataset. Based on the aquaculture perception dataset, unimodal features are extracted, and feature fusion is performed on the extracted unimodal features to generate multimodal aquaculture fusion features. The multimodal aquaculture fusion features are then used to model appetite state to obtain an appetite state index. The feeding strategy is optimized by determining whether feeding is necessary based on the appetite state index and combining the multimodal aquaculture integration characteristics. After each feeding, uneaten feed is collected and the environment is monitored. The correlation between uneaten feed and environmental parameters is analyzed, and a penalty mechanism for uneaten feed is established to optimize the feeding level in aquaculture.

2. The method for optimizing feeding in redfin snapper farming based on dynamic perception as described in claim 1, characterized in that, The process of monitoring fish and water conditions in the target aquaculture area to obtain aquaculture area perception information, and preprocessing this perception information to generate an aquaculture perception dataset, specifically includes: A heterogeneous sensor network is deployed in the target aquaculture area to monitor the aquaculture ponds in the target aquaculture area and obtain aquaculture area perception information, which includes fish body perception information and water body perception information. Adaptive gamma correction technology is used to preprocess the image data in the fish's sensory information. The pixel values ​​are dynamically adjusted according to the brightness distribution of each frame. Then, the structure of similar image blocks is analyzed by non-local mean denoising algorithm to filter out the interference of suspended particles and output an enhanced image sequence. Doppler frequency shift correction is performed on the sonar data in the fish body perception information, coordinate offset is compensated according to the sensor motion posture, and an outlier elimination algorithm based on kurtosis detection is used to eliminate occasional equipment vibration interference, generating a three-dimensional motion trajectory sequence. Based on the enhanced image sequence and the three-dimensional motion trajectory sequence, preprocessed fish body perception information is generated, the standard deviation of the water body perception information is calculated, and outlier detection and missing value supplementation are performed in combination with the preset standard deviation range to generate preprocessed water body perception information. The pre-processed fish body perception information and the pre-processed water body perception information are combined to form an aquaculture perception dataset.

3. The method for optimizing feeding in redfin snapper farming based on dynamic perception as described in claim 1, characterized in that, The step of extracting unimodal features based on the aquaculture perception dataset and then performing feature fusion based on the extracted unimodal features to generate multimodal aquaculture fusion features specifically includes: We acquired an aquaculture perception dataset and input the enhanced image sequence into a pre-trained four-channel YOLOv5 detection network. We set an extended near-infrared spectral channel in the input layer and extracted visible light and infrared band features simultaneously through a four-dimensional convolution kernel. An embedded channel attention module is used in the backbone network to enhance the thermal signal response region of the organism based on spectral energy, while a spatial attention mechanism is introduced to focus on the contour change features of the moving target. The pooling layer is used to obtain the fish body bounding box coordinates, confidence scores, and spatial mask matrix of fish distribution. A dynamic fish density index is generated. The inverse distance weighted interpolation method is used to map the position coordinates of all detected targets into a two-dimensional planar heat map to generate fish aggregation distribution features and obtain the first modality features. The three-dimensional motion trajectory sequence is extracted from the aquaculture perception dataset. The instantaneous velocity of the individual is calculated based on the Euclidean distance between adjacent timestamps. The median value of all target velocities is taken as the average swimming speed index of the fish group. The discrete angular standard deviation of all fish movement directions within a preset period is calculated using the direction vector analysis method to generate the second modal feature. The pre-processed water body sensing information is obtained through the aquaculture sensing dataset and an environmental index matrix is ​​generated. The environmental anomaly index is calculated based on the dissolved oxygen change rate, temperature gradient abrupt change point and turbidity cumulative fluctuation, and the third modal feature is obtained. The first modal feature, the second modal feature, and the third modal feature are input into a gated recurrent unit network and combined with a multi-head attention mechanism to calculate the feature fusion weights, thereby generating multimodal aquaculture fusion features.

4. The method for optimizing feeding in redfin snapper farming based on dynamic perception as described in claim 1, characterized in that, The process of modeling appetite state using the multimodal aquaculture fusion features to obtain an appetite state index specifically includes: Obtain multimodal aquaculture fusion features, calculate the original value of cluster density of the target aquaculture area based on the multimodal aquaculture fusion features, normalize the cluster density and linearly scale it based on the historical maximum density of the day to obtain the cluster density index; Based on the multimodal aquaculture fusion features, the original value of the vertical activity entropy of the target aquaculture area is calculated using the Shannon entropy principle, and then divided by the preset theoretical maximum entropy value to obtain the vertical distribution index. The appetite status of redfin snapper in the target aquaculture area is modeled using the aggregation density index, vertical distribution index, and multimodal aquaculture fusion features based on a deep neural network. A bidirectional GRU unit is set up at the network front end to perform time-series dependency analysis. Two hidden layers are used to capture the precursor patterns of appetite changes. An attention mechanism is introduced at the network back end to build a triple parallel analysis channel, namely the spatial channel, the behavioral channel and the environmental channel. The output values ​​of the triple parallel analysis channels are obtained. The input value aggregation unit performs weighted calculations according to preset weights and uses the Sigmoid function to compress the weighted results to generate an appetite state index.

5. The method for optimizing feeding in redfin snapper farming based on dynamic perception as described in claim 1, characterized in that, The step of determining whether feeding is necessary based on the appetite index and optimizing the feeding strategy by combining the multimodal aquaculture integration characteristics specifically includes: The appetite state index and multimodal aquaculture fusion features are obtained. The appetite state index is compared with the preset feeding judgment rules. If the appetite state index is greater than a certain threshold, it means that feed needs to be fed and the feeding command is triggered. If the appetite index is less than a Class I threshold and greater than a Class II threshold, a delayed feeding instruction is triggered; if the appetite index is less than a Class II threshold, it means that no feeding is needed, and a feeding freeze instruction is triggered. When it is necessary to feed the redfin snapper in the target aquaculture area, the feeding strategy is optimized by using the multimodal aquaculture fusion feature. The feeding area is gridded to generate several grid cells, and the multimodal aquaculture fusion feature is aligned with each grid cell. Acquire the cluster density index and vertical distribution index, define candidate feeding grid cells using the cluster density index and vertical distribution index, and generate the first dataset based on the multimodal aquaculture fusion features corresponding to each candidate feeding grid cell; A Bayesian inference network is introduced. The first dataset is input into the Bayesian inference network for training to obtain the transient posterior distribution. The transient posterior distribution is used to infer the movement state of the fish in each candidate feeding grid cell at the next moment in the Bayesian network. The next moment motion state is obtained by inference and secondary feeding grid cells are defined based on the inferred probability value. The secondary feeding grid cells are merged with the candidate feeding grid cells to generate recommended feeding areas for aquaculture feeding optimization.

6. The method for optimizing feeding in redfin snapper farming based on dynamic perception according to claim 1, characterized in that, The process of collecting uneaten feed and monitoring the environment after each feeding, analyzing the correlation between uneaten feed and environmental parameters, and establishing a penalty mechanism for uneaten feed to optimize the feeding level in aquaculture specifically includes: A grid-type residual feed collection device is laid in the aquaculture pond. The residual feed is collected and weighed within a preset time after each feeding by suction and filtration. At the same time, heterogeneous sensors are deployed to monitor water parameters before and after feeding to obtain aquaculture feeding perception information. Based on the aquaculture feeding perception information, a heat map of residual feed distribution is generated. When the amount of residual feed corresponding to a certain grid unit exceeds a preset threshold, it is marked as an abnormal residual feed area. The corresponding water parameters are extracted through the aquaculture feeding perception information to generate abnormal residual feed area information. Based on the information on abnormal residual bait areas, a covariance matrix is ​​constructed with the amount of residual bait exceeding the standard as the dependent variable and the fluctuation of water body environmental parameters as the independent variable. The singular value decomposition method is used to perform eigenvalue decomposition on the covariance matrix to analyze the correlation strength between the amount of residual bait exceeding the standard and each water body environmental parameter. The correlation strength between the amount of uneaten bait exceeding the standard and various water environmental parameters is judged against the preset threshold. The main water environmental factors that lead to abnormal uneaten bait are analyzed, and information on the cause of abnormal uneaten bait is generated. A residual bait penalty mechanism is established based on the residual bait anomaly attribution information. The residual bait penalty mechanism is established based on the water environment parameters when residual bait anomalies occur, and a preset amount of bait is penalized during the next feeding.

7. A feeding optimization system for redfin snapper farming based on dynamic perception, characterized in that, The system includes a memory and a processor. The memory contains a program for optimizing feeding methods for red snapper farming based on dynamic perception. When the processor executes the program for optimizing feeding methods for red snapper farming based on dynamic perception, it performs the following steps: Fish condition monitoring and water condition monitoring are carried out in the target aquaculture area to obtain aquaculture area perception information. The aquaculture area perception information is preprocessed to generate an aquaculture perception dataset. Based on the aquaculture perception dataset, unimodal features are extracted, and feature fusion is performed on the extracted unimodal features to generate multimodal aquaculture fusion features. The multimodal aquaculture fusion features are then used to model appetite state to obtain an appetite state index. The feeding strategy is optimized by determining whether feeding is necessary based on the appetite state index and combining the multimodal aquaculture integration characteristics. After each feeding, uneaten feed is collected and the environment is monitored. The correlation between uneaten feed and environmental parameters is analyzed, and a penalty mechanism for uneaten feed is established to optimize the feeding level in aquaculture.

Images