Group behavior regulation and control method and system based on multi-dog cooperative training

Abstract

The invention belongs to the technical field of dog training, and particularly relates to a group behavior regulation and control method and system based on multi-dog cooperative training, and the method comprises the following steps: collecting the spatial position and social behavior data of dogs through an intelligent necklace; on the basis, constructing a group behavior map fusing the physical space relationship and the social influence; dynamically identifying a real-time role of each dog by utilizing a graph theory centrality index; according to the role and the current behavior state, a differential control instruction is generated through a state-role-behavior multi-dimensional instruction mapping model; and finally, executing an instruction and monitoring response data so as to optimize the model and realize closed-loop adaptive adjustment. According to the method, through accurate role identification and differentiated instruction distribution, the problem of group behavior conflict caused by unified instructions is effectively solved, the instruction conflict rate is remarkably reduced, and the efficiency and adaptability of group cooperative training are continuously improved by using a closed-loop learning mechanism.

Description

Technical Field

[0001] This invention relates to the field of dog training technology, and in particular to a method and system for regulating group behavior based on multi-dog collaborative training. Background Technology

[0002] With the development of electronic dog training technology, various electronic dog trainers have emerged on the market, capable of "one-to-many" control, meaning a single transmitter can simultaneously send control commands to multiple dog collars. However, these existing technologies have significant limitations in practical applications. First, due to role differences within a dog pack, such as leader dogs, follower dogs, and independent dogs, they may react differently when receiving the same command, leading to group behavior confusion and even exacerbating conflicts. Second, existing technologies lack dynamic and personalized control capabilities, failing to dynamically adjust command distribution based on real-time social relationships and spatial locations among the dogs, significantly limiting training effectiveness. Furthermore, existing solutions are mostly based on individual dog behavior judgment, lacking in-depth analysis and structured behavioral modeling of group interaction patterns, resulting in insufficient group perception capabilities and difficulty in achieving efficient collaborative training.

[0003] Therefore, there is an urgent need in this field for a technical solution that can perceive the group status in real time, accurately identify individual roles, and allocate differentiated instructions as needed, so as to improve the efficiency, stability and adaptability of multi-dog collaborative training, while reducing the risk of behavioral conflicts. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for regulating group behavior based on multi-dog collaborative training. By constructing a two-layer dynamic graph that integrates physical space and social influence, and utilizing multi-dimensional graph centrality indicators to identify the role of each dog in real time, this method achieves precise, differentiated regulation that varies "from one-to-many" uniform commands to "differentiated by dog ​​and state." This approach fundamentally avoids command conflicts caused by differences in role behavior, thus reducing the command conflict rate.

[0005] The technical solution adopted by this invention to solve its technical problem is: to provide a group behavior regulation method based on multi-dog collaborative training, comprising the following steps:

[0006] S1. By wearing smart collars on several dogs i=1,2,...,N, collect spatial location data and social behavior data of each dog, and preprocess the collected data;

[0007] S2. Based on preprocessed spatial location data and social behavior data, construct a group behavior map G={G} representing the interaction relationships among dogs. phy G soc}, where G phy For the physical space relationship map, Gsoc For social influence mapping;

[0008] S3. Based on the topological structure and weight evolution of the group behavior graph, the multi-dimensional graph theory centrality index of each dog node in the graph is calculated and compared with a preset role definition threshold to dynamically identify the real-time role label C of each dog in the group. i (t);

[0009] S4. Monitor the current behavioral status of each dog. i (t), and based on its current role C i (t) and behavioral state S i (t), construct an initial "state-role-behavior" multidimensional instruction mapping model, generate a "state-role-behavior" multidimensional instruction mapping table, and query to obtain the corresponding control instruction D. i (t);

[0010] S5, Control command D i (t) The data is sent to the corresponding smart collar to implement control measures and continuously monitors the dog's behavioral response data R. i (t), the response data R i (t) is uploaded to the cloud server to optimize the initial "state-role-behavior" multi-dimensional instruction mapping model and adjust the intensity of subsequent instructions.

[0011] Furthermore, in step S2, a physical spatial relationship map G is constructed. phy The methods include:

[0012] Calculate the change in relative distance Δd between any two dogs i and j. ij Correlation coefficient r with speed ij ;

[0013] Calculate the weights W of physical spatial relationships ij And based on this, an adjacency matrix is ​​constructed as the graph G. phy The representation of:

[0014]

[0015] Where, Δd ij r represents the change in the relative distance between dogs i and j. ij The velocity correlation coefficient is α, and β are preset weighting coefficients used to balance the effects of distance and velocity, with α+β=1.

[0016] Furthermore, in step S2, a social influence graph G is constructed. soc The methods include:

[0017] Based on social behavior data, quantify and calculate at least one directed social relationship indicator I. k Directed social relationship indicator I k Includes: Imitation delay index M ij Interactive response rate metric R ij Attention Allocation Index A ij Intimacy Index C ij ;

[0018] By normalizing and weighting multiple indicators, the comprehensive social influence weight S from dog i to dog j is calculated. ij And based on this, construct the graph G soc :

[0019]

[0020] Among them, I k,norm Let ω be the k-th normalized social relationship indicator value. k The corresponding preset weight coefficient is K, where K is the total number of indicators used, k=1,2,3…K.

[0021] Furthermore, in step S3, by calculating the following graph theory centrality index and according to preset judgment rules, a corresponding dynamic role label C is assigned to each dog. i (t):

[0022] Out-degree centrality: OD(i) = ΣS ij ;

[0023] In-degree centrality: ID(i) = ΣS ji ;

[0024] Eigenvector centrality: EC(i) = (1 / λ)ΣS ji EC(j);

[0025] Betweenness centrality: BC(i) = Σ(σ) xy (i) / σ xy ); where x and y are any pair of nodes in the network except i, and σ xy Let σ be the total number of shortest paths between x and y. xy (i) represents the number of shortest paths passing through i;

[0026] a) Leader dog: if and only if its out-degree centrality OD(i) is higher than the preset leader threshold Th lead Furthermore, the ratio of its out-degree centrality to its in-degree centrality, OD(i) / ID(i), is greater than 1.

[0027] b) Follower dog: if and only if its in-degree centrality ID(i) is higher than the preset follower threshold Thfollow Furthermore, the ratio of its out-degree centrality to its in-degree centrality, OD(i) / ID(i), is less than 1.

[0028] c) Connector: A connector can be identified as a "connector" if it meets any of the following conditions:

[0029] Influential Connectors: When their eigenvector centrality EC(i) is higher than a preset core threshold Th core ;

[0030] For structural bridges: when their betweenness centrality BC(i) is higher than a preset bridge threshold Th between ;

[0031] d) Independent dog: if and only if its out-degree centrality OD(i), in-degree centrality ID(i), eigenvector centrality EC(i), and betweenness centrality BC(i) are all lower than their respective preset independence thresholds Th independent .

[0032] Furthermore, in step S4, the behavioral state S i (t) includes:

[0033] Boundary crossing status: The dog has exceeded the preset electronic fence range;

[0034] Deviation status: The distance between the dog and its group leader is greater than a preset threshold;

[0035] Abandoned state: The duration of a solitary dog ​​wandering at low speed or aimlessly exceeds a preset threshold;

[0036] Abnormal barking state: The frequency or duration of the dog's barking exceeds the preset threshold.

[0037] Furthermore, in step S4, control command D i (t) generation includes:

[0038] C1. Determine the dog's current behavioral status based on the dog's real-time behavior data and group behavior map;

[0039] C2. Based on the dog's role tag and behavioral status, match the corresponding command combination from the preset command type mapping table;

[0040] C3. When multiple behavioral states need to be regulated, the highest priority control command is selected and issued according to the preset priority decision logic. i (t);

[0041] C4. Based on the dog's historical response rate and learning curve, adjust the intensity or duration of commands to improve training effectiveness and dog acceptance.

[0042] Furthermore, in step C4, the strength of the adjustment instruction follows the following adaptive algorithm:

[0043] ;

[0044] Among them, S new For the strength of the new instruction, S old R is the strength of the previous instruction. hist γ represents the dog's historical average response success rate within the most recent time window, and γ is a preset adjustment factor used to control the magnitude of intensity adjustment.

[0045] Furthermore, in step S3, the preset role definition threshold is dynamically updated based on the historical behavioral data of the dog population through a statistical learning model, including the following steps:

[0046] Collect historical character recognition data and behavioral response data to construct a training dataset;

[0047] The training dataset is trained using machine learning algorithms to predict the optimal threshold;

[0048] The leadership threshold Th is adjusted periodically based on the prediction results. lead Following the threshold Th follow Core threshold Th core Bridge threshold Th between and independent threshold Th independent In order to adapt to changes in group behavior.

[0049] Furthermore, in step S5, model optimization includes: using behavioral response data R i (t) Train a neural network model to optimize the "state-role-behavior" multidimensional instruction mapping table or adjust the instruction generation strategy, including:

[0050] The behavior response data R i (t) is compared with the expected behavior to calculate the response error;

[0051] Based on the response error, the weight parameters of the neural network model are updated using the backpropagation algorithm;

[0052] Based on the updated model, the decision logic for instruction combinations or instruction priorities in the instruction mapping table is adjusted to improve the accuracy of control and dog acceptance.

[0053] The present invention also provides a system for implementing a group behavior regulation method for multi-dog collaborative training, comprising:

[0054] Several smart collars integrate a three-axis accelerometer, GPS module, microphone, UWB module, microcontroller unit, vibration and audio-visual actuator, as well as ZigBee and BLE dual communication modules for data acquisition, command reception and execution;

[0055] The control unit is used to run core algorithms including data preprocessing, two-layer dynamic map construction, multi-dimensional role dynamic recognition, differentiated instruction generation and closed-loop adaptive adjustment, and provides a visual interactive interface.

[0056] Cloud servers are used for large-scale data storage, computation and training of neural network models, and distribution of updated model parameters to the control terminal.

[0057] The present invention has the following beneficial effects:

[0058] (1) This invention constructs a two-layer dynamic graph that integrates physical space and social influence, and uses multi-dimensional graph theory centrality indicators to identify the role of each dog in real time, thereby achieving precise differentiated control from a "one-to-many" unified command to "differentiated control based on the dog and its state". This method fundamentally avoids command conflicts caused by differences in role behavior and can reduce the command conflict rate.

[0059] (2) The system of the present invention treats the dog pack as an organic whole for coordinated guidance, which can efficiently coordinate the behavior of the leader dog, the follower dog and the independent dog, and improve the efficiency of group coordinated training. The system also has strong dynamic adaptation and self-learning capabilities, and can continuously optimize the role recognition threshold and command strategy through closed-loop feedback, and evolve itself as the behavior of the dog pack changes, ensuring the accuracy and effectiveness of long-term application.

[0060] (3) By introducing quantitative analysis of social influence, this invention enhances the perception depth of the deep logic of group interaction and improves the intelligence level of the system. At the communication and control level, a dual-channel low-latency transmission and priority decision-making mechanism are adopted to ensure the real-time performance and reliability of instructions, and the instruction intensity is adaptively adjusted based on the historical response rate, which significantly improves animal welfare while ensuring training effect.

[0061] (4) This invention effectively overcomes the inherent defects of existing dog training technology in group collaborative control, and provides a reliable technical solution for realizing intelligent, precise, efficient and humane large-scale dog group collaborative training and management. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating a group behavior control method based on multi-dog collaborative training according to the present invention.

[0063] Figure 2This is a flowchart of a group behavior control method based on multi-dog collaborative training according to the present invention.

[0064] Figure 3 This is a schematic diagram of the structure of a dog training collar for a group behavior control method and system based on multi-dog collaborative training according to the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1:

[0067] This embodiment provides a group behavior regulation method based on multi-dog collaborative training, such as... Figure 1 and Figure 2 As shown, it includes the following steps:

[0068] S1. Data Acquisition and Preprocessing: Spatial location data and social behavior data of each dog are collected by wearing smart collars on several dogs (i=1,2,…,N). Spatial location data includes GPS coordinates and UWB positioning data. Social behavior data includes acceleration, voice, and interaction frequency. The collected raw data undergoes filtering, noise reduction, and interpolation to improve data quality.

[0069] S2. Based on preprocessed spatial location data and social behavior data, construct a group behavior map G={G} representing the interaction relationships among dogs. phy G soc}, where G phy For the physical space relationship map, G soc For social influence mapping;

[0070] Physical Spatial Relationship Map G phy Calculate the change in relative distance Δd between any two dogs i and j. ij Correlation coefficient r with speed ij Then, the weights of physical spatial relationships are calculated:

[0071]

[0072] Where α and β are preset weighting coefficients, satisfying α+β=1.

[0073] Social Influence Graph G soc Based on social behavior data, multiple indicators of directed social relationships are calculated, such as imitation delay M.ij Interaction response rate R ij Attention Allocation A ij Intimacy Index C ij After normalizing each indicator, a weighted sum is taken to obtain the overall social influence weight:

[0074]

[0075] Among them, I k,norm Let ω be the k-th normalized social relationship indicator value. k The corresponding preset weight coefficient is K, where K is the total number of indicators used, k=1,2,3…K.

[0076] S3. Dynamic Role Recognition: Based on the topological structure and weight evolution of the group behavior graph G, the multi-dimensional graph theory centrality index of each dog node in the graph is calculated and compared with a preset role definition threshold to dynamically identify the real-time role label C of each dog in the group. i (t); specifically including:

[0077] By calculating the following graph theory centrality index and assigning a corresponding dynamic role label C to each dog according to the preset judgment rules, i (t)::

[0078] For each dog node i, calculate the centrality indices for the following four dimensions:

[0079] Out-degree centrality: OD(i) = ΣS ij , representing the sum of the active social influence of dog i on other dogs;

[0080] In-degree centrality: ID(i) = ΣS ji This represents the sum of the passive social influences that dog i receives from other dogs;

[0081] Eigenvector centrality: EC(i) = (1 / λ)ΣS ji EC(j) measures the coreness of dog i's influence in the global network;

[0082] Betweenness centrality: BC(i) = Σ(σ) xy (i) / σ xy ), where x and y are any pair of nodes in the network except i, and σ xy Let σ be the total number of shortest paths between x and y. xy (i) represents the number of shortest paths passing through dog i. This metric measures the degree to which dog i acts as a "bridge" in the transmission of group information or behavior;

[0083] Dynamic Role Determination: The calculated centrality index value is compared with the preset role definition threshold, and a real-time role label C is assigned to each dog according to the following determination rules. i (t):

[0084] a) Leader dog: if and only if its out-degree centrality OD(i) is higher than the preset leader threshold Th lead Furthermore, the ratio of its out-degree centrality to its in-degree centrality, OD(i) / ID(i), is greater than 1. This indicates that the dog has a strong active influence and is a net influence outputter.

[0085] b) Follower dog: if and only if its in-degree centrality ID(i) is higher than the preset follower threshold Th follow Furthermore, the ratio of its out-degree centrality to its in-degree centrality, OD(i) / ID(i), is less than 1. This indicates that the dog is more susceptible to the influence of other dogs.

[0086] c) Connector: A connector can be identified as a "connector" if it meets any of the following conditions:

[0087] Influential Connectors: When their eigenvector centrality EC(i) is higher than a preset core threshold Th core This indicates that it is in the core circle of the social network.

[0088] For structural bridges: when their betweenness centrality BC(i) is higher than a preset bridge threshold Th between This indicates that it plays a key connecting role in the group.

[0089] d) Independent dog: if and only if its out-degree centrality OD(i), in-degree centrality ID(i), eigenvector centrality EC(i), and betweenness centrality BC(i) are all lower than their respective preset independence thresholds Th independent This indicates that the dog interacts less with the group and behaves relatively independently.

[0090] S4. Monitor the current behavioral status of each dog. i (t), and based on its current role C i (t) and behavioral state S i (t), construct an initial "state-role-behavior" multidimensional instruction mapping model, generate a "state-role-behavior" multidimensional instruction mapping table, and query to obtain the corresponding control instruction D. i (t);

[0091] Real-time monitoring and assessment of the current behavioral state S of each dog i (t), the main states include:

[0092] Boundary crossing status: The dog has exceeded the preset geographic boundary of the electronic fence.

[0093] Deviation status: The distance between the dog and the dynamically identified leader dog in its group is greater than the preset deviation threshold.

[0094] Wandering state: Especially for individuals identified as "independent dogs", when the duration of their slow or aimless wandering exceeds a preset threshold.

[0095] Abnormal barking status: Audio is collected through the microphone on the smart collar, and the analysis determines whether the frequency or duration of the dog's barking exceeds the preset safety or quiet threshold.

[0096] S5. Instruction execution and closed-loop model optimization; optimizing control instructions D... i (t) The data is sent to the corresponding smart collar to implement control measures and continuously monitors the dog's behavioral response data R. i (t), the response data R i (t) is uploaded to the cloud server to optimize the initial "state-role-behavior" multi-dimensional instruction mapping model and adjust the intensity of subsequent instructions.

[0097] Command Issuance and Execution: The generated control command D i (t) The command is transmitted to the corresponding smart collar via a wireless communication network (such as ZigBee or BLE). The actuators on the collar (vibration motor, speaker, LED light) execute the command to intervene in the dog's behavior in real time.

[0098] Response Monitoring and Data Upload: After the command is executed, the smart collar continuously monitors the dog's behavioral response data. i (t) (such as whether they return to the boundary, whether they approach the leader, whether they stop barking, etc.), and upload this response data to the cloud server.

[0099] Model optimization: On a cloud server, utilizing the large amount of collected behavioral response data R i (t) Continuously optimize the initial "state-role-behavior" multi-dimensional instruction mapping model. The optimization process includes:

[0100] The actual response data R i (t) is compared with the expected behavioral target to calculate the response error.

[0101] Based on this error, algorithms such as backpropagation are used to update the weight parameters of the neural network model used for instruction mapping or policy generation.

[0102] Based on the updated model, the instruction combinations in the instruction mapping table are adjusted or the instruction priority decision logic is optimized, thereby forming a closed-loop adaptive learning system that continuously improves the accuracy and efficiency of group behavior regulation.

[0103] Example 2: Adaptive threshold update and model iteration.

[0104] This embodiment is a further refinement of steps S3 and S5 in Embodiment 1, illustrating its adaptive evolution mechanism:

[0105] Dynamic updates of character thresholds: preset character definition thresholds (Th lead ,Th follow ,Th core ,Th between ,Th independent The role recognition capabilities are not fixed. The system periodically collects historical role recognition data and corresponding behavioral response data to construct a training dataset. Machine learning algorithms are used to analyze this dataset, predicting the optimal values ​​of various thresholds that will produce the best regulatory effect under the current group behavior pattern, and making periodic adjustments accordingly to ensure that role recognition capabilities adapt to group evolution. Machine learning algorithms include gradient boosting decision trees or support vector machine models.

[0106] Continuous Iteration of the Command Mapping Model: Model optimization in step S5 is an ongoing process. After aggregating the response data of all dogs, the cloud server periodically retrains the command mapping neural network. The training objective is to minimize the "response error," that is, to ensure that the commands output by the model guide the dogs to produce the expected behavior with the highest probability. After updating the network parameters through the backpropagation algorithm, a new version of the command mapping table or decision strategy is generated and incrementally updated to the control end, thereby continuously improving the overall intelligence level of the system.

[0107] Furthermore, in step S2, a physical spatial relationship map G is constructed. phy The methods include:

[0108] Calculate the change in relative distance Δd between any two dogs i and j. ij Correlation coefficient r with speed ij ;

[0109] Calculate the weights W of physical spatial relationships ij And based on this, an adjacency matrix is ​​constructed as the graph G. phy The representation of:

[0110]

[0111] Where, Δd ij r represents the change in the relative distance between dogs i and j. ij The velocity correlation coefficient is α, and β are preset weighting coefficients used to balance the effects of distance and velocity, with α+β=1.

[0112] Furthermore, in step S2, a social influence graph G is constructed. soc The methods include:

[0113] Based on social behavior data, quantify and calculate at least one directed social relationship indicator I. k Directed social relationship indicator I k Includes: Imitation delay index M ij Interactive response rate metric R ij Attention Allocation Index A ij Intimacy Index C ij ;

[0114] By normalizing and weighting multiple indicators, the comprehensive social influence weight S from dog i to dog j is calculated. ij And based on this, construct the graph G soc :

[0115]

[0116] Among them, I k,norm Let ω be the k-th normalized social relationship indicator value. k The corresponding preset weight coefficient is K, where K is the total number of indicators used, k=1,2,3…K.

[0117] Furthermore, in step S3, by calculating the following graph theory centrality index and according to preset judgment rules, a corresponding dynamic role label C is assigned to each dog. i (t):

[0118] Out-degree centrality: OD(i) = ΣS ij ;

[0119] In-degree centrality: ID(i) = ΣS ji ;

[0120] Eigenvector centrality: EC(i) = (1 / λ)ΣS ji EC(j);

[0121] Betweenness centrality: BC(i) = Σ(σ) xy (i) / σ xy ); where x and y are any pair of nodes in the network except i, and σ xy Let σ be the total number of shortest paths between x and y. xy (i) represents the number of shortest paths passing through i;

[0122] a) Leader dog: if and only if its out-degree centrality OD(i) is higher than the preset leader threshold Th lead Furthermore, the ratio of its out-degree centrality to its in-degree centrality, OD(i) / ID(i), is greater than 1.

[0123] b) Follower dog: if and only if its in-degree centrality ID(i) is higher than the preset follower threshold Thfollow Furthermore, the ratio of its out-degree centrality to its in-degree centrality, OD(i) / ID(i), is less than 1.

[0124] c) Connector: A connector can be identified as a "connector" if it meets any of the following conditions:

[0125] Influential Connectors: When their eigenvector centrality EC(i) is higher than a preset core threshold Th core ;

[0126] For structural bridges: when their betweenness centrality BC(i) is higher than a preset bridge threshold Th between ;

[0127] d) Independent dog: if and only if its out-degree centrality OD(i), in-degree centrality ID(i), eigenvector centrality EC(i), and betweenness centrality BC(i) are all lower than their respective preset independence thresholds Th independent .

[0128] Furthermore, in step S4, the behavioral state S i (t) includes:

[0129] Boundary crossing status: The dog has exceeded the preset electronic fence range;

[0130] Deviation status: The distance between the dog and its group leader is greater than a preset threshold;

[0131] Abandoned state: The duration of a solitary dog ​​wandering at low speed or aimlessly exceeds a preset threshold;

[0132] Abnormal barking state: The frequency or duration of the dog's barking exceeds the preset threshold.

[0133] Furthermore, in step S4, based on the dog's real-time role C i (t) and the current behavioral state S i (t), generate control command D i (t). This process specifically includes:

[0134] E1. Status Judgment: Based on real-time data streams and group behavior maps, the behavioral status of dogs is judged in real time.

[0135] E2. Command Matching: Query the preset "Status-Role-Behavior" multidimensional command mapping table. This mapping table defines the command combinations that different roles should receive in a specific state. For example, a guiding vibration may be applied to a "following dog" in a "deviation state", while an audio-visual warning may be used for a "lone dog" in the same state.

[0136] E3. Priority Decision: When the system detects that the same dog has multiple behavioral states that require regulation, it selects and issues the single control command with the highest priority according to the preset priority logic (e.g., safety-related states take precedence over efficiency-related states). i (t).

[0137] E4. Intensity Adaptive Adjustment: Based on the dog's historical response data, dynamically adjust the intensity (such as vibration intensity, sound volume) or duration of commands. The adjustment follows the adaptive algorithm:

[0138]

[0139] Among them, S new For the strength of the new instruction, S old R is the strength of the previous instruction. hist γ represents the dog's historical average response success rate within the most recent time window, and γ is a preset adjustment factor used to control the adjustment range. This effectively improves training effectiveness and dog acceptance.

[0140] This section elaborates on step S4 of claim 1, detailing how the system generates precise, differentiated, and adaptive control commands based on the dog's real-time status, dynamic role, and historical performance. This process is a crucial bridge connecting "system cognition" and "physical control," and its core logic lies in "varying according to state, role, and history."

[0141] The four sub-steps are explained in detail below.

[0142] E1. Behavioral status judgment;

[0143] This step is a prerequisite for triggering command generation. The control unit continuously receives real-time data streams from the smart collar and, combined with the group behavior map G, determines the current behavioral state S of each dog. i (t) performs real-time computation and pattern recognition.

[0144] Boundary Crossing Status: The system compares the dog's real-time GPS coordinates with a preset electronic fence polygonal area. If the coordinates fall outside the area, a "boundary crossing status" is immediately triggered. The electronic fence can be a fixed geographical boundary or a dynamic boundary that changes over time or for specific tasks.

[0145] Deviation Status: The system first obtains the currently identified "leader dog" ID from the role recognition module, and then calculates the real-time Euclidean distance between the dog and the leader dog. If this distance exceeds a preset deviation threshold (e.g., 10 meters), a "deviation status" is triggered. This threshold can be dynamically adjusted according to the group size and environment.

[0146] Loose state: This state primarily applies to individuals identified as "independent dogs." The system analyzes accelerometer data and movement trajectories to determine if the dog is in a low-speed (e.g., speed < 0.5 m / s) and directionless wandering mode. If this state persists for more than a preset threshold (e.g., 30 seconds), the "loose state" is triggered to prevent the dog from completely separating from group interaction.

[0147] Abnormal barking state: The smart collar's microphone collects ambient audio and identifies and counts barking events in real time through cloud-based or local audio analysis models (such as those based on spectral characteristics and energy detection). When the number of barks within a unit of time (e.g., 1 minute) or the duration of a single bark exceeds a preset quiet threshold, an "abnormal barking state" is triggered.

[0148] E2. Instruction matching based on role and state;

[0149] This step enables the differentiation of control strategies. The system maintains a preset "state-role-behavior" instruction type mapping table. This table defines the instruction type and initial intensity to be used for each (behavior state, role) combination.

[0150] Command types include vibration, sound (such as a buzzer at a specific frequency, white noise, or a pre-recorded command word), light (such as an LED blinking), and combinations of these modalities.

[0151] Example of matching logic:

[0152] Scene 1: State = Deviation, Role = Follower Dog.

[0153] Command: Guiding vibration. The vibration motor on the collar's side closest to the leader dog generates pulses to guide the dog toward its return.

[0154] Scene 2: State = Deviation, Role = Independent Dog.

[0155] Command: Audible and visual warning. The collar emits a short warning sound accompanied by a flashing red light, reminding the individual to return to the group area, rather than through physical guidance.

[0156] Scenario 3: Status = Abnormal barking, Role = Leader dog.

[0157] Instructions: Gentle acoustic intervention. Play a specific frequency sound that will attract its attention, avoiding strong stimuli that could affect its leadership and mood.

[0158] Scenario 4: State = Out of bounds, Role = Any.

[0159] Command: Strong vibration + warning sound. For safety reasons, all characters must use a strong and immediate deterrent command.

[0160] E3, Multi-state priority decision-making;

[0161] When the system detects that the same dog simultaneously meets the triggering conditions for multiple behavioral states, an arbitration mechanism is needed to determine which issue should be handled first. The system has a built-in set of preset priority decision logic.

[0162] Priority rules:

[0163] Safety First: Boundary crossing has the highest priority because it is directly related to the dog's safety.

[0164] Collaboration takes priority: deviation from the state is secondary, because it affects the overall collaboration of the group and task execution.

[0165] Order takes precedence: abnormal barking is secondary, as it disrupts group order and quiet.

[0166] Efficiency First: The free state has the lowest priority, and this optimization is mainly aimed at improving the efficiency of independent dogs.

[0167] Decision execution: For example, if a dog simultaneously triggers "deviation" and "abnormal barking," the system will, based on the above logic, prioritize generating and issuing control command D for the "deviation state." i (t). The system will only evaluate and process lower priority states after a high priority state has been cleared.

[0168] E4. Adaptive adjustment of command intensity; this step achieves personalized and refined control. The system maintains an individual file for each dog, recording its historical average response success rate R. hist R hist It refers to the percentage of dogs that respond correctly to commands within the most recent time window (such as the past 10 interventions or 1 hour).

[0169] The adjustment of command strength follows the following adaptive algorithm:

[0170]

[0171] Among them, S new For the strength of the new instruction, S old R is the strength of the previous instruction. hist γ represents the dog's historical average response success rate within the most recent time window, and γ is a preset adjustment factor used to control the magnitude of intensity adjustment.

[0172] γ: A preset adjustment factor (e.g., 0.2) used to control the magnitude of the adjustment and prevent the intensity change from being too drastic.

[0173] Algorithm logic:

[0174] When R histAt lower values ​​(e.g., 0.3, indicating poor response), (1-R) hist The value is relatively large (0.7), indicating the strength S of the new instruction. new It will be significantly enhanced to provide clearer cues.

[0175] When R hist At higher values ​​(e.g., 0.9, indicating a good response), (1-R) hist The value is relatively small (0.1), and the strength S of the new instruction is... new Only slightly increase or even remain unchanged, to show encouragement and avoid overstimulation.

[0176] In addition, the system can adjust the duration of commands based on the learning curve. For example, for dogs that learn quickly, the duration of commands can be gradually shortened to encourage them to form a faster conditioned reflex.

[0177] Furthermore, in step S3, the preset role definition threshold is dynamically updated based on the historical behavioral data of the dog population through a statistical learning model, including the following steps:

[0178] Collect historical character recognition data and behavioral response data to construct a training dataset;

[0179] The training dataset is trained using machine learning algorithms to predict the optimal threshold;

[0180] The leadership threshold Th is adjusted periodically based on the prediction results. lead Following the threshold Th follow Core threshold Th core Bridge threshold Th between and independent threshold Th independent In order to adapt to changes in group behavior.

[0181] This invention introduces a dynamic update mechanism for role definition thresholds based on a statistical learning model, enabling the system to evolve self-evolving and continuously maintain the accuracy of role recognition. This mechanism runs periodically as a background process (e.g., every 24 hours or week), and its process is as follows:

[0182] T1. Collect historical data and construct the training dataset; this step is the data preparation stage for model training. The system purposefully collects two types of historical data from the database:

[0183] Historical role identification data: This data represents the centrality index sequence of all dogs calculated and recorded by the system based on old thresholds within the past period, along with their corresponding real-time role labels C. i (t).

[0184] Historical behavioral response data: Synchronized with the aforementioned role recognition data, this data reflects the dog's response to control command D. iThe response data R of (t) i (t). The key is that the system evaluates the success rate of each intervention (i.e., whether the behavior is corrected within the expected time).

[0185] Based on the above data, a training dataset for threshold optimization is constructed. Each sample in this dataset represents the overall performance of a dog within a time segment, and its features typically include:

[0186] Statistical characteristics of centrality indicators: the mean, variance, maximum, and minimum values ​​of the dog's out-degree centrality, in-degree centrality, eigenvector centrality, and betweenness centrality during this period.

[0187] Role stability characteristics: the frequency of the dog's role switching and the percentage of time it spends in a particular role (such as a leader dog).

[0188] The label for this sample (i.e., the target predicted by the model) is a comprehensive performance score. This score is calculated as follows:

[0189] Overall performance rating = w1 * character stability rating + w2 * command response success rate

[0190] in:

[0191] Role stability score: This measures whether the role assigned to the dog is consistent with its long-term behavioral patterns. For example, a dog that has long been identified as a leader dog will have a low stability score if its role frequently jumps between leader and follower.

[0192] Command Response Success Rate: The percentage of control commands issued to the dog by the system that are successfully executed after the dog has been assigned a role. A higher success rate indicates that the current role recognition and subsequent command matching are more effective.

[0193] w1 and w2 are weighting coefficients used to balance the importance of the two indicators.

[0194] T2. Use a machine learning model to train and predict the optimal threshold; this step is the core computational step for dynamic updates.

[0195] Model selection: Employ machine learning algorithms capable of handling regression problems, such as Gradient Boosting Decision Tree (GBDT), Random Forest, or Support Vector Regression (SVR). These models can effectively capture the non-linear relationship between features and complex objectives.

[0196] Training objective: The model's input consists of features from the aforementioned training dataset, and its output is the corresponding overall performance score. The training objective is for the model to learn: under a given group behavior pattern (represented by features), what threshold setting will maximize the overall performance score of the entire group.

[0197] Finding the optimal threshold: After training the model, the system does not directly use it for prediction, but rather as an evaluation function. Subsequently, optimization algorithms (such as GridSearch or Bayesian Optimization) are used within a reasonable parameter space to search for the set of threshold combinations that maximizes the predicted population average overall performance score. lead ,Th follow ,Th core ,Th between ,Th independent This set of thresholds is the "optimal threshold" for the current period.

[0198] T3. Periodically adjust the role definition thresholds; this step is the decision execution phase of the model.

[0199] Threshold update: The system safely updates the new generation of optimal threshold combinations obtained in step T2 into the core role recognition algorithm module, replacing the old thresholds.

[0200] Gradual updates: To avoid drastic fluctuations in system recognition results caused by sudden threshold changes, a smooth update strategy can be adopted, for example:

[0201] New threshold = η * optimal threshold + (1 - η) * old threshold

[0202] Here, η is a smoothing factor between 0 and 1, which achieves a smooth transition of the threshold.

[0203] By introducing this dynamic update mechanism, the present invention achieves the following beneficial effects:

[0204] Adaptive evolution: The system can automatically adjust its "cognitive model" to follow the natural growth of the dog population and changes in social structure, without the need for manual recalibration, which greatly improves the long-term applicability and intelligence level of the system.

[0205] Closed-loop optimization: This mechanism, together with the S5 instruction mapping model optimization, forms a dual closed-loop optimization system. The former optimizes the accuracy of "role recognition," while the latter optimizes the effectiveness of "behavior control." The two complement each other and jointly improve the performance of the entire system.

[0206] Personalized adaptation: This mechanism can adapt to the unique "culture" of different dog groups. For example, the threshold of the leader dog in a naturally active group may be generally higher than that in a quiet group. The system can capture and adapt to this difference through learning and updating.

[0207] This dynamic update mechanism is one of the key technologies for achieving intelligent and unmanned operation and maintenance in this invention, ensuring the accuracy and robustness of the group behavior control method in long-term operation.

[0208] Furthermore, in step S5, model optimization includes: using behavioral response data R i (t) Training a neural network model to optimize the "state-role-behavior" multidimensional instruction mapping table or adjust the instruction generation strategy, including:

[0209] The behavior response data R i (t) is compared with the expected behavior to calculate the response error;

[0210] Based on the response error, the weight parameters of the neural network model are updated using the backpropagation algorithm;

[0211] Based on the updated model, the decision logic for instruction combinations or instruction priorities in the instruction mapping table is adjusted to improve the accuracy of control and dog acceptance.

[0212] This section elaborates on step S5 of claim 1. The initial "state-role-behavior" multidimensional instruction mapping table is based on expert experience or prior knowledge, which may not cover all complex scenarios or adapt to the individualized learning characteristics of each dog. To address this issue, this invention introduces a continuous model optimization mechanism based on neural networks and closed-loop feedback in step S5. This mechanism enables the system to learn from the results of each intervention, continuously adjusting its decision-making logic, thereby achieving increasingly precise and efficient group behavior control.

[0213] This optimization mechanism is executed on a cloud server, and its workflow is a typical "evaluation-learning-application" cycle. The specific steps are as follows, and you can refer to the model optimization flowchart specially drawn by this invention.

[0214] M1. Data preparation and response error calculation; this step forms the data foundation and defines the objectives for the optimization process.

[0215] Data preparation: The cloud server aggregates behavioral response data uploaded from all smart collars. i (t). These data and the issued control commands D i (t), The dog's role when issuing commands (C) i (t) and behavioral state S i(t) precisely corresponds to form complete training samples (S) i (t),C i (t),D i (t),R i (t)).

[0216] Define the expected behavior: for each (S) i (t),C i (t),D i (t) triples, the system has a clearly defined expected behavior. For example:

[0217] Command: Apply "moderate vibration" to the "follower dog" that is "out of control".

[0218] Expected behavior: The dog will move at least 5 meters toward the leader dog within 10 seconds.

[0219] Calculate the response error: The actual monitored response R... i (t) is compared with the expected behavior, and a response error E is quantified. The method of error calculation can be customized according to the state type:

[0220] For regression problems (such as the regression distance of "deviation from state"): the error E can be the mean square error (MSE) between the actual distance moved and the expected distance moved.

[0221] For classification problems (such as whether "abnormal barking state" has stopped): the error E can be the cross-entropy loss, which measures the difference between the probability distribution of "barking has stopped" and "barking continues" and the actual situation.

[0222] Overall score: A comprehensive score function can also be defined (e.g., successful correction = 0, partial correction = 0.5, failure = 1), with E as the loss value.

[0223] M2. Update the neural network model based on backpropagation; this step is the core of model learning. This invention uses a neural network as the decision-making brain for the "state-role-behavior" mapping.

[0224] Model Input: The input layer of the neural network receives a feature vector, which includes at least:

[0225] Current behavior state S i The one-hot encoding of (t), such as [0,1,0,0], represents "off-state".

[0226] Current character tag C i The one-hot encoding of (t), such as [1,0,0,0], represents "leader dog".

[0227] Additional contextual features, such as dog ID (for personalization), ambient noise level, etc.

[0228] Model Output: Scheme A (Direct Command Generation): The output layer directly corresponds to all possible control commands, such as "weak vibration," "strong vibration," "short beep," "long beep," etc. Each output node represents the probability of selecting that command. The model selects the command with the highest probability and issues it.

[0229] Option B (Q-Learning Value Assessment): The output layer corresponds to a given (S) i (t),C i Given (t), the expected value (Q-value) of executing each possible instruction is determined. The system selects the instruction with the highest Q-value to issue. Here, "value" refers to the predicted effectiveness of the instruction (low response error, high acceptability).

[0230] Model training and updates:

[0231] Forward propagation: This involves transferring (S) from the training samples... i (t),C i (t) Input the current version of the neural network to obtain the result for instruction D i The predicted value (probability or Q value) of (t).

[0232] Loss Calculation: Combine the predicted value with the response error E calculated in step M1 to calculate the loss function L. For example, in scenario B, the loss function could be the mean squared error between the predicted Q value and (1-E), since a low error implies a high value.

[0233] Backpropagation: Using the backpropagation algorithm, the gradient of the loss function L with respect to each weight parameter in the neural network is calculated.

[0234] Parameter update: Using optimizers (such as Adam, SGD), the weight parameters of the neural network are finely adjusted based on the calculated gradient, so that the model can predict lower response errors in future decisions.

[0235] M3, Adjusting Instruction Mapping and Decisions; This step applies the learning outcomes to the actual system.

[0236] Dynamically update the instruction mapping table:

[0237] For scheme A, after training with a large amount of data, the neural network itself becomes a dynamic, non-linear "instruction map". The system can directly use the neural network as an instruction generator to replace or update the old static lookup table.

[0238] For scheme B, the decision results of the neural network for common (state, role) combinations (i.e., the instructions with the highest Q value) can be exported periodically and updated to the static instruction mapping table used by the control end to improve decision-making efficiency.

[0239] Optimizing command priority decision-making logic: The model can also learn more complex strategies. For example, by analyzing large amounts of data, the model might discover that for an isolated dog simultaneously in a state of "crossing the boundary and barking abnormally," prioritizing the "crossing the boundary" action yields a higher overall success rate than prioritizing the "barking" action. This implicit "priority decision-making logic" is internalized into the neural network's weights. When the system deploys the neural network itself, this optimization is automatically implemented.

[0240] By introducing this neural network-based closed-loop optimization mechanism, this invention achieves a leap from "static rules" to "dynamic intelligence":

[0241] Personalized control: The model can learn each dog's sensitivity and acceptance of different types of commands (vibration, sound, light), thus achieving true "training according to the dog's individual needs".

[0242] Continuous evolution: The system no longer relies on fixed expert rules, but is able to continuously learn from actual interactions, adapt to new behavioral patterns and environmental changes, and its performance continues to improve over time.

[0243] Discovering optimal strategies: Neural networks can uncover more efficient combinations of complex instructions and decision sequences that humans have not thought of, thus potentially achieving better group collaborative training results.

[0244] Improving animal welfare: By adaptively adjusting the intensity and type of commands, the minimum stimulus that can effectively guide behavior is found, greatly improving the comfort and acceptance of dogs during training.

[0245] This model optimization mechanism, together with the aforementioned dynamic threshold update mechanism, constitutes the "dual brain" of the system of this invention. They are respectively responsible for the self-evolution of "cognitive" role recognition and "decision" behavior regulation, ensuring the method's advanced nature, adaptability, and efficiency in long-term operation.

[0246] Example 3:

[0247] The present invention also provides a system for implementing a group behavior regulation method for multi-dog collaborative training, comprising:

[0248] Several smart collars integrate a three-axis accelerometer, GPS module, microphone, UWB module, microcontroller unit, vibration and audio-visual actuator, as well as ZigBee and BLE dual communication modules for data acquisition, command reception and execution;

[0249] The control unit is used to run core algorithms including data preprocessing, two-layer dynamic map construction, multi-dimensional role dynamic recognition, differentiated instruction generation and closed-loop adaptive adjustment, and provides a visual interactive interface.

[0250] Cloud servers are used for large-scale data storage, computation and training of neural network models, and distribution of updated model parameters to the control terminal.

[0251] This embodiment provides a group behavior control system for implementing all the steps described in the above method, such as... Figure 3 As shown, the system includes:

[0252] Several smart collars: worn around the neck of each dog, these serve as the system's data acquisition and command execution terminals. Hardware integration:

[0253] Sensing modules: a three-axis accelerometer, a GPS module, a microphone, and a UWB module, used to comprehensively collect data on the dog's movement, position, sound, and precise relative distance.

[0254] Processing and Communication Module: The microcontroller unit (MCU) is responsible for initial data processing and logic control; the dual communication modules of ZigBee and Bluetooth Low Energy (BLE) ensure low-power networking within the group and efficient data transmission with the control terminal.

[0255] Execution unit: vibration motor, speaker, LED light strip, used to receive and execute multimodal control commands such as sound, light, and touch issued by the control terminal.

[0256] Control unit: Typically a mobile device or portable computing device running dedicated software. It is responsible for running the core algorithm, including:

[0257] Preprocess the data uploaded by the collar.

[0258] Construct and update a two-layer dynamic group behavior map (G phy G soc ).

[0259] Execute a multi-dimensional role dynamic recognition algorithm.

[0260] Differentiated control instructions are generated based on the "state-role-behavior" model.

[0261] Implement closed-loop adaptive adjustment logic.

[0262] It provides a visual interactive interface that displays group dynamics, individual status, role distribution, and system intervention logs in real time, facilitating monitoring and interaction by trainers.

[0263] Cloud server: As the backend support platform of the system, it is responsible for:

[0264] Securely store and manage large-scale historical and behavioral response data from the control end.

[0265] It provides powerful computing capabilities for running complex neural network model training and optimization algorithms.

[0266] The parameters of the newly trained model are regularly distributed to each control terminal to ensure that the entire system can continuously learn and evolve to adapt to long-term behavioral changes in different dog groups.

[0267] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for regulating group behavior based on multi-dog cooperative training, characterized in that, Comprising the following steps: S1, collecting spatial position data and social behavior data of each dog through smart collars worn on several dogs i=1, 2,..., N, and preprocessing the collected data; S2, constructing a group behavior graph G={G phy ,G soc} representing the interaction relationship between dogs based on the pre-processed spatial position data and social behavior data, wherein G phy is a physical space relationship graph, and G soc is a social influence graph; S3, based on the topology and weight evolution of the group behavior atlas, by calculating the multidimensional graph theory centrality index of each dog node in the atlas, and comparing with the preset role definition threshold, to dynamically identify the real-time role label C played by each dog in the group i (t); S4, monitoring the current behavior state S of each dog i (t), and according to its current role C i (t) and behavior state S i (t), an initial "state-role-behavior" multi-dimensional instruction mapping model is constructed, a "state-role-behavior" multi-dimensional instruction mapping table is generated, and a corresponding control instruction D is obtained by querying i (t); S5, send control instruction D i (t) to the corresponding smart collar to execute control, and continuously monitor the behavior response data R of the dog i (t), upload the response data R i (t) to the cloud server for model optimization of the initial "state-role-behavior" multi-dimensional instruction mapping model, and adjust the subsequent instruction intensity.

2. The method according to claim 1, wherein, In step S2, the physical space relationship graph G is constructed phy The method comprises: Calculate the relative distance change amount Ad between any two dogs i and j ij and the speed correlation coefficient r ij ; Computing physical space relationship weights W ij and constructing an adjacency matrix as a representation of the graph G phy : ； wherein, Δd ij is the relative distance change amount of dogs i and j; r ij is the speed-related coefficient, and α and β are preset weight coefficients for balancing the distance and speed influences, and α+β=1.

3. The method of claim 1, wherein the plurality of dogs are trained to perform the group behavior. In step S2, the social influence graph G is constructed soc The method comprises: quantifying at least one directed social relationship indicator I based on the social behavior data k directed social relationship indicator I k including: an imitation delay indicator M ij , an interaction response rate indicator R ij , an attention allocation indicator A ij and a closeness index indicator C ij ; By normalizing and weighting summing a plurality of indexes, a comprehensive social influence weight S from dog i to dog j is calculated ij , and a graph G is constructed therefrom soc : ； wherein I k,norm is the kth normalized social relationship index value, ω k is its corresponding preset weight coefficient, K is the total number of indexes used, and k = 1, 2, 3, …, K.

4. The method of claim 1, wherein the plurality of dogs are trained to perform the group behavior. In the step S3, by calculating the following graph theory centrality index, and according to the preset determination rule, a corresponding dynamic role label C is given to each dog i (t): Out-degree centrality: OD(i) =∑S ij ; In-degree centrality: ID(i) =∑S ji ; Eigen vector centrality: EC(i) = (1 / λ)∑S ji • EC(j); Betweenness centrality: BC(i) =∑(σ xy (i) / σ xy ); where x, y are any pair of nodes in the network other than i, σ xy is the total number of shortest paths between x and y, and σ xy (i) is the number of shortest paths that pass through i. a) leader dog: if and only if its out-degree centrality OD(i) is higher than a pre-set leadership threshold Th lead , and its ratio of out-degree centrality to in-degree centrality OD(i) / ID(i) is greater than 1 ; b) Follows dog: if and only if its in-degree centrality ID(i) is higher than a pre-set follow threshold Th follow , and its out-degree centrality ratio to in-degree centrality OD(i) / ID(i) is less than 1. c) Connector: If any of the following conditions is met, it can be determined as "connector": Influential connector: when its eigenvector centrality EC(i) is higher than a preset core threshold Th core ; Structural bridge: when its betweenness centrality BC(i) is higher than a preset bridge threshold Th between ; d) Independent dog: if and only if its out-degree centrality OD(i), in-degree centrality ID(i), eigenvector centrality EC(i) and betweenness centrality BC(i) are all lower than the respective pre-set independent threshold Th independent .

5. The method of claim 1, wherein, In step S4, the behavior state S i (t) comprises: Out-of-bound state: the dog exceeds the preset electronic fence range; Deviation state: the distance between the dog and the leader of the group is greater than the preset threshold; Free state: the independent dog is in low-speed or purposeless wandering for a long time exceeding the preset threshold; Abnormal barking state: the barking frequency or duration of the dog exceeds the preset threshold.

6. The method of claim 1, wherein, In step S4, the control command D i (t) generating comprising: C1, judging the current behavior state of the dog according to the real-time behavior data of the dog and the group behavior graph; C2, according to the role label and behavior state of the dog, matching the corresponding instruction combination from the preset instruction type mapping table; C3. When there are multiple behavior states that need to be regulated, the control instruction D with the highest priority is selected and issued according to the preset priority decision logic i (t); C4, based on the historical response rate and learning curve of the dog, adjusting the intensity or duration of the instruction to improve the training effect and dog acceptance.

7. The method of claim 1, wherein the plurality of dogs are trained to perform the group behavior. In step C4, the intensity of the instruction is adjusted according to the following adaptive algorithm: ； where S new is the intensity of the new instruction, S old is the intensity of the previous instruction, R hist is the historical average response success rate of the dog in the most recent time window, and γ is a preset adjustment factor to control the magnitude of the intensity adjustment.

8. The method of claim 1, wherein the plurality of dogs are trained to perform the group behavior. In step S3, the preset role definition threshold is dynamically updated based on the historical behavior data of the dog group through a statistical learning model, including the following steps: Collect historical role recognition data and behavior response data to build a training data set; Use machine learning algorithms to train the training data set to predict the optimal threshold; periodically adjust the leadership threshold Th according to the prediction results lead , the following threshold Th follow , the core threshold Th core , the bridge threshold Th between and the independent threshold Th independent to adapt to changes in group behavior.

9. The method of claim 1, wherein, In step S5, the model optimization includes: using the behavioral response data R i (t) training the neural network model to optimize the "state-role-behavior" multidimensional instruction mapping table or adjust the instruction generation strategy, including: The behavioral response data R i (t) comparing to the expected behavior, calculating response error; Based on the response error, update the weight parameters of the neural network model through the back propagation algorithm; According to the updated model, adjust the instruction combination or instruction priority decision logic in the instruction mapping table to improve the control accuracy and dog acceptance.

10. A group behavior regulation system for implementing the method of any one of claims 1 to 9, characterized by Comprising: A number of smart collars integrated with three-axis accelerometer, GPS module, microphone, UWB module, micro control unit, vibration and sound and light execution unit, and ZigBee and BLE dual communication module, for executing data collection, instruction reception and execution; Control end, for running core algorithms including data preprocessing, double-layer dynamic graph construction, multi-dimensional role dynamic identification, differentiated instruction generation and closed-loop adaptive adjustment, and providing visual interactive interface; Cloud server: for performing large-scale storage of data, calculation and training of neural network model, and distributing updated model parameters to the control end.

Images