Precise feeding method of feeding robot based on pattern recognition

Abstract

The invention discloses a mode recognition-based precise feeding method of a feeding robot, which comprises the following steps of: acquiring behavior observation data of a breeding object to obtain an effective behavior sequence; constructing a behavior time sequence characterization module to obtain a behavior potential characterization sequence; organizing and forming a behavior unit representation sequence, and constructing a coupling relationship between behavior phase variables; inputting an improved Kuramoto model to obtain a synchronization degree index and a synchronization degree evolution sequence; performing mode judgment on the group feeding behavior, and generating a feeding trigger signal; a feeding control parameter set is generated, and the feeding robot is controlled to execute feeding operation. According to the invention, by introducing the comparative predictive codes and combining with the improved Kuramoto model, accurate identification of real feeding behaviors and intelligent accurate feeding control of the feeding robot in a group breeding scene are realized.

Description

Technical Field

[0001] This invention relates to the field of pattern recognition technology, and in particular to a method for precise feeding of a feeding robot based on pattern recognition. Background Technology

[0002] Currently, with the continuous improvement of the scale and automation level of aquaculture, the intelligent and precise feeding operation has gradually become an important technical direction in aquaculture management. In traditional technology, feeding operations mostly adopt timed and quantitative feeding or feeding methods triggered by simple sensor signals, and control feeding is achieved by preset feeding schedules, fixed feeding parameters, or single behavioral thresholds. Traditional methods usually cannot fully perceive the behavioral changes of farmed animals at the group level, especially when group behavior is affected by environmental fluctuations, individual differences, or external stimuli, making it difficult to accurately reflect the true feeding status.

[0003] At the level of behavioral analysis and intelligent decision-making, existing technologies are beginning to incorporate visual perception and motion analysis to identify the behavioral characteristics of farmed animals and trigger feeding operations accordingly. However, existing solutions lack the ability to model the continuous evolution of group behavior over time, making it difficult to distinguish between short-term behavioral activity triggered by environmental disturbances or conditioned reflexes and group feeding behavior formed and developed from genuine feeding intentions. Furthermore, existing methods still have shortcomings in multi-source behavioral feature fusion, characterization of group behavior consistency, and analysis of behavioral synchronization states. Feeding decisions often rely on judgments based on single moments or local features, which can easily lead to misjudgments of feeding timing or unreasonable feeding parameter settings in complex farming scenarios, thus limiting the precise feeding capabilities of feeding robots in group farming scenarios.

[0004] Therefore, how to provide a precise feeding method for feeding robots based on pattern recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a pattern recognition-based method for precise feeding of a feeding robot. This invention utilizes a combination of multi-source behavioral perception, temporal pattern modeling, and group synchronization analysis to accurately identify feeding behavior and make feeding decisions in group farming scenarios. The method collects visual behavior data, group motion characteristic data, spatial distribution characteristic data, and environmental correlation data of the farmed animals to construct a temporal representation of behavior. It introduces contrastive predictive coding to predict the evolutionary trend of group behavior and further combines an improved Kuramoto model to analyze and determine the group's feeding synchronization state. Based on the synchronization degree, stability, and spatial distribution characteristics of the group's feeding behavior, feeding control parameters such as feeding intensity, feeding rhythm, and spatial allocation are dynamically generated to achieve precise control of the feeding robot. This invention can effectively distinguish between short-term behavioral disturbances and genuine feeding intentions, improves the accuracy of feeding timing judgment, and has the advantages of high reliability in feeding decisions, high feed utilization efficiency, and strong adaptability to complex farming environments.

[0006] A method for precise feeding of a feeding robot based on pattern recognition according to an embodiment of the present invention includes:

[0007] Behavioral observation data of farmed animals in the feeding area of ​​a group farming scenario are collected, and the behavioral observation data is preprocessed to obtain effective behavioral sequences;

[0008] A behavior time series representation module based on contrastive predictive coding is constructed to model the temporal prediction relationship between adjacent time segments in the effective behavior sequence, and to obtain a behavior potential representation sequence that corresponds one-to-one with each time segment.

[0009] The potential behavioral representation sequences are organized into multiple behavioral unit representation sequences according to individual identifiers, and each behavioral unit representation sequence is mapped to a corresponding behavioral phase variable. At the same time, the coupling relationship between each behavioral phase variable is constructed based on the behavioral similarity relationship.

[0010] By inputting the behavioral phase variables and their corresponding coupling relationships into the improved Kuramoto model, population synchronization evolution calculations are performed to obtain synchronization indexes and synchronization evolution sequences.

[0011] The feeding behavior of the group is determined based on the synchronization index and the synchronization evolution sequence. When the synchronization index reaches the synchronization threshold and the synchronization maintenance condition is met within the continuous determination time window, a feeding trigger signal is generated.

[0012] Upon receiving a feeding trigger signal, a set of feeding control parameters corresponding to the current group feeding mode is generated, and the feeding robot is controlled to perform feeding operations according to the set of feeding control parameters.

[0013] Optionally, the behavioral observation data specifically includes visual behavioral data, group movement characteristic data, spatial distribution characteristic data, and environmental correlation data.

[0014] Optionally, the preprocessing of the behavioral observation data specifically includes noise suppression, anomaly processing, time alignment, spatial region constraints, serialization, and standardization.

[0015] Optionally, obtaining the behavioral potential representation sequence corresponding one-to-one with each time segment includes:

[0016] A behavioral temporal representation module based on contrastive predictive coding is constructed, which includes a temporal segment coding unit, a contextual temporal modeling unit, a future representation prediction unit, and a contrastive constraint representation learning unit.

[0017] The temporal segment coding unit divides the effective behavior sequence into time segments and performs coding processing on the behavior observation data corresponding to each time segment to obtain a behavior segment representation that corresponds one-to-one with each time segment.

[0018] The contextual temporal modeling unit performs temporal correlation modeling on the behavioral segment representations corresponding to multiple adjacent time segments in chronological order. At the same time, it introduces a time interval embedding channel to encode the time interval information between adjacent time segments, and then jointly represents the encoded time interval information with the temporal correlation results to form a contextual behavioral representation containing time interval features.

[0019] The future behavior representation prediction unit predicts the behavior representation corresponding to the next time segment based on different semantic level representations. It sets up semantic level sub-projection paths to perform hierarchical mapping of context behavior representations, and obtains predicted behavior representations that respectively represent local behavior change features and group behavior evolution trend features.

[0020] The contrastive constraint representation learning unit pairs the predicted behavior representation with the corresponding actual behavior segment representation of the next time segment, calculates the attractive error term between similar pairs and the repulsive error term between dissimilar pairs, and constructs a contrastive loss function based on the difference between the attractive and repulsive error terms.

[0021] Based on the contrastive loss function, the behavioral representation is updated, and the representation mapping space is optimized to keep the behavioral representations compact between similar behaviors and separate between dissimilar behaviors, and outputs a sequence of potential behavioral representations that corresponds one-to-one with each time segment.

[0022] Optionally, the step of constructing the coupling relationship between each behavioral phase variable based on behavioral similarity includes:

[0023] Based on behavioral observation data, individual identifiers of each aquaculture object in the feeding area are obtained, and the individual identifiers are bound to the correspondence between each time segment in the potential behavioral representation sequence;

[0024] The potential behavioral representation sequences are aggregated according to individual identifiers. The potential behavioral representations of the same breeding object in consecutive time segments are organized in chronological order into behavioral unit representation sequences corresponding to the breeding object, forming a set of behavioral unit representation sequences corresponding to multiple breeding objects respectively.

[0025] Phase mapping processing is performed on each behavioral unit representation sequence in the behavioral unit representation sequence set to map the behavioral unit representation corresponding to each time segment to the corresponding behavioral phase variable, thereby obtaining the behavioral phase variable sequence corresponding to each breeding object in each time segment.

[0026] The behavioral similarity relationship between each pair of aquaculture objects is calculated based on the behavioral unit representation sequence set, and the coupling relationship between each behavioral phase variable is determined based on the behavioral similarity relationship.

[0027] Optionally, obtaining the synchronization index and the synchronization evolution sequence includes:

[0028] The behavioral phase variables and their corresponding coupling relationships are input into the improved Kuramoto model. Each aquaculture object in the feeding area is constructed as an oscillator, forming an oscillator set. A unique correspondence is established for each oscillator, and each oscillator corresponds one-to-one with a behavioral unit representation sequence.

[0029] Based on the sequence of behavioral phase variables, each oscillator is assigned a corresponding behavioral phase variable, and the behavioral phase variable is used as the state representation of the oscillator at the current time step.

[0030] Based on the behavioral potential characterization sequence and environmental correlation data, a corresponding dynamic evolution frequency is determined for each oscillator, and it serves as the basic evolution input for the oscillator under the condition that it is not affected by the remaining oscillators.

[0031] Based on the behavioral phase variables and their corresponding coupling relationships, the corresponding coupling strength is determined for any two oscillators. A phase coupling function term is constructed based on the behavioral phase difference between the two oscillators, and the coupling function term is combined with the corresponding coupling strength to form the coupling input.

[0032] During each time update process, a synchronous evolution update process is performed on the behavior phase variable of each oscillator. The dynamic evolution frequency of the oscillator is used as the basic evolution component, the coupling input is used as the interaction component, and the behavior phase variable is updated by combining the historical phase state and historical synchronization state of the oscillator to obtain the updated behavior phase variable.

[0033] Based on the updated behavioral phase variables of all oscillators, the synchronization index of the group as a whole at the current moment is calculated, and a synchronization evolution sequence is formed during the continuous time update process. The changing trend and fluctuation of the synchronization evolution sequence are statistically processed to obtain the synchronization evolution sequence.

[0034] Optionally, generating the feeding trigger signal includes:

[0035] Obtain the synchronization index and synchronization evolution sequence, and arrange the synchronization evolution sequence in chronological order to form the basic judgment sequence;

[0036] Based on the basic decision sequence, a continuous decision time window is constructed. Within each continuous decision time window, the synchronization index within the corresponding time period is extracted. Based on the synchronization index, the synchronization level, synchronization change trend and synchronization fluctuation degree within the time window are calculated respectively, forming a synchronization decision feature set corresponding to the time window.

[0037] Pre-set synchronization maintenance conditions, including synchronization level threshold conditions, synchronization change trend conditions, and synchronization fluctuation constraint conditions;

[0038] The synchronization judgment feature set is compared with the synchronization maintenance condition window by window. The group feeding behavior execution mode is judged based on the window by window comparison result. When the current continuous judgment time window meets the synchronization maintenance condition, the corresponding feeding trigger candidate signal is generated.

[0039] Perform continuous confirmation processing on the candidate feed trigger signal. When the synchronization maintenance condition is met in the continuous decision time window corresponding to multiple consecutive adjacent time update steps, output the feed trigger signal.

[0040] Optionally, generating the set of feeding control parameters corresponding to the current group feeding pattern includes:

[0041] Obtain the group feeding behavior pattern determination results, synchronization index and synchronization evolution sequence corresponding to the feeding trigger signal, and retrieve the current time data of visual behavior data, group movement characteristic data, spatial distribution characteristic data and environmental correlation data from the behavior observation data;

[0042] Based on the group feeding behavior pattern determination results and the spatial distribution characteristic data at the current moment, the set of target feeding sub-regions within the feeding area is determined, and the feeding priority order of each target feeding sub-region is determined based on the set of target feeding sub-regions;

[0043] Feeding intensity parameters are determined based on synchronization index and synchronization evolution sequence, feeding rhythm parameters are determined based on group movement characteristic data, and feeding compensation parameters are determined based on environmental correlation data, forming a set of feeding control parameters. The set of feeding control parameters includes feeding intensity parameters, feeding rhythm parameters, feeding compensation parameters, and regional feeding allocation parameters corresponding to the target feeding sub-region set.

[0044] The set of feeding control parameters is converted into a sequence of feeding control instructions that can be executed by the feeding robot. The sequence of feeding control instructions includes an arrival instruction for the target feeding sub-region, a feeding path instruction, a feeding discharge instruction, and a feeding rhythm control instruction.

[0045] The feeding robot is controlled to perform feeding operations according to the sequence of feeding control instructions.

[0046] The beneficial effects of this invention are:

[0047] This invention proposes a pattern recognition-based method for precise feeding of a feeding robot. By combining multi-source behavioral perception with temporal modeling of group behavior, it analyzes and determines the feeding behavior of farmed animals. The method provides a unified representation of group behavior and models the evolutionary relationship of group behavior over time by constructing a behavior temporal representation module based on contrastive predictive coding. This yields latent behavioral representations, maps individual behaviors to phase variables, and introduces an improved Kuramoto model to analyze the synchronization and evolutionary process of group behavior. This accurately identifies group feeding behaviors driven by genuine feeding intentions and possessing continuous characteristics.

[0048] This invention can stably determine the feeding status of a group in both time and space. Furthermore, by combining spatial distribution characteristics and environmental conditions, it dynamically generates control parameters such as feeding intensity, feeding rhythm, feeding compensation, and regional feeding allocation, achieving precise control of the feeding robot. This invention can effectively distinguish between short-term behavioral activity caused by environmental disturbances or conditioned reflexes and genuine feeding behavior, reducing the risk of feeding misjudgment, improving the accuracy of feeding timing judgment and the rationality of feeding resource allocation, and enhancing feed utilization efficiency and the level of intelligent aquaculture management. It has good practical value and promising prospects for widespread application. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 This is a flowchart of a precise feeding method for a feeding robot based on pattern recognition proposed in this invention;

[0051] Figure 2This is a schematic diagram of the behavior time sequence representation module of a pattern recognition-based precise feeding method for a feeding robot proposed in this invention.

[0052] Figure 3 This is a flowchart of the processing of the improved Kuramoto model for a pattern recognition-based precise feeding method for a feeding robot proposed in this invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figure 1 , Figure 2 and Figure 3 A method for precise feeding of a feeding robot based on pattern recognition, comprising:

[0055] Behavioral observation data of farmed animals in the feeding area of ​​a group farming scenario are collected, and the behavioral observation data is preprocessed to obtain effective behavioral sequences;

[0056] A behavior time series representation module based on contrastive predictive coding is constructed to model the temporal prediction relationship between adjacent time segments in the effective behavior sequence, and to obtain a behavior potential representation sequence that corresponds one-to-one with each time segment.

[0057] The potential behavioral representation sequences are organized into multiple behavioral unit representation sequences according to individual identifiers, and each behavioral unit representation sequence is mapped to a corresponding behavioral phase variable. At the same time, the coupling relationship between each behavioral phase variable is constructed based on the behavioral similarity relationship.

[0058] By inputting the behavioral phase variables and their corresponding coupling relationships into the improved Kuramoto model, population synchronization evolution calculations are performed to obtain synchronization indexes and synchronization evolution sequences.

[0059] The feeding behavior of the group is determined based on the synchronization index and the synchronization evolution sequence. When the synchronization index reaches the synchronization threshold and the synchronization maintenance condition is met within the continuous determination time window, a feeding trigger signal is generated.

[0060] Upon receiving a feeding trigger signal, a set of feeding control parameters corresponding to the current group feeding mode is generated, and the feeding robot is controlled to perform feeding operations according to the set of feeding control parameters.

[0061] In this embodiment, the behavioral observation data specifically includes visual behavioral data, group movement characteristic data, spatial distribution characteristic data, and environmental correlation data, wherein:

[0062] Visual behavioral data includes data on changes in the posture of farmed animals, frequency of feeding actions, changes in head orientation, amplitude of body swaying, and local aggregation of the population;

[0063] Group movement characteristic data refers to the movement speed data, movement direction change data, overall group movement consistency data, group movement synchronization data, and movement change data of the group transitioning from a static state to an active state.

[0064] Spatial distribution characteristic data includes spatial location data of the farmed animals within the feeding area, distance distribution data between individuals, population density data within a local area, and population aggregation center location data;

[0065] Environmental data includes ambient temperature data, ambient lighting conditions data, ambient ventilation status data, and environmental conditions changing over time.

[0066] In this embodiment, the preprocessing of behavioral observation data specifically includes noise suppression, anomaly processing, time alignment, spatial region constraints, serialization, and standardization.

[0067] In this embodiment, obtaining the behavioral potential representation sequence corresponding one-to-one with each time segment includes:

[0068] A behavior temporal representation module based on contrastive predictive coding is constructed. This module includes a temporal segment coding unit, a contextual temporal modeling unit, a future representation prediction unit, and a contrastive constraint representation learning unit. Specifically, the construction of this behavior temporal representation module based on contrastive predictive coding is as follows:

[0069] The temporal segment encoding unit, the contextual temporal modeling unit, the future representation prediction unit, and the contrastive constraint representation learning unit are sequentially connected to form the behavioral temporal representation module;

[0070] The temporal segment encoding unit divides the effective behavior sequence into time segments and performs encoding processing on the behavior observation data corresponding to each time segment to obtain a behavior segment representation that corresponds one-to-one with each time segment. Specifically, the encoding processing on the behavior observation data corresponding to each time segment is as follows:

[0071] For each time segment, the visual behavior features, group movement features, spatial distribution features, and environmental association features within the time segment are represented as corresponding feature vectors. The feature vectors are scaled and combined according to a unified feature arrangement order. The combined feature vectors are dimensionally compressed and reconstructed by combining linear combination and nonlinear transformation. Feature vectors of different types and dimensions are converted into feature representations of a unified dimension. The feature representations are fused within the same feature representation space to obtain the behavioral segment representation.

[0072] The contextual temporal modeling unit performs temporal correlation modeling on the behavioral segment representations corresponding to multiple adjacent time segments in chronological order. Simultaneously, it introduces a time interval embedding channel to encode the time interval information between adjacent time segments. The encoded time interval information is then jointly represented with the temporal correlation results to form a contextual behavioral representation that includes time interval features.

[0073] The temporal correlation modeling is performed as follows:

[0074] According to the time sequence, the sequential relationship and dependency between the behavioral segment representations corresponding to multiple adjacent time segments are modeled and processed. The sequence of behavioral segment representations corresponding to consecutive time segments is taken as input, and the behavioral segment representations of adjacent time segments are correlated and processed to transform the discrete behavioral segment representations into continuous behavioral representations containing temporal continuity information.

[0075] The time interval embedding channel refers to an independent information representation channel that represents the time interval information between adjacent time segments. It reflects the time interval difference between adjacent time segments in the time dimension and does not directly participate in the extraction of behavioral features themselves. The time interval information participates in time series modeling in a form independent of behavioral features.

[0076] The introduced time interval embedding channel encodes the time interval information between adjacent time segments, specifically as follows:

[0077] Based on the time interval embedding channel, the time identifier information corresponding to adjacent time segments is obtained, and the time interval between adjacent time segments is calculated. The time interval is numerically normalized, and the normalized time interval is feature-extended. The single time interval value is converted into a time interval feature representation containing multi-dimensional time attribute information, and the time interval feature representation is dimensionally aligned.

[0078] The formation of the contextual behavior representation containing time interval features specifically includes:

[0079] The continuous behavior representation and the corresponding time interval feature representation are jointly represented and processed, and the evolution relationship of behavior over time and the time interval difference between adjacent time segments are fused together to form a contextual behavior representation that includes time interval features.

[0080] The future behavior representation prediction unit predicts the behavior representation corresponding to the next time segment based on different semantic level representations. It sets up semantic level sub-projection paths to perform hierarchical mapping of contextual behavior representations, obtaining predicted behavior representations that respectively represent local behavior change characteristics and group behavior evolution trend characteristics. Among these, the following is a summary:

[0081] The setting of the semantic hierarchical sub-projection path is specifically as follows:

[0082] To address the different semantic levels of information contained in the contextual behavior representation, the contextual behavior representation is divided into two semantic levels according to the difference in the semantic granularity of the behavior. Corresponding feature projection structures are set for each semantic level. One semantic level describes the local behavioral change features of the breeding object in a short time scale, while the other semantic level describes the overall evolution trend of the group behavior in a longer time scale.

[0083] The setting of semantic hierarchical projection paths for hierarchical mapping of contextual behavior representations is specifically as follows:

[0084] The same contextual behavior representation is input into feature projection structures corresponding to different semantic levels. In the semantic level representing local behavior change features, the local feature representation is formed by retaining some information in the contextual behavior representation that reflects short-term behavior fluctuations, local action changes and immediate response features. In the semantic level representing the evolutionary trend of group behavior, the evolutionary feature representation is formed by extracting some information in the contextual behavior representation that reflects the consistency, coordinated changes and overall evolutionary direction of behavior across multiple time segments. The same contextual behavior representation forms feature representations with different emphases at different semantic levels.

[0085] The predicted behavioral representations obtained, which respectively characterize local behavioral change features and group behavioral evolution trend features, are specifically as follows:

[0086] For the semantic level that represents the features of local behavioral changes, the future behavior representation prediction unit uses the local feature representation as the representation basis of the current short time scale behavior state, calculates the continuous changes of the behavior state in the time dimension, and generates the local behavior change prediction behavior representation for the next time segment.

[0087] For the semantic level that represents the evolutionary trend of group behavior, the future behavior representation prediction unit takes the evolutionary feature representation as the representation basis of the current overall behavior state of the group, calculates the evolutionary direction of the behavior state across time segments, and generates the evolutionary trend prediction behavior representation for the next time segment.

[0088] The predictive behavioral representation is formed by combining the predictive behavioral representation of local behavioral changes and the predictive behavioral representation of evolutionary trends.

[0089] The contrastive constraint representation learning unit pairs the predicted behavior representation with the corresponding actual behavior segment representation for the next time segment, calculates the attractive error term between similar pairs and the repulsive error term between dissimilar pairs, and constructs a contrastive loss function based on the difference between the attractive and repulsive error terms, where:

[0090] Similarity pairs refer to representation pairs that have temporal consistency and behavioral semantic consistency with the corresponding actual time segment representations of behavioral segments;

[0091] Dissimilar pairs refer to representation pairs formed between predicted behavioral representations and behavioral segment representations that do not correspond to specific time segments.

[0092] The calculation of the attraction error term between the similar pairs is as follows:

[0093] When calculating the attraction error term between similar pairs, the difference between the predicted behavior representation and the corresponding actual behavior segment representation is measured, and the representation difference is used as the basis for the value of the attraction error term. The smaller the representation difference, the smaller the value of the corresponding attraction error term.

[0094] The calculation of the repulsion error term between the dissimilar pairs is as follows:

[0095] When calculating the exclusion error term between dissimilar pairs, the representation difference between the predicted behavior representation and the dissimilar behavior segment representation is measured, and the representation difference is used as the basis for the value of the exclusion error term. The larger the representation difference, the smaller the value of the corresponding exclusion error term; the smaller the representation difference, the larger the value of the corresponding exclusion error term.

[0096] The comparison loss function constructed based on the difference between the attraction error term and the repulsion error term is as follows:

[0097] Within the same update step, the attraction error term is calculated for each pair of representations in the similar pair set, and the attraction error terms of all similar pairs are summarized to obtain the total attraction error. The repulsion error term is calculated for each pair of representations in the dissimilar pair set, and the repulsion error terms of all dissimilar pairs are summarized to obtain the total repulsion error. When constructing the loss, the total attraction error is used as the aggregation term to be minimized, and the total repulsion error is used in the loss calculation in the opposite direction. The optimization process promotes the continuous reduction of the representation difference of similar pairs on the one hand, and the continuous increase of the representation difference of dissimilar pairs on the other hand, forming a contrastive loss function that simultaneously has similar compactness and dissimilar separation constraints.

[0098] Based on the contrastive loss function, the behavioral representation is updated, optimizing the representation mapping space to maintain compactness among similar behaviors and separation among dissimilar behaviors, and outputting a sequence of potential behavioral representations corresponding one-to-one with each time segment. Specifically, the updating of the behavioral representation involves:

[0099] The predicted behavior representation and the actual behavior segment representation corresponding to the current time segment are used as update objects. Based on the feedback results of the contrastive loss function, the representation position of the behavior representation in the feature space is adjusted. When the similarity between the predicted behavior representation and the behavior segment representation of the corresponding actual time segment is insufficient, the predicted behavior representation moves towards the corresponding actual behavior representation by reducing the representation difference between the two. When the discriminability between the predicted behavior representation and the dissimilar behavior segment representation is insufficient, the predicted behavior representation moves away from the dissimilar behavior representation by increasing the representation difference between the two.

[0100] Update operations are repeatedly performed on the behavioral representations corresponding to all time segments. During the update process, the behavioral representations gradually form a distribution structure that clusters among similar behaviors and separates among dissimilar behaviors. The behavioral representations corresponding to each time segment maintain a stable and distinguishable representation in the feature space, and finally output a sequence of potential behavioral representations that corresponds one-to-one with each time segment.

[0101] In this embodiment, the step of constructing the coupling relationship between each behavioral phase variable based on behavioral similarity includes:

[0102] Based on behavioral observation data, individual identifiers of each aquaculture object in the feeding area are obtained, and the individual identifiers are bound to the correspondence between each time segment in the potential behavioral representation sequence;

[0103] The potential behavioral representation sequences are aggregated according to individual identifiers. The potential behavioral representations of the same breeding object in consecutive time segments are organized in chronological order into behavioral unit representation sequences corresponding to the breeding object, forming a set of behavioral unit representation sequences corresponding to multiple breeding objects respectively.

[0104] Phase mapping is performed on each behavioral unit representation sequence in the behavioral unit representation sequence set, mapping the behavioral unit representation corresponding to each time segment to the corresponding behavioral phase variable, thus obtaining the behavioral phase variable sequence for each cultured object in each time segment, where:

[0105] The phase mapping process performed on each behavioral unit representation sequence in the behavioral unit representation sequence set is specifically as follows:

[0106] For the behavioral unit representation sequence corresponding to the breeding object, the behavioral unit representation corresponding to each time segment is read in chronological order. The position mapping processing of each behavioral unit representation is performed in a unified behavioral representation space, and the behavioral unit representation is mapped to the phase value representing the position in the behavioral state change cycle.

[0107] The specific steps for obtaining the behavioral phase variable sequences of each cultured object at each time segment are as follows:

[0108] The phase values ​​corresponding to each time segment of the behavioral unit representation sequence are arranged in chronological order to form the corresponding behavioral phase variable sequence. Phase mapping and time arrangement processing are performed on each behavioral unit representation sequence in the behavioral unit representation sequence set to obtain the behavioral phase variable sequence corresponding to each breeding object in each time segment.

[0109] The behavioral similarity relationship between each pair of aquaculture objects is calculated based on the set of behavioral unit representation sequences, and the coupling relationship between each behavioral phase variable is determined based on the behavioral similarity relationship, wherein:

[0110] The calculation of the behavioral similarity between each pair of the farmed animals is as follows:

[0111] For any two aquaculture objects, the corresponding behavioral unit representation sequences are obtained respectively. The two behavioral unit representation sequences are aligned at the same time segment position. The similarity between the behavioral unit representations corresponding to each time segment after alignment is compared segment by segment to obtain the local behavioral similarity. The local behavioral similarities on multiple time segments are combined to obtain the similarity that reflects the overall behavioral similarity between the two aquaculture objects.

[0112] The determination of the coupling relationship between each behavioral phase variable based on behavioral similarity is specifically as follows:

[0113] When the behavioral similarity between two farmed objects is higher than the similarity threshold, a coupling relationship stronger than the coupling relationship threshold is established between the corresponding behavioral phase variables. When the behavioral similarity between two farmed objects is lower than the similarity threshold, no coupling relationship is established between the corresponding behavioral phase variables.

[0114] In this embodiment, obtaining the synchronization index and the synchronization evolution sequence includes:

[0115] The behavioral phase variables and their corresponding coupling relationships are input into the improved Kuramoto model. Each cultured object within the feeding area is constructed as an oscillator, forming an oscillator set. A unique correspondence is established for each oscillator, with each oscillator corresponding one-to-one with a behavioral unit representation sequence. Specifically, constructing each cultured object within the feeding area as an oscillator involves:

[0116] Each identifiable aquaculture object is assigned a unique object identifier within the feeding area, and the corresponding behavioral unit representation sequence is obtained based on the object identifier. The behavioral unit representation sequence is used as a description of the behavioral evolution of the aquaculture object within the observation time range. The aquaculture object is abstractly represented as an independent oscillator individual, and a unique association is established between each oscillator and the corresponding aquaculture object. The change of the oscillator state over time directly corresponds to the change of the behavioral state of the aquaculture object at different times. The phase state of the oscillator is represented by the behavioral phase variable under the corresponding time segment. The evolutionary characteristics of the oscillator are driven by the behavioral evolutionary characteristics of the aquaculture object.

[0117] Based on the sequence of behavioral phase variables, each oscillator is assigned a corresponding behavioral phase variable, and the behavioral phase variable is used as the state representation of the oscillator at the current time step. Specifically, assigning a corresponding behavioral phase variable to each oscillator involves:

[0118] Based on the one-to-one correspondence established between the oscillator and the aquaculture object, the phase values ​​in the behavioral phase variable sequence are mapped to the corresponding oscillator. For each oscillator, the object identifier corresponding to the oscillator is first determined, and the behavioral phase variable sequence corresponding to the aquaculture object is obtained based on the object identifier. At the current time step, the behavioral phase variable corresponding to the time step is read from the behavioral phase variable sequence, and the behavioral phase variable is assigned to the corresponding oscillator as the phase state representation of the oscillator at the current time step.

[0119] Based on the behavioral potential representation sequence and environmental correlation data, a corresponding dynamic evolution frequency is determined for each oscillator, and this frequency serves as the basic evolutionary input for the oscillator under conditions unaffected by the remaining oscillators. Specifically, determining the corresponding dynamic evolution frequency for each oscillator involves:

[0120] Based on the behavioral evolution state of the corresponding aquaculture object and its environmental conditions, the natural evolution rhythm of the oscillator at the current time step is characterized. For each oscillator, based on the corresponding potential behavioral representation sequence, information reflecting the activity level, intensity of behavioral changes, and behavioral evolution trend of the aquaculture object at the current time step is obtained. Combined with the environmental correlation data corresponding to the aquaculture object, the external conditions affecting the behavioral rhythm of the aquaculture object are comprehensively characterized. The behavioral state information and environmental state information form a joint description at the same time scale. Based on the comprehensive representation of the current behavioral state and environmental state, the dynamic evolution frequency of the natural evolution speed of the oscillator at the current time step is determined. Specifically, the determination of the dynamic evolution frequency of the natural evolution speed of the oscillator at the current time step is as follows:

[0121] The representation quantities directly related to the speed of the behavior rhythm are extracted from the joint description of the behavior state and the environment state, and the behavior rhythm component and the environment rhythm component are formed respectively. The scale of the two components is unified. The behavior rhythm component is used as the basic term and the environment rhythm component is used as the adjustment term. The two are jointly synthesized to obtain the natural evolution speed representation of the oscillator at the current time step, and the natural evolution speed representation is mapped to the corresponding dynamic evolution frequency.

[0122] Based on the behavioral phase variables and their corresponding coupling relationships, the corresponding coupling strength is determined for any two oscillators. A phase coupling function term is constructed according to the behavioral phase difference between the two oscillators, and the coupling function term is combined with the corresponding coupling strength to form the coupling input, where:

[0123] The determination of the corresponding coupling strength for any two oscillators is specifically as follows:

[0124] Based on the set of behavioral unit representation sequences, the overall behavioral similarity quantitative representation of two corresponding breeding objects is obtained, reflecting the degree of similarity between the two in terms of behavioral patterns, behavioral change rhythm and behavioral evolution trend. The behavioral similarity quantitative representation is mapped to the corresponding coupling strength value. The oscillators corresponding to breeding objects with behavioral similarity higher than the threshold have a coupling strength greater than the strength threshold, and the oscillators corresponding to breeding objects with behavioral similarity lower than the threshold have a coupling strength less than the strength threshold.

[0125] The construction of the phase coupling function term based on the behavioral phase difference between the two oscillators is specifically as follows:

[0126] In each time update step, the behavior phase variables of any two oscillators at the current time step are obtained, and the relative difference between them in the phase space is calculated to obtain the phase difference. Based on the magnitude and direction of the phase difference, the interaction trend between the oscillators is characterized. When the phase difference is less than a threshold, it is considered that the behavior rhythms of the two oscillators are similar, and the interaction promotes synchronization in phase update. When the phase difference is greater than the threshold, it is considered that the behavior rhythms of the two oscillators deviate significantly, and the interaction inhibits synchronization in phase update. Based on the correspondence between the phase difference and the interaction trend, the phase difference is mapped to a coupling influence quantity representing the strength and direction of the interaction, and the coupling influence quantity is used as the phase coupling function term.

[0127] During each time update process, a synchronous evolution update process is performed on the behavioral phase variable of each oscillator. The dynamic evolution frequency of the oscillator is used as the basic evolution component, the coupling input is used as the interaction component, and the behavioral phase variable is updated by combining the historical phase state and historical synchronization state of the oscillator to obtain the updated behavioral phase variable. The synchronous evolution update process for the behavioral phase variable of each oscillator is specifically as follows:

[0128] During each time update process, the dynamic evolution frequency corresponding to the oscillator is used as the basic evolution component of the oscillator at the current time step to describe the natural trend of the oscillator behavior phase change over time without considering the group influence. The coupling input formed by the action of the remaining oscillators is introduced as the interaction component to characterize the direction and degree of influence of the remaining oscillators in the group on the change of the current oscillator behavior phase.

[0129] When performing phase updates, the historical behavior phase state of the oscillator in the previous time update step and the historical synchronization state of the group in the previous time update step are further combined to adjust the current phase change. When the historical behavior phase state is consistent with the phase change trend in the current time step, the behavior evolution of the oscillator is considered to have strong temporal continuity, and the current phase change amplitude is moderately enhanced during phase updates. When the historical behavior phase state deviates significantly from the current phase change trend, the behavior evolution of the oscillator is considered to have a sudden risk of mutation, and the current phase change amplitude is suppressed during phase updates.

[0130] The phase update is adjusted as a whole by combining the historical synchronization state of the group. When the group is in a synchronization state above the threshold level in the previous time update step, the individual phase change is smoothed. When the synchronization state of the group is below the threshold level, the individual phase change constraint is relaxed, and the oscillator responds more flexibly to its own behavioral state changes.

[0131] Based on the updated behavioral phase variables of all oscillators, the overall synchronization index of the group at the current moment is calculated, and a synchronization evolution sequence is formed during the continuous time update process. The changing trend and fluctuation of the synchronization evolution sequence are statistically processed to obtain the synchronization evolution sequence, where:

[0132] The calculation of the overall synchronization index of the group at the current moment is specifically as follows:

[0133] In the current time update step, the updated behavior phase variables of all oscillators in the feeding area are collected. By comparing the concentration of each oscillator behavior phase variable in the phase space, it is determined whether the oscillator phases in the group tend to be in the same direction and whether they show obvious phase clustering. When the behavior phase variables of more than a threshold number of oscillators are relatively concentrated in the phase space, a synchronization index higher than the index threshold is obtained. When the oscillator behavior phase variables are relatively dispersed, a synchronization index lower than the threshold is obtained.

[0134] The obtained synchronization evolution sequence is specifically as follows:

[0135] In multiple consecutive time update steps, the calculation of the synchronization index is repeated, and the synchronization index corresponding to each time update step is arranged in chronological order to form a synchronization evolution sequence that reflects the change of the group's synchronization state over time.

[0136] In this embodiment, generating the feeding trigger signal includes:

[0137] Obtain the synchronization index and synchronization evolution sequence, and arrange the synchronization evolution sequence in chronological order to form the basic judgment sequence;

[0138] A continuous decision time window is constructed based on the basic decision sequence. Within each continuous decision time window, a synchronization index for the corresponding time period is extracted. Based on the synchronization index, the synchronization level, synchronization trend, and synchronization fluctuation degree within the time window are calculated, forming a synchronization decision feature set corresponding to the time window. Wherein:

[0139] The construction of a continuous decision time window based on the basic decision sequence is specifically as follows:

[0140] Using the synchronization evolution sequence as the basic decision sequence, the synchronization index is segmented and organized according to the chronological order. A starting time position is selected in the basic decision sequence, and starting from the starting time position, multiple synchronization indices corresponding to adjacent time update steps are selected to form a continuous decision time window. The decision time window is slid forward along the time axis. While keeping the window length unchanged, multiple continuous decision time windows that are connected in time are formed in sequence.

[0141] The calculation of the synchronization level, synchronization trend, and synchronization fluctuation degree within the time window based on the synchronization index is as follows:

[0142] In the process of calculating the synchronization level, the values ​​of the synchronization index corresponding to all time update steps within the continuous decision time window are summed, and the summation result is divided by the number of time update steps within the continuous time window to obtain the average value of the synchronization index within the decision time window. The average value is used as the synchronization level.

[0143] In the process of calculating the synchronous change trend, the change in the synchronization index between two adjacent time update steps within the continuous judgment time window is calculated sequentially according to the time order, and all changes within the time window are accumulated. When the accumulated result is positive, it indicates that the synchronization degree is generally on an upward trend within the time window. When the accumulated result is close to zero, it indicates that the synchronization degree is generally stable within the time window. When the accumulated result is negative, it indicates that the synchronization degree is generally on a downward trend within the time window, thus obtaining the synchronous change trend.

[0144] In the process of calculating the degree of synchronization fluctuation, the maximum and minimum values ​​of the synchronization index within the continuous judgment time window are obtained, and the difference between the two is calculated. The difference is used as the degree of synchronization fluctuation within the time window, which characterizes the stability of the synchronization state within the time period. When the difference is less than the threshold, it indicates that the synchronization degree changes slowly within the time window and the synchronization state is relatively stable. When the difference is greater than the threshold, it indicates that the synchronization degree fluctuates significantly within the time window and the synchronization state is less stable.

[0145] Pre-defined synchronization maintenance conditions include synchronization level threshold conditions, synchronization change trend conditions, and synchronization fluctuation constraint conditions, wherein:

[0146] The synchronization level threshold condition is constrained within a continuous judgment time window. The overall synchronization degree reflected by the group synchronization index must reach the set synchronization level requirement. When the synchronization index is generally higher than the synchronization level threshold within the time window, it indicates that the behavior phase of most aquaculture objects is in a relatively consistent state, and the synchronization level threshold condition is considered to be met.

[0147] The synchronous change trend condition is constrained within a continuous judgment time window. The directionality of the synchronous index changes over time. When the synchronous index shows a continuous upward and stable trend within the time window, it indicates that the group's synchronous state is being formed, has been formed and is maintained, and is considered to meet the synchronous change trend condition. When the synchronous index shows a significant downward trend within the time window, it indicates that the group's synchronous state is weakening or being destroyed, and is considered not to meet the synchronous change trend condition.

[0148] The synchronization fluctuation constraint condition is imposed on the fluctuation range of the synchronization degree index around the overall level within a continuous judgment time window. When the fluctuation range of the synchronization degree index within the time window is less than the fluctuation threshold, it indicates that the synchronization state of the group is relatively stable and not easily affected by short-term disturbances, and it is considered to meet the synchronization fluctuation constraint condition. When the synchronization degree index fluctuates drastically within the time window, it indicates that the synchronization state of the group is unstable and easily affected by instantaneous behavior, and it is considered not to meet the synchronization fluctuation constraint condition.

[0149] The synchronization determination feature set is compared with the synchronization maintenance condition window by window. Based on the window-by-window comparison results, the group feeding behavior execution mode is determined. When the current continuous determination time window meets the synchronization maintenance condition, a corresponding feeding trigger candidate signal is generated. The determination of the group feeding behavior execution mode based on the window-by-window comparison results is specifically as follows:

[0150] Within each continuous judgment time window, the synchronization judgment feature set corresponding to the time window is obtained, including synchronization level, synchronization change trend and synchronization fluctuation degree. The synchronization judgment feature set is compared with the set synchronization level threshold condition, synchronization change trend condition and synchronization fluctuation constraint condition to determine whether the group synchronization state within the time window simultaneously meets the synchronization maintenance conditions.

[0151] When the set of synchronous judgment features within a certain continuous judgment time window simultaneously meets the synchronization maintenance condition, it is considered that the group behavior within the time window has formed a synchronous feature with continuity and stability. The group behavior corresponding to the time window is judged as feeding synchronous behavior, and a corresponding feeding trigger candidate signal is generated.

[0152] Perform continuous confirmation processing on the candidate feed trigger signal. When the synchronization maintenance condition is met in the continuous decision time window corresponding to multiple consecutive adjacent time update steps, output the feed trigger signal.

[0153] In this embodiment, generating the set of feeding control parameters corresponding to the current group feeding pattern includes:

[0154] Obtain the group feeding behavior pattern determination results, synchronization index and synchronization evolution sequence corresponding to the feeding trigger signal, and retrieve the current time data of visual behavior data, group movement characteristic data, spatial distribution characteristic data and environmental correlation data from the behavior observation data;

[0155] Based on the group feeding behavior pattern determination results and the spatial distribution characteristic data at the current moment, a set of target feeding sub-regions within the feeding area is determined, and the feeding priority order of each target feeding sub-region is determined based on the set of target feeding sub-regions. Specifically, determining the set of target feeding sub-regions within the feeding area involves:

[0156] Based on the spatial distribution characteristics data at the current moment, the feeding area is spatially divided into multiple distinguishable sub-regions. The distribution density, aggregation degree and spatial continuity characteristics of the farmed objects in each sub-region are obtained. Combined with the group feeding behavior pattern judgment results, sub-regions corresponding to the current synchronous feeding behavior spatial location are selected. The focus is on the spatial sub-regions where the farmed objects are spatially concentrated and match the synchronous feeding behavior of the group.

[0157] Feeding intensity parameters are determined based on synchronization indices and synchronization evolution sequences, feeding rhythm parameters are determined based on group movement characteristic data, and feeding compensation parameters are determined based on environmental correlation data, forming a set of feeding control parameters. This set of feeding control parameters includes feeding intensity parameters, feeding rhythm parameters, feeding compensation parameters, and regional feeding allocation parameters corresponding to the target feeding sub-region set, wherein:

[0158] The method for determining the feeding intensity parameters based on the synchronization index and the synchronization evolution sequence is as follows:

[0159] Obtain the synchronization index at the current moment, and combine it with the synchronization evolution sequence to determine the degree of formation and stability of the group synchronization state in the time dimension. When the synchronization index is higher than the index threshold and the synchronization evolution sequence shows a trend of continuous maintenance and enhancement, determine the feeding intensity parameter that is higher than the parameter threshold. When the synchronization index is lower than the index threshold and the synchronization state is still in the formation stage, determine the feeding intensity parameter that is lower than the parameter threshold.

[0160] The determination of feeding rhythm parameters based on group movement characteristic data specifically includes:

[0161] Obtain the group movement characteristic data at the current moment, and determine the execution rhythm of the feeding operation in the time dimension based on the movement rhythm characteristics reflected in the group movement characteristic data. When the group movement shows an activity level and consistency higher than the activity threshold, it is determined to be a dense feeding rhythm; otherwise, a relatively gentle feeding rhythm is determined.

[0162] The determination of feeding compensation parameters based on environmental correlation data specifically includes:

[0163] Obtain environmental data at the current moment, and adjust feeding parameters to compensate for the potential impact of environmental conditions on feeding efficiency and feed utilization. When environmental conditions have an adverse effect on feeding, determine the corresponding feeding compensation parameters to compensate for the feeding amount. When environmental conditions are relatively ideal, no additional compensation is introduced.

[0164] Regional feeding allocation parameters refer to control parameters that describe the proportional relationship of feeding resources allocated among multiple target feeding sub-regions. They represent the proportion of feeding amount, feeding time, or feeding order weight of different target feeding sub-regions in the same feeding operation.

[0165] The set of feeding control parameters is converted into a sequence of feeding control instructions that can be executed by the feeding robot. The sequence of feeding control instructions includes an arrival instruction for the target feeding sub-region, a feeding path instruction, a feeding discharge instruction, and a feeding rhythm control instruction.

[0166] The feeding robot is controlled to perform feeding operations according to the sequence of feeding control commands.

[0167] Example 1:

[0168] To verify the feasibility of this invention in practice, it was applied to an intensive aquaculture pond in a coastal area. This pond uses a feeding robot to automatically feed and manage the cultured organisms. The pond area is approximately 1200 square meters, with about 8500 organisms per pond. The feeding area is significantly affected by changes in light, water disturbance, and human activity. Group behavior is prone to short-term aggregation or activity even outside of feeding periods. Traditional feeding methods based on timed feeding or simple behavioral thresholds are prone to accidental triggering of feeding under non-realistic feeding conditions, resulting in feed waste and water quality fluctuations.

[0169] In this scenario, the feeding robot continuously collects visual behavior data, group movement characteristic data, spatial distribution characteristic data, and environmental correlation data of the farmed animals through deployed visual acquisition devices and environmental perception devices. After preprocessing the collected data, it forms an effective behavior sequence. This invention's method performs temporal representation modeling of group behavior within continuous time segments, obtaining potential behavioral representations reflecting behavioral evolution trends. Furthermore, it maps individual behaviors to behavioral phase variables, constructs coupling relationships between group behaviors, and continuously analyzes the synchronization state of group behavior. When the synchronization degree of group behavior remains stable and shows an increasing trend within a continuous time window, this state is determined to be genuine feeding behavior, and a feeding trigger signal is generated. When only short-term behavioral activity occurs and the synchronization state is unstable, no feeding operation is triggered.

[0170] During 30 consecutive days of operation, a total of 126 valid feeding trigger events were recorded. Comparing data from the same aquaculture pond before and after implementing the method of this invention, the results showed that before implementation, there were an average of about 1.9 false feeding triggers per day, while after implementation, the number of false triggers decreased to an average of 0.4 per day. The average daily feed intake per pond decreased from about 145 kg to 132 kg, and the feed utilization rate increased by about 8.6%. The average fluctuation range of turbidity in the aquaculture pond water within 2 hours after feeding decreased from 22% to 13%, and the duration of group feeding became more concentrated and stable.

[0171] Table 1. Statistical comparison of feeding effects and group behavior stability before and after application of the method of the present invention.

[0172] Statistical indicators Before applying the method of this invention After applying the method of the present invention Improvement Status Description Statistical period 30 consecutive days 30 consecutive days Same breeding pond, same breeding batch Average number of feeding triggers per day 6.8 5.2 Fewer accidental triggers, more focused feeding Average number of times feeding is accidentally triggered per day (times / day) 1.9 0.4 False triggers have decreased significantly. Percentage of false triggers (%) 27.9% 7.7% Improved accuracy in determining group feeding Average daily feed intake per pond (kg / day) 145 132 Feed distribution is more in line with actual needs Feed utilization rate (%) 78.3% 86.9% Utilization efficiency improved by approximately 8.6 percentage points. Average duration of a single feeding (minutes) 18.5 14.2 The feeding process is more concentrated and efficient Duration of group feeding synchronization (minutes) 9.6 15.4 Synchronized feeding behavior is more stable Fluctuation in water turbidity (%) 2 hours after feeding 22% 13% Feed residues were significantly reduced. Frequency of abnormal behavior after feeding (times / week) 6.3 2.1 Increased stability of group behavior Accuracy rate (%) in determining group feeding behavior 82.4% 94.1% Pattern recognition performance has been significantly improved.

[0173] The data comparison in Table 1 shows that, under the same rearing pond, the same rearing batch, and continuous 30-day operation conditions, the feeding effect and the stability of group behavior were significantly improved before and after applying the method of this invention. Regarding feeding trigger behavior, after applying the method of this invention, the average number of feeding triggers per day decreased from 6.8 to 5.2, indicating that the method of this invention can effectively reduce unnecessary feeding operations.

[0174] Regarding feed input and utilization, the average daily feed input per pond decreased from 145 kg to 132 kg. While the feed input decreased, the feed utilization rate increased from 78.3% to 86.9%, an increase of approximately 8.6 percentage points. This indicates that the present invention, through more precise control of feeding timing and parameters, makes feed input more aligned with the actual feeding needs of the population, reducing ineffective feeding and feed waste.

[0175] From the perspective of environmental and group behavior stability, the fluctuation range of water turbidity decreased from 22% to 13% within 2 hours after feeding, indicating a significant reduction in feed residue and disorderly feeding. The frequency of abnormal behavior after feeding decreased from 6.3 times per week to 2.1 times per week, and the stability of group behavior was significantly enhanced.

[0176] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A precision feeding method of a feeding robot based on pattern recognition, characterized by, include: Behavioral observation data of farmed animals in the feeding area of ​​a group farming scenario are collected, and the behavioral observation data are preprocessed to obtain effective behavioral sequences; A behavior time series representation module based on contrastive predictive coding is constructed to model the temporal prediction relationship between adjacent time segments in the effective behavior sequence, and to obtain a behavior potential representation sequence that corresponds one-to-one with each time segment. The potential behavioral representation sequences are organized into multiple behavioral unit representation sequences according to individual identifiers, and each behavioral unit representation sequence is mapped to a corresponding behavioral phase variable. At the same time, the coupling relationship between each behavioral phase variable is constructed based on the behavioral similarity relationship. By inputting the behavioral phase variables and their corresponding coupling relationships into the improved Kuramoto model, population synchronization evolution calculations are performed to obtain synchronization indexes and synchronization evolution sequences. The feeding behavior of the group is patterned based on the synchronization index and the synchronization evolution sequence. When the synchronization index reaches the synchronization threshold and the synchronization maintenance condition is met within the continuous judgment time window, a feeding trigger signal is generated. Upon receiving a feeding trigger signal, a set of feeding control parameters corresponding to the current group feeding mode is generated, and the feeding robot is controlled to perform feeding operations according to the set of feeding control parameters.

2. The method according to claim 1, wherein, The behavioral observation data specifically includes visual behavioral data, group movement characteristic data, spatial distribution characteristic data, and environmental correlation data.

3. The method according to claim 1, wherein, The preprocessing of behavioral observation data specifically includes noise suppression, anomaly processing, time alignment, spatial region constraints, serialization, and standardization.

4. The method according to claim 1, wherein, The process of obtaining the behavioral latent representation sequence corresponding one-to-one with each time segment includes: A behavioral temporal representation module based on contrastive predictive coding is constructed, which includes a temporal segment coding unit, a contextual temporal modeling unit, a future representation prediction unit, and a contrastive constraint representation learning unit. The temporal segment coding unit divides the effective behavior sequence into time segments and performs coding processing on the behavior observation data corresponding to each time segment to obtain a behavior segment representation that corresponds one-to-one with each time segment. The contextual temporal modeling unit performs temporal correlation modeling on the behavioral segment representations corresponding to multiple adjacent time segments in chronological order. At the same time, it introduces a time interval embedding channel to encode the time interval information between adjacent time segments, and then jointly represents the encoded time interval information with the temporal correlation results to form a contextual behavioral representation containing time interval features. The future behavior representation prediction unit predicts the behavior representation corresponding to the next time segment based on different semantic level representations. It sets up semantic level sub-projection paths to perform hierarchical mapping of context behavior representations, and obtains predicted behavior representations that respectively represent local behavior change features and group behavior evolution trend features. The contrastive constraint representation learning unit pairs the predicted behavior representation with the corresponding actual behavior segment representation of the next time segment, calculates the attractive error term between similar pairs and the repulsive error term between dissimilar pairs, and constructs a contrastive loss function based on the difference between the attractive and repulsive error terms. Based on the contrastive loss function, the behavioral representation is updated, and the representation mapping space is optimized to keep the behavioral representations compact between similar behaviors and separate between dissimilar behaviors, and outputs a sequence of potential behavioral representations that corresponds one-to-one with each time segment.

5. The method according to claim 1, wherein, The construction of coupling relationships between behavioral phase variables based on behavioral similarity includes: Based on behavioral observation data, individual identifiers of each aquaculture object in the feeding area are obtained, and the individual identifiers are bound to the correspondence between each time segment in the potential behavioral representation sequence; The potential behavioral representation sequences are aggregated according to individual identifiers. The potential behavioral representations of the same breeding object in consecutive time segments are organized in chronological order into behavioral unit representation sequences corresponding to the breeding object, forming a set of behavioral unit representation sequences corresponding to multiple breeding objects respectively. Phase mapping processing is performed on each behavioral unit representation sequence in the behavioral unit representation sequence set to map the behavioral unit representation corresponding to each time segment to the corresponding behavioral phase variable, thereby obtaining the behavioral phase variable sequence corresponding to each breeding object in each time segment. The behavioral similarity relationship between each pair of aquaculture objects is calculated based on the behavioral unit representation sequence set, and the coupling relationship between each behavioral phase variable is determined based on the behavioral similarity relationship.

6. The method for precise feeding of a feeding robot based on pattern recognition according to claim 1, characterized in that, The process of obtaining the synchronization index and the synchronization evolution sequence includes: The behavioral phase variables and their corresponding coupling relationships are input into the improved Kuramoto model. Each aquaculture object in the feeding area is constructed as an oscillator, forming an oscillator set. A unique correspondence is established for each oscillator, and each oscillator corresponds one-to-one with a behavioral unit representation sequence. Based on the sequence of behavioral phase variables, each oscillator is assigned a corresponding behavioral phase variable, and the behavioral phase variable is used as the state representation of the oscillator at the current time step. Based on the behavioral potential characterization sequence and environmental correlation data, a corresponding dynamic evolution frequency is determined for each oscillator, and it serves as the basic evolution input for the oscillator under the condition that it is not affected by the remaining oscillators. Based on the behavioral phase variables and their corresponding coupling relationships, the corresponding coupling strength is determined for any two oscillators. A phase coupling function term is constructed based on the behavioral phase difference between the two oscillators, and the coupling function term is combined with the corresponding coupling strength to form the coupling input. During each time update process, a synchronous evolution update process is performed on the behavior phase variable of each oscillator. The dynamic evolution frequency of the oscillator is used as the basic evolution component, the coupling input is used as the interaction component, and the behavior phase variable is updated by combining the historical phase state and historical synchronization state of the oscillator to obtain the updated behavior phase variable. Based on the updated behavioral phase variables of all oscillators, the overall synchronization index of the group at the current moment is calculated, and a synchronization evolution sequence is formed during the continuous time update process. The changing trend and fluctuation of the synchronization evolution sequence are statistically processed to obtain the synchronization evolution sequence.

7. The method for precise feeding of a feeding robot based on pattern recognition according to claim 1, characterized in that, The generation of the feeding trigger signal includes: Obtain the synchronization index and synchronization evolution sequence, and arrange the synchronization evolution sequence in chronological order to form the basic judgment sequence; A continuous decision time window is constructed based on the basic decision sequence. Synchronization indexes for the corresponding time period are extracted within each continuous decision time window. Based on the synchronization indexes, the synchronization level, synchronization change trend, and synchronization fluctuation degree within the time window are calculated to form a set of synchronization decision features corresponding to the time window. Pre-set synchronization maintenance conditions, including synchronization level threshold conditions, synchronization change trend conditions, and synchronization fluctuation constraint conditions; The synchronization judgment feature set is compared with the synchronization maintenance condition window by window. The group feeding behavior execution mode is judged based on the window by window comparison result. When the current continuous judgment time window meets the synchronization maintenance condition, the corresponding feeding trigger candidate signal is generated. Perform continuous confirmation processing on the candidate feed trigger signal. When the synchronization maintenance condition is met in the continuous decision time window corresponding to multiple consecutive adjacent time update steps, output the feed trigger signal.

8. The method for precise feeding of a feeding robot based on pattern recognition according to claim 1, characterized in that, The generation of the feeding control parameter set corresponding to the current group feeding pattern includes: Obtain the group feeding behavior pattern determination results, synchronization index and synchronization evolution sequence corresponding to the feeding trigger signal, and retrieve the current time data of visual behavior data, group movement characteristic data, spatial distribution characteristic data and environmental correlation data from the behavior observation data; Based on the group feeding behavior pattern determination results and the spatial distribution characteristic data at the current moment, the set of target feeding sub-regions within the feeding area is determined, and the feeding priority order of each target feeding sub-region is determined based on the set of target feeding sub-regions; Feeding intensity parameters are determined based on synchronization index and synchronization evolution sequence, feeding rhythm parameters are determined based on group movement characteristic data, and feeding compensation parameters are determined based on environmental correlation data, forming a set of feeding control parameters. The set of feeding control parameters includes feeding intensity parameters, feeding rhythm parameters, feeding compensation parameters, and regional feeding allocation parameters corresponding to the target feeding sub-region set. The set of feeding control parameters is converted into a sequence of feeding control instructions that can be executed by the feeding robot. The sequence of feeding control instructions includes an arrival instruction for the target feeding sub-region, a feeding path instruction, a feeding discharge instruction, and a feeding rhythm control instruction. The feeding robot is controlled to perform feeding operations according to the sequence of feeding control instructions.

Images