Optimization system and method for sound-light cooperative bird prevention strategy based on deep reinforcement learning

Abstract

This invention discloses a system and method for optimizing a sound-optical coordinated bird control strategy based on deep reinforcement learning, comprising the following steps: collecting bird flock monitoring data and bird control data, and performing preprocessing; extracting bird flock behavioral features and constructing a dynamic bird flock propagation network; modeling the propagation of the bird flock's startled response process; identifying a set of key propagation nodes; calculating the bird flock behavioral information entropy parameters and generating a critical judgment result for flock escape; establishing an ecological game relationship between the set of bird control strategies and the set of bird flock behavioral responses, and calculating bird flock behavioral adaptation parameters; constructing a reinforcement learning environment state representation vector to generate a sound-optical coordinated bird control strategy; implementing sound-optical directional stimulation on the set of key propagation nodes, and iteratively executing until the optimized result of the sound-optical coordinated bird control strategy is output. This invention uses deep reinforcement learning and flock propagation modeling methods to achieve sound-optical coordinated bird control strategy optimization, possessing the advantages of high bird control efficiency, strong adaptability, and precise control.

Description

Technical Field

[0001] This invention relates to the field of biological bird deterrence, and more particularly to an optimization system and method for a sound-optical coordinated bird deterrence strategy based on deep reinforcement learning. Background Technology

[0002] With frequent bird activity in airports, farmland, landfills, and open urban areas, bird gatherings and flight behaviors can easily impact aviation safety, agricultural production, and the public environment. Therefore, bird control technology has gradually become an important research direction in this field. Existing bird control technologies mainly employ sound waves, lasers, or a combination of sound and light, using acoustic or optical devices with fixed parameters to stimulate bird flocks and achieve a repelling effect. Some systems trigger bird control devices after obtaining the location of bird flocks through video or radar monitoring, but overall, they still rely primarily on rule-based trigger control or simple parameter adjustments, lacking systematic modeling and dynamic optimization control capabilities based on the propagation patterns of bird flock behavior.

[0003] However, existing bird control technologies still have significant shortcomings in practical applications. On the one hand, existing systems typically rely solely on the location or number of birds for bird control, lacking analysis of the transmission mechanism of startled behavior within the flock, resulting in low bird control efficiency and high energy consumption. On the other hand, traditional bird control strategies struggle to characterize the behavioral adaptation process of bird flocks to bird control measures, and cannot dynamically optimize bird control strategies based on changes in flock behavior, thus easily leading to diminishing bird control effectiveness or repeated return of flocks. Summary of the Invention

[0004] One objective of this invention is to propose an acoustic-optical coordinated bird deterrence strategy optimization system and method based on deep reinforcement learning. This invention uses deep reinforcement learning and swarm propagation modeling methods to achieve acoustic-optical coordinated bird deterrence strategy optimization, which has the advantages of high bird deterrence efficiency, strong adaptability and precise control.

[0005] The method for optimizing acousto-optical bird deterrence strategy based on deep reinforcement learning according to an embodiment of the present invention includes the following steps:

[0006] Collect bird flock monitoring data and bird control data within the target area, and perform preprocessing to generate a bird flock status data set;

[0007] Bird flock behavior characteristics were extracted from bird flock monitoring data in the bird flock status dataset, and a dynamic bird flock propagation network was constructed.

[0008] Based on the dynamic propagation network of bird flocks, the propagation model of the bird flock startled response process is constructed, the propagation parameters of startled behavior are calculated, and a set of bird flock startled propagation states is generated.

[0009] Based on the set of bird flock startled propagation states, identify the set of key propagation nodes in the dynamic propagation network of bird flocks;

[0010] Based on the set of bird flock startled propagation states and the set of key propagation nodes, calculate the bird flock behavior information entropy parameters, generate a characterization result of fear information propagation efficiency, and generate a critical judgment result for group escape.

[0011] Based on the set of bird flock startled propagation states and the results of fear information propagation efficiency characterization, an ecological game relationship is established between the set of bird deterrence strategies and the set of bird flock behavioral responses. Bird flock behavioral adaptation parameters are calculated, and bird flock behavioral adaptation prediction results are generated.

[0012] By integrating bird flock state data set, key propagation node set, fear information propagation efficiency representation results, group escape criticality judgment results and bird flock behavior adaptation prediction results, a reinforcement learning environment state representation vector is constructed, and an acoustic-light coordinated bird deterrence strategy is generated through a deep reinforcement learning policy network.

[0013] Based on the sound-light coordinated bird deterrence strategy, sound-light directional stimulation is applied to the key propagation node set, and the parameters of the deep reinforcement learning strategy network are updated. This process is repeated until the optimized result of the sound-light coordinated bird deterrence strategy is output.

[0014] Optionally, the preprocessing includes time synchronization, spatial registration, anomaly removal, and normalization. The bird flock monitoring data includes the spatial location of the bird flock, the flight speed of the bird flock, and the population density of the bird flock. The bird deterrence control data includes the sound source output parameters, the light source output parameters, and the sound and light triggering sequence.

[0015] Optionally, the generation of the bird flock dynamic propagation network specifically includes:

[0016] Read the bird flock status data set, and arrange the bird flock spatial location, bird flock flight speed and bird flock density according to a unified time index to generate a bird flock time-series status data sequence;

[0017] Bird flock behavior features are extracted from bird flock time-series state data sequences to generate a set of bird flock behavior features;

[0018] Bird flock behavior characteristics include the spatial distribution of individual birds, the speed distribution of individual birds, and the population density distribution in local areas;

[0019] Based on the spatial location of each bird within the same time segment, calculate the spatial distance between individual birds, determine the spatial relationships between individual birds, and generate a set of spatial relationships.

[0020] Based on the flock flight speeds of individual birds within the same time segment, calculate the speed difference and speed direction difference between individual birds, determine the motion correlation between individual birds, and generate a set of motion correlation relationships.

[0021] Based on the set of bird flock behavior characteristics, calculate the corresponding values ​​of behavioral changes of different individual birds in continuous time segments, determine the response correlation between individual birds, and generate a set of response correlations.

[0022] By combining the sets of spatial associations, motion associations, and response associations, a bird flock behavior association network is constructed. The node connection relationships and edge weight parameters are updated according to continuous time segments to generate a bird flock dynamic propagation network.

[0023] Optionally, the generation of the set of bird flock startled propagation states specifically includes:

[0024] Read the node connection relationships, edge weight parameters, and bird flock behavior feature sets corresponding to each bird individual in the bird flock dynamic propagation network to determine the basic state of behavior propagation of each bird individual in the current time segment.

[0025] Each individual bird is taken as the current propagation source. Based on the node connection relationship and edge weight parameters between the current propagation source and the other individual birds, the set of direct propagation objects corresponding to the current propagation source is determined.

[0026] Read the set of bird flock behavior characteristics corresponding to the current propagation source and the set of bird flock behavior characteristics corresponding to the set of direct propagation objects, and calculate the intensity of the behavioral influence of the current propagation source on each direct propagation object.

[0027] Obtain the edge weight parameters corresponding to the current propagation source and each direct propagation object, and determine the propagation intensity of the current propagation source to each direct propagation object according to the intensity of the behavior influence and the corresponding edge weight parameters. Then, summarize the results to generate the propagation parameters of the startled behavior corresponding to each direct propagation object of the current propagation source.

[0028] The parameters of the startled behavior propagation are repeatedly obtained for each individual bird, and the parameters of the startled behavior propagation for each individual bird are arranged in order of time segments to generate a set of startled behavior propagation states of the bird flock.

[0029] Optionally, the generation of the key propagation node set specifically includes:

[0030] Read the set of bird flock startled propagation states, extract the startled behavior propagation parameters of each bird in each time segment according to a unified time index, and generate a set of propagation parameter sequences.

[0031] Bird individuals whose startled behavior propagation parameters are greater than or equal to a preset propagation intensity screening threshold are selected from the propagation parameter sequence set to generate a candidate key node set, and the startled behavior propagation parameter sequence corresponding to each bird individual in the candidate key node set is recorded synchronously.

[0032] For each individual bird in the candidate key node set, read the node connection relationship and edge weight parameters corresponding to that individual bird in the bird flock dynamic propagation network, count the number of adjacent nodes to obtain the node connectivity parameter, and sum the adjacent edge weight parameters to obtain the node edge weight summary value.

[0033] The cumulative value of the propagation contribution is calculated for the propagation parameter sequence of startled behavior corresponding to each individual bird, and the difference of the cumulative value of the propagation contribution is calculated to generate the propagation persistence parameter;

[0034] The cumulative value of propagation contribution, propagation persistence parameter, node connectivity parameter, and node edge weight summation value are merged to generate a comprehensive score value for key nodes. The candidate key node set is then sorted based on the comprehensive score value to generate a key node ranking result.

[0035] Extract the set of key propagation nodes from the key node sorting results.

[0036] Optionally, the generation of the group escape criticality determination result specifically includes:

[0037] Obtain the set of bird flock startled propagation states, extract the startled behavior propagation parameters of each node pair in each time segment according to a unified time index, and filter the set of node pairs containing key propagation nodes based on the set of key propagation nodes to generate the set of key node propagation parameters.

[0038] The propagation parameter set of key nodes is summarized and statistically analyzed in each time segment to generate the total propagation value corresponding to that time segment. Based on the total propagation value, the propagation parameters of each node for startled behavior are proportionalized to generate the propagation probability distribution of key nodes corresponding to that time segment.

[0039] The key node propagation probability distributions corresponding to each time segment are arranged in time index order to generate a key node propagation probability sequence. Based on the key node propagation probability sequence, the information entropy parameter of bird flock behavior is calculated for each time segment to generate an information entropy sequence.

[0040] The difference between the information entropy parameters of adjacent time segments in the information entropy sequence is calculated to generate an information entropy change sequence, and the result of the fear information propagation efficiency is generated based on the information entropy sequence and the information entropy change sequence.

[0041] The results of the fear information dissemination efficiency characterization include a dissemination efficiency score sequence and an efficiency stability score sequence;

[0042] Based on the results of the fear information propagation efficiency characterization, the critical trigger time segment is determined, and the critical judgment result of the group escape is generated;

[0043] Arrange the critical trigger time segment index and the critical judgment flag value in a corresponding manner, and output the critical judgment result of the group escaping.

[0044] Optionally, the generation of the bird flock behavior adaptation prediction results specifically includes:

[0045] Organize the spatial location, flight speed, and population density of bird flocks according to a unified time index to form a sequence of bird flock status characteristics, and perform differential calculations on adjacent time segments to obtain the changes in bird flock spatial location, flight speed, and population density.

[0046] Organize the propagation parameters of startled behavior of each node according to a unified time index, and use the set of key propagation nodes to filter the node pairs containing key propagation nodes to form the key node startled propagation feature sequence. In each time segment, sum the key node startled propagation feature sequence to obtain the propagation intensity summary value, and perform difference operation on the propagation intensity summary values ​​of adjacent time segments to obtain the propagation intensity change.

[0047] The difference between the propagation efficiency score sequence and the efficiency stability score sequence is used to obtain the change in propagation efficiency and the change in efficiency stability. These changes are then aligned with the changes in bird flock spatial location, bird flock flight speed, bird flock population density, and propagation intensity by time index to form a game state sequence.

[0048] The bird control data is discretized and encoded to generate bird deterrence action identifiers for each discretized combination, forming a set of bird deterrence action identifiers;

[0049] Within each time segment, the bird deterrence action identifier of that time segment is matched one-to-one with the changes in the spatial location of the bird flock, the changes in the flight speed of the bird flock, the changes in the population density of the bird flock, the changes in the propagation intensity, the changes in the propagation efficiency, and the changes in the efficiency stability of the bird flock, and the bird flock behavior response set is generated by summarizing them.

[0050] An ecological game relationship is established based on the game state sequence, the set of bird deterrence action identifiers, and the set of bird flock behavior responses. The set of bird flock behavior responses corresponding to the bird deterrence action identifiers is statistically analyzed and bird flock behavior adaptation parameters are generated. Based on the bird flock behavior adaptation parameters, bird flock behavior adaptation prediction results are generated.

[0051] Optionally, the generation of the acoustic-optical coordinated bird-repelling strategy specifically includes:

[0052] Time alignment is performed on the spatial location, flight speed, and population density of bird flocks in the bird flock state dataset to generate a bird flock state vector.

[0053] Perform set representation on the set of key propagation nodes to generate key node representation vectors;

[0054] The propagation efficiency score and efficiency stability score in the performance representation of fear information propagation efficiency are combined to generate an efficiency representation vector.

[0055] Encode the critical decision identifier value and the critical trigger time segment index in the group escape critical decision result to generate a critical decision vector;

[0056] The optimal response bird-driving action identifier in the bird flock behavior adaptation prediction result is encoded to generate an action identifier vector, and then concatenated with the corresponding response mean vector to generate an adaptation prediction vector.

[0057] Within each time segment, the flock state vector, key node representation vector, efficiency representation vector, critical decision vector, and adaptive prediction vector are respectively dimensionally aligned and normalized, and then spliced ​​and fused to generate reinforcement learning environment state representation vectors, which are arranged in time index order to generate a sequence of environment state representation vectors.

[0058] A set of bird-repelling actions is constructed based on the sound source output parameters, light source output parameters, and the audio-visual triggering time sequence. The reinforcement learning environment state representation vector corresponding to the current time segment in the environmental state representation vector sequence is input into the deep reinforcement learning policy network, and the corresponding audio-visual collaborative bird-repelling policy is output.

[0059] Optionally, the generation of the optimized results of the acoustic-optical coordinated bird deterrence strategy specifically includes:

[0060] Within the current time segment, the sound source output parameters, light source output parameters, and sound and light triggering sequence are determined based on the sound and light coordinated bird deterrence strategy. The sound source emission direction and the light coverage area are pointed to the spatial region corresponding to the set of key propagation nodes, and sound and light directional stimulation is performed.

[0061] After the audio-visual directional stimulation ends, a set of bird flock state data for the next time segment is generated, and the set of bird flock startled propagation state, the result of the fear information propagation efficiency characterization, and the result of the group escape threshold judgment are updated simultaneously.

[0062] The current state is represented by the reinforcement learning environment state vector corresponding to the current time segment, and the current action is represented by the bird-driving action identifier corresponding to the sound-light coordinated bird-driving strategy. The next state is represented by the reinforcement learning environment state vector corresponding to the next time segment, and a reward value is generated.

[0063] The current state, current action, reward value, and next state are combined to form a state transition sample. The state transition samples corresponding to each time segment are then summarized to generate a state transition sample set.

[0064] The parameters of the deep reinforcement learning policy network are updated based on the state transition sample set, and the corresponding acoustic-optical bird deterrence policy is output after the update.

[0065] The process involves iteratively executing audio-visual directional stimulation, reward value calculation, state transition sample set generation, and deep reinforcement learning strategy network parameter updates until the optimized result of the audio-visual collaborative bird deterrence strategy is output.

[0066] An acoustic-optical coordinated bird-repelling strategy optimization system based on deep reinforcement learning according to an embodiment of the present invention includes:

[0067] The bird flock data acquisition module is used to collect bird flock monitoring data and bird control data within the target area, and perform preprocessing to generate a bird flock status data set.

[0068] The bird flock dynamic propagation network construction module is used to extract bird flock behavior characteristics based on bird flock monitoring data in the bird flock status dataset and construct the bird flock dynamic propagation network.

[0069] The startled behavior propagation modeling module is used to model the propagation of bird flock startled response process, calculate startled behavior propagation parameters, and generate a set of bird flock startled propagation states.

[0070] The critical propagation node identification module is used to identify the set of critical propagation nodes in the dynamic propagation network of a flock of birds based on the set of bird flock startled propagation states.

[0071] The Fear Information Efficiency and Critical Judgment Module is used to calculate the information entropy parameter of bird flock behavior, generate the fear information propagation efficiency representation result, and generate the critical judgment result of group escape.

[0072] The Ecological Game Modeling and Adaptation Prediction Module is used to establish the ecological game relationship between the set of bird deterrence strategies and the set of bird flock behavior responses, calculate bird flock behavior adaptation parameters, and generate bird flock behavior adaptation prediction results.

[0073] The state fusion and policy generation module is used to construct the state representation vector of the reinforcement learning environment and input the state representation vector of the reinforcement learning environment into the deep reinforcement learning policy network to generate an acoustic-optical collaborative bird deterrence policy.

[0074] The acoustic-optical directional stimulation and strategy update module is used to implement acoustic-optical directional stimulation on the set of key propagation nodes according to the acoustic-optical coordinated bird repelling strategy, and update the parameters of the deep reinforcement learning strategy network. This process is repeated until the optimized results of the acoustic-optical coordinated bird repelling strategy are output.

[0075] The beneficial effects of this invention are:

[0076] This invention proposes a method and system for optimizing aesthetic-optical bird control strategies based on deep reinforcement learning. By systematically modeling the propagation characteristics of bird flock behavior and combining this with intelligent strategy optimization, a more efficient and adaptive bird control mechanism is achieved. By collecting bird flock monitoring data and constructing a dynamic bird flock propagation network, the spatial, movement, and behavioral response relationships among individual birds are comprehensively analyzed, thus accurately depicting the behavioral propagation patterns of the flock at the group level. Based on this, by modeling the propagation process of startled bird responses and calculating the propagation parameters of startled behavior, the system can identify the propagation path and intensity of startled behavior within the flock. Furthermore, it identifies key propagation nodes in the dynamic bird flock propagation network, enabling targeted intervention by bird control on key individuals that significantly influence flock behavior. This significantly improves the propagation efficiency of bird-repelling stimuli within the flock, reduces unnecessary equipment energy consumption, and enhances the overall bird-repelling effect.

[0077] Furthermore, this invention constructs a representation of the efficiency of fear information propagation by calculating the information entropy parameter of bird flock behavior, and determines the critical state of flock escape by combining the trend of information entropy changes. This enables the system to identify the key time stage in the transition from localized fright and spread to flock escape, thus providing a more time-series basis for bird control strategies. Simultaneously, by establishing an ecological game relationship between the set of bird control strategies and the set of bird flock behavior responses, statistical modeling is performed on the changes in bird flock behavior caused by different bird control actions, generating bird flock behavior adaptation prediction results. This allows the system to predict the adaptation trend of bird flocks to bird control measures in advance, avoiding the problem of diminishing bird control effectiveness caused by a single strategy in traditional bird control methods. Based on this, a reinforcement learning environment state representation vector is constructed by fusing the bird flock state data set, the set of key propagation nodes, the representation of fear information propagation efficiency, the determination of the critical state of flock escape, and the bird flock behavior adaptation prediction results. A deep reinforcement learning policy network is then used to generate a sound-light coordinated bird control strategy, enabling the bird control system to continuously learn and optimize control strategies in a dynamic environment, thereby achieving coordinated optimization control of sound source output parameters, light source output parameters, and sound-light triggering timing. By continuously executing directional acoustic and optical stimuli and using a strategy update mechanism for iterative optimization, this invention can form a stable and convergent optimal acoustic and optical synergistic bird-repelling strategy, making the bird-repelling process more precise, efficient, and intelligent, and effectively improving the long-term operational stability and bird-repelling effect of the bird-repelling system in complex environments. Attached Figure Description

[0078] 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:

[0079] Figure 1This is a flowchart of the sound-optical coordinated bird-repelling strategy optimization method based on deep reinforcement learning proposed in this invention;

[0080] Figure 2 This diagram illustrates the identification of key propagation nodes in the proposed deep reinforcement learning-based acoustic-optical coordinated bird control strategy optimization method. Detailed Implementation

[0081] 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.

[0082] refer to Figure 1 and Figure 2 The method for optimizing the acoustic-optical coordinated bird-repelling strategy based on deep reinforcement learning includes the following steps:

[0083] Collect bird flock monitoring data and bird control data within the target area, and perform preprocessing to generate a bird flock status data set;

[0084] Bird flock behavior characteristics were extracted from bird flock monitoring data in the bird flock status dataset, and a dynamic bird flock propagation network was constructed.

[0085] Based on the dynamic propagation network of bird flocks, the propagation model of the bird flock startled response process is constructed, the propagation parameters of startled behavior are calculated, and a set of bird flock startled propagation states is generated.

[0086] Based on the set of bird flock startled propagation states, identify the set of key propagation nodes in the dynamic propagation network of bird flocks;

[0087] Based on the set of bird flock startled propagation states and the set of key propagation nodes, calculate the bird flock behavior information entropy parameters, generate a characterization result of fear information propagation efficiency, and generate a critical judgment result for group escape.

[0088] Based on the set of bird flock startled propagation states and the results of fear information propagation efficiency characterization, an ecological game relationship is established between the set of bird deterrence strategies and the set of bird flock behavioral responses. Bird flock behavioral adaptation parameters are calculated, and bird flock behavioral adaptation prediction results are generated.

[0089] By integrating bird flock state data set, key propagation node set, fear information propagation efficiency representation results, group escape criticality judgment results and bird flock behavior adaptation prediction results, a reinforcement learning environment state representation vector is constructed, and an acoustic-light coordinated bird deterrence strategy is generated through a deep reinforcement learning policy network.

[0090] Based on the sound-light coordinated bird deterrence strategy, sound-light directional stimulation is applied to the key propagation node set, and the parameters of the deep reinforcement learning strategy network are updated. This process is repeated until the optimized result of the sound-light coordinated bird deterrence strategy is output.

[0091] In this embodiment, preprocessing includes time synchronization, spatial registration, anomaly removal, and normalization. Bird flock monitoring data includes the spatial location, flight speed, and population density of the bird flock. Bird deterrence control data includes sound source output parameters, light source output parameters, and acoustic-optical triggering timing. The bird flock monitoring data is a set of data collected by bird flock monitoring units deployed in the target area to characterize the activity status of the bird flock. The sound source output parameters are a set of data used to characterize the acoustic output status of the sound drive device, including sound pressure level, sound source frequency, and sound source emission direction. The light source output parameters are a set of data used to characterize the optical output status of the optical drive device, including light source flicker frequency, beam scanning trajectory, and illumination coverage area. The acoustic-optical triggering timing is a set of data used to characterize the coordinated triggering sequence and triggering time relationship between the sound drive device and the optical drive device, including the sound drive triggering time, the optical drive triggering time, and the acoustic-optical triggering interval.

[0092] In this embodiment, the generation of the bird flock dynamic propagation network specifically includes:

[0093] Read the bird flock status data set, and arrange the bird flock spatial location, bird flock flight speed and bird flock density according to a unified time index to generate a bird flock time-series status data sequence;

[0094] Bird flock behavior features are extracted from bird flock time-series state data sequences to generate a set of bird flock behavior features;

[0095] Bird flock behavior characteristics include the spatial distribution of individual birds, the speed distribution of individual birds, and the population density distribution in local areas;

[0096] The spatial location, flight speed, and population density of the bird flocks are read according to a unified time index for each time segment. The spatial distribution of individual birds is determined based on the spatial location of the bird flocks for each time segment. The speed distribution of individual birds is determined based on the flight speed of the bird flocks for each time segment. The population density distribution of local areas is determined based on the population density of the bird flocks for each time segment. The spatial distribution of individual birds, the speed distribution of individual birds, and the population density distribution of local areas for the same time segment are integrated to generate a set of bird flock behavior features.

[0097] The spatial distribution of individual birds is determined by the spatial location of each bird in the target area based on the corresponding flock spatial location within the same time segment. The speed distribution of individual birds is determined by the speed distribution of each bird in the target area based on the corresponding flock flight speed within the same time segment. The population density distribution of local areas is determined by the population density of bird flocks in each local area within the same time segment.

[0098] Based on the spatial location of each bird within the same time segment, calculate the spatial distance between individual birds, determine the spatial relationships between individual birds, and generate a set of spatial relationships.

[0099] The process involves obtaining the spatial location of each individual bird within the same time segment and determining the spatial coordinates of each individual bird within the target area. Then, using each individual bird as the current calculation object, the spatial coordinates of the current calculation object and the remaining individual birds are read, and the spatial distance between the current calculation object and the remaining individual birds is calculated one by one, generating a spatial distance sequence for the current calculation object. Next, the spatial distance calculation is repeated for all individual birds, and the spatial distance sequences are arranged according to the bird individual index order to generate a spatial distance matrix. The spatial distances of each pair of individual birds are read and compared with a preset spatial association distance threshold. When the spatial distance is less than or equal to the preset spatial association distance threshold, a spatial association relationship is determined for the corresponding pair of individual birds, and a corresponding spatial association relationship identifier is generated. When the spatial distance is greater than the preset spatial association distance threshold, a spatial association relationship is determined for the corresponding pair of individual birds, and a corresponding spatial association relationship identifier is generated. Finally, the spatial association relationship identifiers for each pair of individual birds are summarized to generate a spatial association relationship set.

[0100] Based on the flock flight speeds of individual birds within the same time segment, calculate the speed difference and speed direction difference between individual birds, determine the motion correlation between individual birds, and generate a set of motion correlation relationships.

[0101] The algorithm acquires the flock flight speeds of individual birds within the same time segment, determining the flight speed and direction of each individual bird. Using each individual bird as the current calculation object, it reads the flight speeds and directions of the current calculation object relative to the other individual birds, calculating the speed difference and speed-direction difference between the current calculation object and each other, generating a speed difference sequence and a speed-direction difference sequence for the current calculation object. This speed difference and speed-direction difference calculation is repeated for all individual birds, and the speed difference sequences and speed-direction difference sequences are arranged according to the bird's index, generating a speed difference matrix and a speed-direction difference matrix. Finally, it reads the corresponding speed difference and speed-direction difference for each individual bird. The data is compared with preset speed difference thresholds and preset speed direction difference thresholds. When the speed difference is less than or equal to the preset speed difference threshold and the speed direction difference is less than or equal to the preset speed direction difference threshold, it is determined that the corresponding bird pair has a motion association relationship and a corresponding motion association relationship identifier is generated. When the speed difference is greater than the preset speed difference threshold or the speed direction difference is greater than the preset speed direction difference threshold, it is determined that the corresponding bird pair does not have a motion association relationship and a corresponding motion association relationship identifier is generated. The motion association relationship identifiers for each bird pair are summarized to generate a motion association relationship set. The speed direction difference is the angle difference between the flight directions of two bird pairs within the same time segment.

[0102] Based on the set of bird flock behavior characteristics, calculate the corresponding values ​​of behavioral changes of different individual birds in continuous time segments, determine the response correlation between individual birds, and generate a set of response correlations.

[0103] The system reads the spatial distribution and velocity distribution of each bird individual within adjacent time segments, and also reads the local population density distribution of each local region. It compares the changes in spatial distribution and velocity distribution of any two birds within adjacent time segments to generate spatial distribution variation difference and velocity variation difference. Then, based on the changes in local population density distribution of the local regions where the two birds are located, it generates density distribution variation difference. The spatial distribution variation difference, velocity variation difference, and density variation difference are weighted and summed to generate a corresponding behavioral change value for each bird pair. The behavioral change value for each bird pair is compared with a preset behavioral change threshold. When the behavioral change value is less than or equal to the preset behavioral change threshold, a response correlation is determined for the corresponding bird pair, and a corresponding response correlation identifier is generated. The response correlation identifiers for each bird pair are summarized to generate a response correlation set.

[0104] By combining the sets of spatial association relationships, motion association relationships, and response association relationships, a bird flock behavior association network is constructed, and the node connection relationships and edge weight parameters are updated according to continuous time segments to generate a bird flock dynamic propagation network.

[0105] The process involves reading the spatial association identifier, motion association identifier, and response association identifier for each bird individual pair and arranging them according to a unified bird individual index. Then, a weighted sum of these identifiers is performed on the same bird individual pair to generate the association strength for that pair. Next, using each bird individual as a network node and the association strength of each pair as the connection strength between nodes, a bird flock behavior association network for the current time segment is constructed. The process also involves reading the bird flock behavior association networks for adjacent time segments, calculating the difference in association strength between the same bird individual pair in adjacent time segments to generate edge weight changes, and updating the edge weight parameters between corresponding nodes based on these changes. Next, node connections are updated based on newly added and disappeared association identifiers within adjacent time segments, generating updated node connections. Finally, the update results for each consecutive time segment are sequentially integrated to generate a bird flock dynamic propagation network, where the presence of a spatial association identifier is 1 and its absence is 0, the presence of a motion association identifier is 1 and its absence is 0, and the presence of a response association identifier is 1 and its absence is 0.

[0106] In this embodiment, the generation of the set of bird flock startled propagation states specifically includes:

[0107] Read the node connection relationship, edge weight parameters and bird flock behavior feature set corresponding to each bird individual in the bird flock dynamic propagation network, and determine the basic propagation state of each bird individual in the current time segment. The basic propagation state of behavior is the initial propagation state of each bird individual in the bird flock dynamic propagation network determined by the node connection relationship, edge weight parameters and bird flock behavior feature set corresponding to each bird individual in the current time segment.

[0108] Each individual bird is taken as the current propagation source. Based on the node connection relationship and edge weight parameters between the current propagation source and the other individual birds, the set of direct propagation objects corresponding to the current propagation source is determined.

[0109] The process involves obtaining the bird flock dynamic propagation network corresponding to the current time segment and selecting one bird individual as the current propagation source. Then, it reads the node connections between the current propagation source and the remaining birds, filters out birds with existing node connections, and generates a current propagation candidate object set. Next, it reads the edge weight parameters between each bird individual in the current propagation candidate object set and the current propagation source, and compares each edge weight parameter with a preset propagation weight threshold. When the edge weight parameter is greater than or equal to the preset propagation weight threshold, the corresponding bird individual is determined to be a direct propagation object of the current propagation source, and a corresponding direct propagation object identifier is generated. When the edge weight parameter is less than the preset propagation weight threshold, the corresponding bird individual is determined not to be a direct propagation object of the current propagation source, and a corresponding non-direct propagation object identifier is generated. Finally, the bird individuals corresponding to each direct propagation object identifier are summarized to generate the direct propagation object set corresponding to the current propagation source.

[0110] Read the set of bird flock behavior characteristics corresponding to the current propagation source and the set of bird flock behavior characteristics corresponding to the set of direct propagation objects, and calculate the intensity of the behavioral influence of the current propagation source on each direct propagation object.

[0111] The system reads the spatial distribution, velocity distribution, and local population density distribution of individual birds corresponding to the current propagation source, and also reads the spatial distribution, velocity distribution, and local population density distribution of individual birds corresponding to each directly propagated object. It compares the spatial distribution of individual birds corresponding to the current propagation source with that of each directly propagated object to generate a spatial distribution impact difference; it compares the velocity distribution of individual birds corresponding to the current propagation source with that of each directly propagated object to generate a velocity distribution impact difference; and it compares the local population density distribution of the current propagation source with that of each directly propagated object to generate a density distribution impact difference. The system then performs linear mapping on these spatial distribution impact differences, velocity distribution impact differences, and density distribution impact differences, and performs a weighted sum of the mapped results to generate the behavioral impact strength of the current propagation source on each directly propagated object.

[0112] Obtain the edge weight parameters corresponding to the current propagation source and each direct propagation object, and determine the propagation intensity of the current propagation source to each direct propagation object according to the intensity of the behavior influence and the corresponding edge weight parameters. Then, summarize the results to generate the propagation parameters of the startled behavior corresponding to each direct propagation object of the current propagation source.

[0113] The intensity of the impact of the behavior is weighted and summed with the corresponding edge weight parameters to generate the propagation intensity of the current propagation source to each directly propagated object;

[0114] The parameters of the startled behavior propagation are repeatedly obtained for each individual bird, and the parameters of the startled behavior propagation for each individual bird are arranged in order of time segments to generate a set of startled behavior propagation states of the bird flock.

[0115] In this embodiment, the generation of the key propagation node set specifically includes:

[0116] Read the set of bird flock startled propagation states, extract the startled behavior propagation parameters of each bird in each time segment according to a unified time index, and generate a set of propagation parameter sequences.

[0117] Bird individuals whose startled behavior propagation parameters are greater than or equal to a preset propagation intensity screening threshold are selected from the propagation parameter sequence set to generate a candidate key node set, and the startled behavior propagation parameter sequence corresponding to each bird individual in the candidate key node set is recorded synchronously.

[0118] For each individual bird in the candidate key node set, read the node connection relationship and edge weight parameters corresponding to that individual bird in the bird flock dynamic propagation network, count the number of adjacent nodes to obtain the node connectivity parameter, and sum the adjacent edge weight parameters to obtain the node edge weight summary value.

[0119] The cumulative value of the propagation contribution is calculated for the propagation parameter sequence of startled behavior corresponding to each individual bird, and the difference of the cumulative value of the propagation contribution is calculated to generate the propagation persistence parameter;

[0120] When generating the cumulative propagation contribution value, the propagation parameters of startled behavior of node pairs arranged by a unified time index are read, and the incoming and outgoing edge sets corresponding to each bird are determined according to the node connection relationship in the bird flock dynamic propagation network. Then, taking a single bird as the current calculation object, the propagation parameters of startled behavior of each node pair in the outgoing edge set corresponding to the bird in each time segment are read one by one and summed to obtain the output propagation contribution value of the bird in that time segment. At the same time, the propagation parameters of startled behavior of each node pair in the incoming edge set corresponding to the bird in each time segment are read one by one and summed to obtain the input propagation contribution value of the bird in that time segment. Next, the difference between the output propagation contribution value and the input propagation contribution value is calculated to generate the net propagation contribution value of the bird in that time segment. The net propagation contribution values ​​corresponding to each time segment are arranged in time index order to generate the net propagation contribution sequence corresponding to the bird. The net propagation contribution sequence is accumulated to generate the cumulative propagation contribution value corresponding to the bird.

[0121] The cumulative value of propagation contribution, propagation persistence parameter, node connectivity parameter, and node edge weight summation value are fused to generate a comprehensive score value for key nodes. The candidate key node set is then sorted based on the comprehensive score value to generate a key node ranking result. The fusion adopts normalization and weighted summation.

[0122] Extract the set of key propagation nodes from the key node sorting results;

[0123] Read the key node sorting results, extract the individual bird identifiers one by one according to the sorting order and read the corresponding key node comprehensive score value; compare the key node comprehensive score value with the preset key node comprehensive score threshold. When the key node comprehensive score value is greater than or equal to the preset key node comprehensive score threshold, write the corresponding individual bird identifier into the key propagation node set. When the key node comprehensive score value is less than the preset key node comprehensive score threshold, do not write it.

[0124] In this embodiment, the generation of the group escape criticality determination result specifically includes:

[0125] Obtain the set of bird flock startled propagation states, extract the startled behavior propagation parameters of each node pair in each time segment according to a unified time index, and filter the set of node pairs containing key propagation nodes based on the set of key propagation nodes to generate the set of key node propagation parameters.

[0126] The propagation parameter set of key nodes is summarized and statistically analyzed in each time segment to generate the total propagation value corresponding to that time segment. Based on the total propagation value, the propagation parameters of each node for startled behavior are proportionalized to generate the propagation probability distribution of key nodes corresponding to that time segment.

[0127] Read all node pairs’ startled behavior propagation parameters in the current time segment from the key node propagation parameter set, and sum them one by one to obtain the total propagation value; traverse the startled behavior propagation parameters of each node pair in the same time segment again, and perform proportionalization processing with the total propagation value as the denominator to generate the key node propagation probability value corresponding to each node pair; arrange the key node propagation probability values ​​of each node pair in the same time segment according to the node pair index order to generate the key node propagation probability distribution corresponding to the time segment.

[0128] The key node propagation probability distributions corresponding to each time segment are arranged in time index order to generate a key node propagation probability sequence. Based on the key node propagation probability sequence, the information entropy parameter of bird flock behavior is calculated for each time segment to generate an information entropy sequence.

[0129] Read the key node propagation probability distribution corresponding to each time segment and write it into the time series container according to a unified time index to generate a key node propagation probability sequence; take a single time segment as the current calculation object, read all key node propagation probability values ​​in the key node propagation probability distribution corresponding to that time segment, perform logarithmic mapping on each key node propagation probability value and multiply it by the key node propagation probability value to generate the entropy contribution value corresponding to that key node propagation probability value; sum all entropy contribution values ​​within the same time segment and take the negative sign to generate the bird flock behavior information entropy parameter corresponding to that time segment; arrange the bird flock behavior information entropy parameters corresponding to each time segment in time index order to generate an information entropy sequence;

[0130] The difference between the information entropy parameters of adjacent time segments in the information entropy sequence is calculated to generate an information entropy change sequence, and the result of the fear information propagation efficiency is generated based on the information entropy sequence and the information entropy change sequence.

[0131] The results of the fear information dissemination efficiency characterization include a dissemination efficiency score sequence and an efficiency stability score sequence;

[0132] The process involves reading the information entropy sequence, sequentially retrieving the information entropy parameters corresponding to adjacent time segments according to a unified time index, subtracting the information entropy parameter of the previous time segment from the information entropy parameter of the subsequent time segment to obtain the information entropy difference, and arranging each information entropy difference in time index order to generate an information entropy change sequence. Then, within a preset statistical window, the information entropy sequence is read, and the average value of the information entropy parameters within the window is calculated to generate a propagation efficiency score. Within the same statistical window, the information entropy change sequence is read, and the absolute value of each information entropy difference within the window is first taken, then the average value is calculated to generate an efficiency stability score. The propagation efficiency score and efficiency stability score are then arranged in a corresponding manner to generate a fear information propagation efficiency representation result, which is then output according to the time index.

[0133] Based on the results of the fear information propagation efficiency characterization, the critical trigger time segment is determined, and the critical judgment result of the group escape is generated;

[0134] The propagation efficiency score and efficiency stability score of the fear information propagation efficiency representation result are read time segment by time segment according to the unified time index, and compared with the preset efficiency critical threshold and preset stability threshold respectively. When the propagation efficiency score is greater than or equal to the preset efficiency critical threshold and the efficiency stability score is less than or equal to the preset stability threshold, and the propagation efficiency score corresponding to the previous time segment is less than the preset efficiency critical threshold or the efficiency stability score corresponding to the previous time segment is greater than the preset stability threshold, the current time segment is determined to be the critical trigger time segment, and a group escape critical judgment result containing the critical trigger time segment index and critical judgment identifier value is generated.

[0135] The propagation efficiency score is a score obtained by summarizing the information entropy sequence within a preset statistical window. A higher propagation efficiency score indicates that the spread of frightened behavior related to key propagation nodes is more extensive and more fully diffused within the statistical window. The efficiency stability score is a score obtained by summarizing the information entropy change sequence within a preset statistical window. A higher efficiency stability score indicates that the information entropy change is greater and the propagation process is more volatile. A lower efficiency stability score indicates that the information entropy change is smaller and the propagation process is more stable. The propagation efficiency score and efficiency stability score together constitute the representation of the efficiency of fear information propagation.

[0136] The critical escape determination result is the output of the determination of the critical time segment from the startled spread to the group escape based on the fear information propagation efficiency representation result. It includes the critical trigger time segment index and the critical determination identifier value.

[0137] Arrange the critical trigger time segment index and the critical judgment flag value in a corresponding manner, and output the critical judgment result of the group escaping.

[0138] In this embodiment, the generation of bird flock behavior adaptation prediction results specifically includes:

[0139] Organize the spatial location, flight speed, and population density of bird flocks according to a unified time index to form a sequence of bird flock status characteristics, and perform differential calculations on adjacent time segments to obtain the changes in bird flock spatial location, flight speed, and population density.

[0140] Organize the propagation parameters of startled behavior of each node according to a unified time index, and use the set of key propagation nodes to filter the node pairs containing key propagation nodes to form the key node startled propagation feature sequence. In each time segment, sum the key node startled propagation feature sequence to obtain the propagation intensity summary value, and perform difference operation on the propagation intensity summary values ​​of adjacent time segments to obtain the propagation intensity change.

[0141] The difference between the propagation efficiency score sequence and the efficiency stability score sequence is used to obtain the change in propagation efficiency and the change in efficiency stability. These changes are then aligned with the changes in bird flock spatial location, bird flock flight speed, bird flock population density, and propagation intensity by time index to form a game state sequence.

[0142] The bird control data is discretized and encoded to generate bird deterrence action identifiers for each discretized combination, forming a set of bird deterrence action identifiers;

[0143] Within each time segment, the bird deterrence action identifier of that time segment is matched one-to-one with the changes in the spatial location of the bird flock, the changes in the flight speed of the bird flock, the changes in the population density of the bird flock, the changes in the propagation intensity, the changes in the propagation efficiency, and the changes in the efficiency stability of the bird flock, and the bird flock behavior response set is generated by summarizing them.

[0144] An ecological game relationship is established based on the game state sequence, the bird deterrence action identifier set and the bird flock behavior response set. The bird flock behavior response set corresponding to the bird deterrence action identifier is statistically analyzed and bird flock behavior adaptation parameters are generated. Bird flock behavior adaptation prediction results are generated based on the bird flock behavior adaptation parameters.

[0145] Ecological game theory is used to describe the mapping relationship between the bird-driving actions in the bird-driving action set and the bird flock behavior response set;

[0146] Using time segments as units, the state segment index corresponding to each time segment in the game state sequence, the bird-repelling action identifier in the bird-repelling action identifier set, and the bird flock behavior response record corresponding to that time segment in the bird flock behavior response set are bound into triples to generate an ecological game sample set. The ecological game sample set is grouped according to the bird-repelling action identifier, and the mean values ​​of the changes in bird flock spatial location, bird flock flight speed, bird flock density, propagation intensity, propagation efficiency, and efficiency stability corresponding to each bird-repelling action identifier are calculated and normalized to obtain the response mean vector corresponding to the bird-repelling action identifier. The response mean vector is weighted and summed to obtain the utility value corresponding to the bird-repelling action identifier, and the utility value is generated repeatedly for all bird-repelling action identifiers to form a utility value set. The utility value set is normalized to obtain the bird flock behavior adaptation parameter set, and the bird-repelling action identifier with the largest bird flock behavior adaptation parameter under the game state sequence conditions corresponding to the current time segment is selected as the optimal response bird-repelling action identifier. The optimal response bird-repelling action identifier and the corresponding response mean vector are output as the bird flock behavior adaptation prediction result.

[0147] In this embodiment, the generation of the sound-light coordinated bird-repelling strategy specifically includes:

[0148] Time alignment is performed on the spatial location, flight speed, and population density of bird flocks in the bird flock state dataset to generate a bird flock state vector.

[0149] Perform set representation on the set of key propagation nodes to generate key node representation vectors;

[0150] Based on the set of key propagation nodes, the identifiers of individual birds within the set are determined, and the spatial location of each bird identifier corresponding to the flock is located in the flock status data set. The mean values ​​of all flock spatial locations within the set of key propagation nodes are calculated along the three dimensions of spatial coordinates to generate spatial center coordinates. The maximum and minimum values ​​of all flock spatial locations within the set of key propagation nodes are calculated along the three dimensions of spatial coordinates, and the difference is calculated to generate spatial range parameter values. The spatial distance between each flock spatial location and the spatial center coordinate value within the set of key propagation nodes is calculated, and the mean of all spatial distances is calculated to generate spatial dispersion parameter values. The spatial center coordinate values, spatial range parameter values, and spatial dispersion parameter values ​​are concatenated in a fixed order to generate a key node representation vector.

[0151] The propagation efficiency score and efficiency stability score in the performance representation of fear information propagation efficiency are combined to generate an efficiency representation vector.

[0152] Encode the critical decision identifier value and the critical trigger time segment index in the group escape critical decision result to generate a critical decision vector. When the critical decision identifier value is 1, the critical trigger time segment index is taken as the actual trigger time segment index. When the critical decision identifier value is 0, the critical trigger time segment index is taken as zero.

[0153] The optimal response bird-driving action identifier in the bird flock behavior adaptation prediction result is encoded to generate an action identifier vector, and then concatenated with the corresponding response mean vector to generate an adaptation prediction vector.

[0154] Within each time segment, the flock state vector, key node representation vector, efficiency representation vector, critical decision vector, and adaptive prediction vector are respectively dimensionally aligned and normalized, and then spliced ​​and fused to generate reinforcement learning environment state representation vectors, which are arranged in time index order to generate a sequence of environment state representation vectors.

[0155] A set of bird-repelling actions is constructed based on the sound source output parameters, light source output parameters, and the sound and light triggering time sequence. The reinforcement learning environment state representation vector corresponding to the current time segment in the environmental state representation vector sequence is input into the deep reinforcement learning policy network, and the corresponding sound and light coordinated bird-repelling strategy is output. The sound and light coordinated bird-repelling strategy includes the target parameter combination of the sound source output parameters, light source output parameters, and the sound and light triggering time sequence.

[0156] The deep reinforcement learning policy network is implemented using the PPO model. The policy network is input with the reinforcement learning environment state representation vector, which is passed through three fully connected feature extraction layers and one gated recurrent unit layer to obtain the state representation. Then it is connected to the action output layer. The output dimension of the action output layer is consistent with the number of bird-repelling action icons in the bird-repelling action set. It outputs the selection probability corresponding to each bird-repelling action icon. Based on the maximum probability, the bird-repelling action icon corresponding to the sound-light coordinated bird-repelling strategy is selected. The target parameter combination of sound source output parameters, light source output parameters and sound-light triggering timing is obtained by mapping the bird-repelling action icon.

[0157] In this embodiment, the generation of the optimization results of the sound-light coordinated bird deterrence strategy specifically includes:

[0158] Within the current time segment, the sound source output parameters, light source output parameters, and sound and light triggering sequence are determined based on the sound and light coordinated bird deterrence strategy. The sound source emission direction and the light coverage area are pointed to the spatial region corresponding to the set of key propagation nodes, and sound and light directional stimulation is performed.

[0159] After the audio-visual directional stimulation ends, a set of bird flock state data for the next time segment is generated, and the set of bird flock startled propagation state, the result of the fear information propagation efficiency characterization, and the result of the group escape threshold judgment are updated simultaneously.

[0160] The current state is represented by the reinforcement learning environment state vector corresponding to the current time segment, and the current action is represented by the bird-driving action identifier corresponding to the sound-light coordinated bird-driving strategy. The next state is represented by the reinforcement learning environment state vector corresponding to the next time segment, and a reward value is generated.

[0161] During the reward value generation process, based on the bird flock monitoring data corresponding to the current time segment and the next time segment, the changes in bird flock spatial location, bird flock flight speed, and bird flock density are calculated respectively. Based on the fear information propagation efficiency representation results, the changes in propagation efficiency score and efficiency stability score are calculated, and the critical judgment value for flock escape is obtained. The changes in bird flock spatial location, bird flock flight speed, bird flock density, propagation efficiency score, and efficiency stability score are normalized respectively, and a weighted sum is performed to obtain the reward value.

[0162] The current state, current action, reward value, and next state are combined to form a state transition sample. The state transition samples corresponding to each time segment are then summarized to generate a state transition sample set.

[0163] The parameters of the deep reinforcement learning policy network are updated based on the state transition sample set, and the corresponding acoustic-optical bird deterrence policy is output after the update.

[0164] Based on the state transition sample set, the deep reinforcement learning policy network is subjected to mini-batch sampling and iterative updates. The policy network parameters are updated through gradient backpropagation and the loss is minimized until convergence.

[0165] The process involves iteratively executing audio-visual directional stimulation, reward value calculation, state transition sample set generation, and deep reinforcement learning policy network parameter updates until the optimized result of the audio-visual collaborative bird deterrence strategy is output. The optimized result of the audio-visual collaborative bird deterrence strategy is the optimal parameter combination strategy output by the deep reinforcement learning policy network after convergence, which controls the output parameters of the sound source, the output parameters of the light source, and the timing of the audio-visual triggering.

[0166] A deep reinforcement learning-based acoustic-optical coordinated bird control strategy optimization system includes:

[0167] The bird flock data acquisition module is used to collect bird flock monitoring data and bird control data within the target area, and perform preprocessing to generate a bird flock status data set.

[0168] The bird flock dynamic propagation network construction module is used to extract bird flock behavior characteristics based on bird flock monitoring data in the bird flock status dataset and construct the bird flock dynamic propagation network.

[0169] The startled behavior propagation modeling module is used to model the propagation of bird flock startled response process, calculate startled behavior propagation parameters, and generate a set of bird flock startled propagation states.

[0170] The critical propagation node identification module is used to identify the set of critical propagation nodes in the dynamic propagation network of a flock of birds based on the set of bird flock startled propagation states.

[0171] The Fear Information Efficiency and Critical Judgment Module is used to calculate the information entropy parameter of bird flock behavior, generate the fear information propagation efficiency representation result, and generate the critical judgment result of group escape.

[0172] The Ecological Game Modeling and Adaptation Prediction Module is used to establish the ecological game relationship between the set of bird deterrence strategies and the set of bird flock behavior responses, calculate bird flock behavior adaptation parameters, and generate bird flock behavior adaptation prediction results.

[0173] The state fusion and policy generation module is used to construct the state representation vector of the reinforcement learning environment and input the state representation vector of the reinforcement learning environment into the deep reinforcement learning policy network to generate an acoustic-optical collaborative bird deterrence policy.

[0174] The acoustic-optical directional stimulation and strategy update module is used to implement acoustic-optical directional stimulation on the set of key propagation nodes according to the acoustic-optical coordinated bird repelling strategy, and update the parameters of the deep reinforcement learning strategy network. This process is repeated until the optimized results of the acoustic-optical coordinated bird repelling strategy are output.

[0175] Example 1: To verify the feasibility of this invention in practice, it was applied to a bird control scenario at a coastal airport. The area is near a wetland environment, where large flocks of birds forage, rest, and take off near the runway, posing a safety hazard during peak flight hours. Traditional bird control methods rely mainly on manual patrols and periodic use of fixed-parameter audio-visual equipment. When bird numbers are large or their behavior spreads rapidly, unidirectional or fixed-frequency audio-visual stimulation is insufficient to produce effective bird control in a timely manner. Furthermore, after repeated stimulation, birds gradually adapt, leading to a gradual weakening of the control effect and the birds returning shortly afterward. This invention models the propagation patterns of bird behavior and optimizes bird control strategies using deep reinforcement learning. This allows the bird control system to automatically adjust its audio-visual coordinated control method based on changes in the dynamic behavior of the bird flock, thus solving the problems of unstable control efficiency and high energy consumption in traditional bird control methods.

[0176] In this scenario, bird monitoring units are first deployed on both sides of the airport runway. Multi-source sensors continuously collect information on the spatial location, flight speed, and flock density of the birds, while simultaneously recording the sound source output status, light source output status, and audio-visual trigger timing of the bird deterrent equipment. After acquiring the monitoring data, the system performs time synchronization, spatial registration, and anomaly removal processing, enabling analysis of monitoring information from different sources on a unified time scale. Subsequently, the system constructs a dynamic bird flock propagation network based on changes in the bird flock's spatial location and flight speed. By analyzing the spatial, motion, and behavioral response relationships among individual birds, the system characterizes the propagation path of startled behavior within the flock. Based on this, the system models the propagation process of startled behavior within the network and calculates the behavioral impact and propagation intensity of each individual bird during the propagation process, thereby identifying key propagation nodes that play a crucial role in the diffusion of flock behavior. Once key propagation nodes are identified, the system calculates the relevant propagation probability distribution of these nodes and further calculates the bird flock behavior information entropy parameters to quantify the efficiency of fear information propagation. Simultaneously, the system combines the information entropy change trend to determine whether the flock has entered a critical stage of mass flight. When the system detects an increase in propagation efficiency and approaches the stage of flock flight, it automatically generates a combined acoustic and optical bird-repelling strategy through a reinforcement learning policy network. This strategy prioritizes the direction of sound source emission and the scanning area of ​​the beam towards the spatial region where key propagation nodes are located, thereby amplifying the repelling effect by utilizing the internal behavioral propagation mechanism of the bird flock. As the system continues to operate, the stimuli generated by the bird-repelling equipment are fed back into the system. The system continuously updates the parameters of the reinforcement learning policy network based on changes in the spatial distribution of the bird flock, changes in flight speed, and changes in flock density, enabling the bird-repelling strategy to continuously adapt to changes in bird flock behavior.

[0177] To verify the performance of the present invention, it was compared with the traditional method. The comparison results are shown in Table 1.

[0178] Table 1. Comparison of the overall operational performance of the sound-light coordinated bird deterrence strategy optimization method and traditional bird deterrence methods.

[0179]

[0180] As can be seen from Table 1, the proposed method for optimizing the acoustic-optical bird deterrence strategy based on deep reinforcement learning outperforms the traditional acoustic-optical bird deterrence method in several key performance indicators.

[0181] Regarding the average repelling response time, traditional methods mainly rely on fixed triggering rules or human experience to set bird deterrence device parameters. When a flock of birds enters the area, it often takes multiple rounds of sound and light stimulation to produce a significant repelling effect. However, this invention constructs a dynamic bird flock propagation network and identifies key propagation nodes, enabling the system to quickly locate individuals that have an influence on the group's behavior in the early stages of bird flock behavior propagation and implement targeted stimulation. Therefore, the average repelling response time is significantly shortened. The data in the table shows that the repelling response time has decreased significantly, indicating that the system can trigger the group's diffusion behavior more quickly.

[0182] In terms of bird deterrence effectiveness, both the success rate of a single bird deterrence attempt and the rate of decrease in bird density in key areas showed significant improvements. Traditional bird deterrence methods typically employ regional or random directional stimuli, resulting in low propagation efficiency of sound and light stimuli within the flock. This leads to some individual birds not being effectively affected, thus reducing the overall deterrence effect. This invention models the propagation process of startled behavior in bird flocks and calculates the propagation efficiency of fear information using behavioral information entropy. Targeted stimuli are implemented at the stage when the flock's behavior is about to spread, causing significant behavioral changes at key nodes first. The startled behavior is then rapidly spread to the entire flock through the group propagation mechanism, thereby increasing the deterrence success rate and accelerating the overall departure speed of the flock.

[0183] From the perspective of system operational stability, this invention also has advantages in terms of the rate of secondary regression of bird flocks and the stability index of bird deterrence effect. Traditional bird deterrence systems, due to their single bird deterrence strategy, allow birds to gradually adapt after repeated exposure to sound and light stimuli, resulting in short-term re-regression. This invention establishes an ecological game relationship between the set of bird deterrence strategies and the set of bird flock behavioral responses, statistically analyzes the changes in bird flock behavior caused by different bird deterrence actions, and continuously optimizes bird deterrence parameters using a reinforcement learning strategy network. This enables the system to dynamically adjust the sound source output parameters, light source output parameters, and sound and light triggering sequence according to changes in bird flock behavior, thereby reducing the degree of adaptation of bird flocks to fixed bird deterrence methods and reducing the occurrence of bird flock re-gathering.

[0184] Furthermore, in terms of equipment operating efficiency, this invention also demonstrates significant optimization effects in terms of energy consumption per unit area and the number of equipment triggers. Traditional bird-repelling methods typically require frequent equipment triggers to maintain the bird-repelling effect. However, this invention, by identifying key propagation nodes and providing targeted stimulation, concentrates bird-repelling energy on nodes that have a propagating effect on group behavior, thereby expanding the bird-repelling effect through the group propagation mechanism. Therefore, while reducing the number of equipment triggers, it can still maintain high bird-repelling efficiency, ultimately achieving a synergistic optimization of reduced bird-repelling energy consumption and improved bird-repelling effect. These results indicate that this invention, by introducing bird flock behavior propagation modeling, information entropy propagation efficiency analysis, and deep reinforcement learning strategy optimization mechanisms, effectively solves the problems of low bird-repelling efficiency, high energy consumption, and unstable bird-repelling effect in traditional bird-repelling methods.

[0185] The above are merely preferred embodiments 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 method for optimizing acousto-optical bird-repelling strategies based on deep reinforcement learning, characterized in that, Includes the following steps: Collect bird flock monitoring data and bird control data within the target area, and perform preprocessing to generate a bird flock status data set; Bird flock behavior characteristics were extracted from bird flock monitoring data in the bird flock status dataset, and a dynamic bird flock propagation network was constructed. Based on the dynamic propagation network of bird flocks, the propagation model of the bird flock startled response process is constructed, the propagation parameters of startled behavior are calculated, and a set of bird flock startled propagation states is generated. Based on the set of bird flock startled propagation states, identify the set of key propagation nodes in the dynamic propagation network of bird flocks; Based on the set of bird flock startled propagation states and the set of key propagation nodes, calculate the bird flock behavior information entropy parameters, generate a characterization result of fear information propagation efficiency, and generate a critical judgment result for group escape. Based on the set of bird flock startled propagation states and the results of fear information propagation efficiency characterization, an ecological game relationship is established between the set of bird deterrence strategies and the set of bird flock behavioral responses. Bird flock behavioral adaptation parameters are calculated, and bird flock behavioral adaptation prediction results are generated. By integrating bird flock state data set, key propagation node set, fear information propagation efficiency representation results, group escape criticality judgment results and bird flock behavior adaptation prediction results, a reinforcement learning environment state representation vector is constructed, and an acoustic-light coordinated bird deterrence strategy is generated through a deep reinforcement learning policy network. Based on the sound-light coordinated bird deterrence strategy, sound-light directional stimulation is applied to the key propagation node set, and the parameters of the deep reinforcement learning strategy network are updated. This process is repeated until the optimized result of the sound-light coordinated bird deterrence strategy is output.

2. The method for optimizing the acoustic-optical coordinated bird-repelling strategy based on deep reinforcement learning according to claim 1, characterized in that, The preprocessing includes time synchronization, spatial registration, anomaly removal, and normalization. The bird flock monitoring data includes the bird flock's spatial location, flight speed, and population density. The bird deterrence control data includes sound source output parameters, light source output parameters, and sound and light triggering timing.

3. The method for optimizing the acoustic-optical coordinated bird-repelling strategy based on deep reinforcement learning according to claim 1, characterized in that, The generation of the bird flock dynamic propagation network specifically includes: Read the bird flock status data set, and arrange the bird flock spatial location, bird flock flight speed and bird flock density according to a unified time index to generate a bird flock time-series status data sequence; Bird flock behavior features are extracted from bird flock time-series state data sequences to generate a set of bird flock behavior features; Bird flock behavior characteristics include the spatial distribution of individual birds, the speed distribution of individual birds, and the population density distribution in local areas; Based on the spatial location of each bird within the same time segment, calculate the spatial distance between individual birds, determine the spatial relationships between individual birds, and generate a set of spatial relationships. Based on the flock flight speeds of individual birds within the same time segment, calculate the speed difference and speed direction difference between individual birds, determine the motion correlation between individual birds, and generate a set of motion correlation relationships. Based on the set of bird flock behavior characteristics, calculate the corresponding values ​​of behavioral changes of different individual birds in continuous time segments, determine the response correlation between individual birds, and generate a set of response correlations. By combining the sets of spatial associations, motion associations, and response associations, a bird flock behavior association network is constructed. The node connection relationships and edge weight parameters are updated according to continuous time segments to generate a bird flock dynamic propagation network.

4. The method for optimizing the acoustic-optical coordinated bird-repelling strategy based on deep reinforcement learning according to claim 1, characterized in that, The generation of the set of bird flock startled propagation states specifically includes: Read the node connection relationships, edge weight parameters and bird flock behavior feature sets corresponding to each bird individual in the bird flock dynamic propagation network, and determine the basic state of behavior propagation of each bird individual in the current time segment; Each individual bird is taken as the current propagation source. Based on the node connection relationship and edge weight parameters between the current propagation source and the other individual birds, the set of direct propagation objects corresponding to the current propagation source is determined. Read the set of bird flock behavior characteristics corresponding to the current propagation source and the set of bird flock behavior characteristics corresponding to the set of direct propagation objects, and calculate the intensity of the behavioral influence of the current propagation source on each direct propagation object; Obtain the edge weight parameters corresponding to the current propagation source and each direct propagation object, and determine the propagation intensity of the current propagation source to each direct propagation object according to the intensity of the behavior influence and the corresponding edge weight parameters. Then, summarize the results to generate the propagation parameters of the startled behavior corresponding to each direct propagation object of the current propagation source. The parameters of the startled behavior propagation are repeatedly obtained for each individual bird, and the parameters of the startled behavior propagation for each individual bird are arranged in order of time segments to generate a set of startled behavior propagation states of the bird flock.

5. The method for optimizing the acoustic-optical coordinated bird-repelling strategy based on deep reinforcement learning according to claim 1, characterized in that, The generation of the key propagation node set specifically includes: Read the set of bird flock startled propagation states, extract the startled behavior propagation parameters of each bird in each time segment according to a unified time index, and generate a set of propagation parameter sequences. In the set of propagation parameter sequences, individual birds whose startled behavior propagation parameters are greater than or equal to a preset propagation intensity screening threshold are selected, a set of candidate key nodes is generated, and the propagation parameter sequence of startled behavior corresponding to each individual bird in the set of candidate key nodes is recorded synchronously. For each individual bird in the candidate key node set, read the node connection relationship and edge weight parameters corresponding to that individual bird in the bird flock dynamic propagation network, count the number of adjacent nodes to obtain the node connectivity parameter, and sum the adjacent edge weight parameters to obtain the node edge weight summary value. The cumulative value of the propagation contribution is calculated for the propagation parameter sequence of startled behavior corresponding to each individual bird, and the difference of the cumulative value of the propagation contribution is calculated to generate the propagation persistence parameter; The cumulative value of propagation contribution, propagation persistence parameter, node connectivity parameter, and node edge weight summation value are merged to generate a comprehensive score value for key nodes. The candidate key node set is then sorted based on the comprehensive score value to generate a key node ranking result. Extract the set of key propagation nodes from the key node sorting results.

6. The method for optimizing the acoustic-optical coordinated bird-repelling strategy based on deep reinforcement learning according to claim 1, characterized in that, The generation of the group escape criticality determination result specifically includes: Obtain the set of bird flock startled propagation states, extract the startled behavior propagation parameters of each node pair in each time segment according to a unified time index, and filter the set of node pairs containing key propagation nodes based on the set of key propagation nodes to generate the set of key node propagation parameters. The propagation parameter set of key nodes is summarized and statistically analyzed in each time segment to generate the total propagation value corresponding to that time segment. Based on the total propagation value, the propagation parameters of each node for startled behavior are proportionalized to generate the propagation probability distribution of key nodes corresponding to that time segment. The key node propagation probability distributions corresponding to each time segment are arranged in time index order to generate a key node propagation probability sequence. Based on the key node propagation probability sequence, the information entropy parameter of bird flock behavior is calculated for each time segment to generate an information entropy sequence. The difference between the information entropy parameters of adjacent time segments in the information entropy sequence is calculated to generate an information entropy change sequence. Based on the information entropy sequence and the information entropy change sequence, a characterization result of the fear information propagation efficiency is generated. The results of the fear information dissemination efficiency characterization include a dissemination efficiency score sequence and an efficiency stability score sequence; Based on the results of the fear information propagation efficiency characterization, the critical trigger time segment is determined, and the critical judgment result of the group escape is generated; Arrange the critical trigger time segment index and the critical judgment flag value in a corresponding manner, and output the critical judgment result of the group escaping.

7. The method for optimizing the acoustic-optical coordinated bird-repelling strategy based on deep reinforcement learning according to claim 1, characterized in that, The generation of the bird flock behavior adaptation prediction results specifically includes: Organize the spatial location, flight speed, and population density of bird flocks according to a unified time index to form a sequence of bird flock status characteristics, and perform differential calculations on adjacent time segments to obtain the changes in bird flock spatial location, flight speed, and population density. Organize the propagation parameters of startled behavior of each node according to a unified time index, and use the set of key propagation nodes to filter the node pairs containing key propagation nodes to form the key node startled propagation feature sequence. In each time segment, sum the key node startled propagation feature sequence to obtain the propagation intensity summary value, and perform difference operation on the propagation intensity summary values ​​of adjacent time segments to obtain the propagation intensity change. The difference between the propagation efficiency score sequence and the efficiency stability score sequence is used to obtain the change in propagation efficiency and the change in efficiency stability. These changes are then aligned with the changes in bird flock spatial location, bird flock flight speed, bird flock population density, and propagation intensity by time index to form a game state sequence. The bird control data is discretized and encoded to generate bird deterrence action identifiers for each discretized combination, forming a set of bird deterrence action identifiers; Within each time segment, the bird deterrence action identifier of that time segment is matched one-to-one with the changes in the spatial location of the bird flock, the changes in the flight speed of the bird flock, the changes in the population density of the bird flock, the changes in the propagation intensity, the changes in the propagation efficiency, and the changes in the efficiency stability of the bird flock, and the bird flock behavior response set is generated by summarizing them. An ecological game relationship is established based on the game state sequence, the set of bird deterrence action identifiers, and the set of bird flock behavior responses. The set of bird flock behavior responses corresponding to the bird deterrence action identifiers is statistically analyzed and bird flock behavior adaptation parameters are generated. Based on the bird flock behavior adaptation parameters, bird flock behavior adaptation prediction results are generated.

8. The method for optimizing the acoustic-optical coordinated bird-repelling strategy based on deep reinforcement learning according to claim 1, characterized in that, The generation of the acoustic-optical coordinated bird-repelling strategy specifically includes: Time alignment is performed on the spatial location, flight speed, and population density of bird flocks in the bird flock state dataset to generate a bird flock state vector. Perform set representation on the set of key propagation nodes to generate key node representation vectors; The propagation efficiency score and efficiency stability score in the performance representation of fear information propagation efficiency are combined to generate an efficiency representation vector. Encode the critical decision identifier value and the critical trigger time segment index in the group escape critical decision result to generate a critical decision vector; The optimal response bird-driving action identifier in the bird flock behavior adaptation prediction result is encoded to generate an action identifier vector, and then concatenated with the corresponding response mean vector to generate an adaptation prediction vector. Within each time segment, the flock state vector, key node representation vector, efficiency representation vector, critical decision vector, and adaptive prediction vector are respectively dimensionally aligned and normalized, and then spliced ​​and fused to generate reinforcement learning environment state representation vectors, which are arranged in time index order to generate a sequence of environment state representation vectors. A set of bird-repelling actions is constructed based on the sound source output parameters, light source output parameters, and the audio-visual triggering time sequence. The reinforcement learning environment state representation vector corresponding to the current time segment in the environmental state representation vector sequence is input into the deep reinforcement learning policy network, and the corresponding audio-visual collaborative bird-repelling policy is output.

9. The method for optimizing acousto-optical bird-repelling strategies based on deep reinforcement learning according to claim 1, characterized in that, The generation of the optimized results of the acoustic-optical coordinated bird deterrence strategy specifically includes: Within the current time segment, the sound source output parameters, light source output parameters, and sound and light triggering sequence are determined based on the sound and light coordinated bird deterrence strategy. The sound source emission direction and the light coverage area are pointed to the spatial region corresponding to the set of key propagation nodes, and sound and light directional stimulation is performed. After the audio-visual directional stimulation ends, a set of bird flock state data for the next time segment is generated, and the set of bird flock startled propagation state, the result of the fear information propagation efficiency characterization, and the result of the group escape threshold judgment are updated simultaneously. The current state is represented by the reinforcement learning environment state vector corresponding to the current time segment, and the current action is represented by the bird-driving action identifier corresponding to the sound-light coordinated bird-driving strategy. The next state is represented by the reinforcement learning environment state vector corresponding to the next time segment, and a reward value is generated. The current state, current action, reward value, and next state are combined to form a state transition sample. The state transition samples corresponding to each time segment are then summarized to generate a state transition sample set. The parameters of the deep reinforcement learning policy network are updated based on the state transition sample set, and the corresponding acoustic-optical bird deterrence policy is output after the update. The process involves iteratively executing audio-visual directional stimulation, reward value calculation, state transition sample set generation, and deep reinforcement learning strategy network parameter updates until the optimized result of the audio-visual collaborative bird deterrence strategy is output.

10. A deep reinforcement learning-based acoustic-optical coordinated bird-repelling strategy optimization system, comprising executing the deep reinforcement learning-based acoustic-optical coordinated bird-repelling strategy optimization method according to any one of claims 1 to 9, characterized in that, include: The bird flock data acquisition module is used to collect bird flock monitoring data and bird control data within the target area, and perform preprocessing to generate a bird flock status data set. The bird flock dynamic propagation network construction module is used to extract bird flock behavior characteristics based on bird flock monitoring data in the bird flock status dataset and construct the bird flock dynamic propagation network. The startled behavior propagation modeling module is used to model the propagation of bird flock startled response process, calculate startled behavior propagation parameters, and generate a set of bird flock startled propagation states. The critical propagation node identification module is used to identify the set of critical propagation nodes in the dynamic propagation network of a flock of birds based on the set of bird flock startled propagation states. The Fear Information Efficiency and Critical Judgment Module is used to calculate the information entropy parameter of bird flock behavior, generate the fear information propagation efficiency representation result, and generate the critical judgment result of group escape. The Ecological Game Modeling and Adaptation Prediction Module is used to establish the ecological game relationship between the set of bird deterrence strategies and the set of bird flock behavior responses, calculate bird flock behavior adaptation parameters, and generate bird flock behavior adaptation prediction results. The state fusion and policy generation module is used to construct the state representation vector of the reinforcement learning environment and input the state representation vector of the reinforcement learning environment into the deep reinforcement learning policy network to generate an acoustic-optical collaborative bird deterrence policy. The acoustic-optical directional stimulation and strategy update module is used to implement acoustic-optical directional stimulation on the set of key propagation nodes according to the acoustic-optical coordinated bird repelling strategy, and update the parameters of the deep reinforcement learning strategy network. This process is repeated until the optimized results of the acoustic-optical coordinated bird repelling strategy are output.

Images