Bird intelligent observation and identification method based on reinforcement learning

Abstract

The application discloses a kind of bird intelligent observation identification methods based on reinforcement learning, specifically includes: step 1: image stream and acoustic stream are collected, generate observation unit identification set and adjacent matrix;Step 2: in time index window, generate class identification vector, behavior identification vector, write observation event sequence table;Step 3: establish HDP parameter table;Step 4: adjacent observation unit pair is executed graph coupling sampling update, write back theme mixing proportion vector table;Step 5: generate long-term ecological benefit posterior distribution parameter table;Step 6: according to group level thompson sampling, generate observation path decision sequence;Step 7: according to decision sequence, perform observation, write back observation event sequence table.The application realizes observation path adaptive optimization and collaborative decision distribution, improves coverage efficiency and scheduling stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent bird observation and recognition technology, and in particular to an intelligent bird observation and recognition method based on reinforcement learning. Background Technology

[0002] In the field of ecological and environmental monitoring, birds are important indicator species of ecosystems, and their population size, activity frequency, and behavioral patterns reflect the stability of the regional ecological structure. Traditional bird observation methods rely on manual patrols and recording, with observers collecting samples along fixed routes or within experienced areas. This approach suffers from limited observation coverage, uneven time distribution, and significant differences in subjective judgment. With the increasing deployment of automated equipment, fixed cameras, mobile observation platforms, and acoustic sensors are gradually being used for field data acquisition, enabling continuous acquisition of image and acoustic streams. However, a lack of collaborative scheduling mechanisms among numerous observation units remains a significant challenge.

[0003] Existing technologies typically employ static inspection paths or greedy selection strategies based on single-point identification probabilities to visit observation units point by point. This approach does not consider the spatial adjacency relationships between observation units, nor does it perform hierarchical statistical modeling of historical observation information across multiple units, resulting in low efficiency in updating topic distribution and insufficient sample utilization. In scenarios with multiple observation areas, some areas are repeatedly visited, while areas with potential ecological value are not covered in a timely manner.

[0004] Meanwhile, existing benefit assessment methods mostly employ fixed weights or simple mean calculations, failing to establish long-term benefit distribution models and thus unable to dynamically balance uncertainty and expected benefits. This leads to randomness and instability in observation path selection. Therefore, constructing a hierarchical topic model that integrates spatial adjacency relationships, followed by generating observation path decision sequences using a posterior distribution sampling mechanism, has become a pressing issue in the field of intelligent bird observation and recognition. Summary of the Invention

[0005] One objective of this invention is to propose a bird intelligent observation and identification method based on reinforcement learning. This invention introduces a graph-coupled hierarchical Dirichlet process parameter table structure and a group-level Thompson sampling decision mechanism. It performs spatial adjacency constraint updates and posterior distribution sampling calculations on the observation event sequences of multiple observation units, constructs a continuously iterative observation path optimization process, realizes joint control of topic statistical modeling and payoff uncertainty, and forms a continuously adaptive observation scheduling structure with the characteristics of coherent path generation, full sample utilization, and stable decision-making process.

[0006] A bird intelligent observation and identification method based on reinforcement learning according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Acquire the observation image stream and acoustic stream, generate the observation unit identifier set, read the spatial coordinates of the observation units, and calculate the adjacency matrix of the observation units;

[0008] Step 2: Generate category identifier vectors and behavior identifier vectors within the time index window and write them into the observation event sequence table;

[0009] Step 3: Establish the Hierarchical Dirichlet Process (HDP) parameter table;

[0010] Step 4: Select adjacent observation unit pairs according to the adjacency matrix, read the topic mixing ratio vector and topic counting vector, perform component coupling calculation to generate coupling ratio vector, combine topic matching count to generate topic sampling score vector, perform random sampling after normalization to generate topic identifier, update the topic counting table and global topic parameter table, and then write back to the topic mixing ratio vector table;

[0011] Step 5: Calculate the posterior distribution parameter table of long-term ecological benefits;

[0012] Step 6: Establish an observation unit group table, perform Thompson sampling on the group posterior distribution and the observation unit posterior distribution, generate group sampling payoff value and unit sampling payoff value, and generate observation path decision sequence according to the sampling payoff value;

[0013] Step 7: Execute the observations according to the observation path decision sequence and then write them into the observation event sequence table.

[0014] Optionally, step 1 specifically includes:

[0015] Acquire observation image streams and acoustic streams; perform frame extraction on the observation image streams to generate image frame sequences; and perform frame segmentation on the acoustic streams to generate acoustic frame sequences.

[0016] Write timestamps to the image frame sequence and the acoustic frame sequence to generate a unified time index sequence; read the spatial coordinate table of the observation unit and perform Euclidean distance calculation on the spatial coordinates to generate the distance matrix of the observation unit;

[0017] A threshold comparison is performed on the distance matrix of the observation units to generate an adjacency marker matrix. A symmetric verification is performed on the adjacency marker matrix to generate an adjacency matrix of the observation units. An observation unit time index mapping table is constructed according to the observation unit identifier set and the unified time index sequence, and then written to the data buffer.

[0018] Optionally, step 2 specifically includes:

[0019] Select a time index window within a unified time index sequence, and read the image frame sequence and acoustic frame sequence corresponding to the time index window;

[0020] Perform forward inference operations on the image frame sequence using a convolutional neural network to generate bird category probability vectors. Then, perform maximum probability index reading operations on the bird category probability vectors to generate bird category identification vectors.

[0021] Perform Mel-frequency transform on the acoustic frame sequence to generate a spectrum matrix, perform forward inference on the spectrum matrix using a recurrent neural network to generate a behavior probability vector, and perform maximum probability index reading on the behavior probability vector to generate a behavior identifier vector;

[0022] Perform time index alignment operation on the bird category identifier vector and behavior identifier vector to generate category identifier vector table and behavior identifier vector table, and write the category identifier vector table and behavior identifier vector table into the observation event sequence table in time index order.

[0023] Optionally, step 3 specifically includes:

[0024] Read the set of observation unit identifiers and the sequence table of observation events, perform grouping and slicing operations on the sequence table of observation events according to the observation unit identifiers, and generate a set of sub-tables of observation unit events.

[0025] Extract the category identifier vector table and behavior identifier vector table from the event sub-table set of the observation unit, and perform concatenation operation on the category identifier vector and behavior identifier vector in time index order to generate event identifier sequence;

[0026] Create a topic index table and then write the initial topic identifiers. Create a global topic parameter table and then write the category component count vector and behavior component count vector for each topic identifier.

[0027] Establish a subject-mixing ratio vector table for observation units, and then write initial values ​​for the subject-mixing ratio vector for each observation unit identifier; establish a subject counting table, and then write initial values ​​for the subject counting vector for each observation unit identifier. The subject counting vector contains the counting positions consistent with the subject index table.

[0028] The event identifier sequence is counted and accumulated according to the topic index table. The count accumulation result is written into the global topic parameter table and the topic count table. The initial value of the topic mixing ratio vector is normalized and then written into the topic mixing ratio vector table of the observation unit to form the hierarchical Dirichlet process HDP parameter table.

[0029] Optionally, step 4 specifically includes:

[0030] Read the observation unit adjacency matrix, observation unit topic mixing ratio vector table, topic count table, global topic parameter table, and topic index table; generate a set of adjacent observation unit pairs according to the position of the adjacency marker value of 1 in the observation unit adjacency matrix;

[0031] Perform sequential traversal of the set of adjacent observation units, read the event sub-table set of the corresponding observation unit of the adjacent observation unit, extract the event identifier sequence, establish the event topic identifier cache table, and write the initial topic identifier into the event topic identifier cache table according to the time index position.

[0032] For each time index position within the event identifier sequence, read the event identifier vector. Traverse the topic identifier set according to the topic index table. Read the category component count vector and behavior component count vector corresponding to the topic identifier in the global topic parameter table. Read the count according to the event identifier vector position and then perform summation to generate the event matching count value vector. Read the topic mixing ratio vector of the first observation unit and the topic mixing ratio vector of the second observation unit in the adjacent observation unit pair. Perform coupling coefficient product and component addition on the two topic mixing ratio vectors to generate the coupling ratio vector.

[0033] The coupling ratio vector and the event matching count vector are multiplied by components to generate a topic sampling score vector. The topic sampling score vector is summed and normalized by components to generate a topic sampling probability vector. Random sampling is performed on the topic sampling probability vector to obtain the topic sampling identifier.

[0034] Read the original topic identifier at the current time index position from the event topic identifier cache table, perform deduction update and write update on the topic count table, perform deduction update and write update on the global topic parameter table, and write the topic sampling identifier to the event topic identifier cache table;

[0035] After the traversal is complete, normalization updates are performed on the topic count vectors of the first and second observation units, the topic mixing ratio vector table of the observation units is written back, and the topic mixing ratio vector table of spatial adjacency constraints is output.

[0036] Optionally, step 5 specifically includes:

[0037] Read the spatial adjacency constraint topic mixing ratio vector table and the global topic parameter table, and read the topic mixing ratio vector one by one according to the observation unit identifier;

[0038] Traverse the set of topic identifiers in the topic index table, read the corresponding category component count vector and behavior component count vector for each topic identifier, and calculate the topic statistical weight value.

[0039] The topic mixing ratio vector and the topic statistical weight value are multiplied together to generate the topic return component vector.

[0040] Perform component summation on the theme benefit component vector to generate the benefit statistics of the observation unit; perform variance calculation on the benefit statistics of the observation unit and the theme count vector to generate the uncertainty statistics; write the benefit statistics into the expected benefit component according to the observation unit identifier, write the uncertainty statistics into the uncertainty component, and generate the long-term ecological benefit posterior distribution parameter table of the observation unit.

[0041] Optionally, step 6 specifically includes:

[0042] Read the long-term ecological benefit posterior distribution parameter table and the observation unit identifier set, perform a partitioning operation on the observation unit identifier set according to the preset grouping rules, and generate an observation unit group table; establish a group posterior parameter cache for each group identifier, the group posterior parameter is obtained by summing the expected benefit components of the observation units in the group to obtain the group expected benefit value, and by summing the uncertainty components of the observation units in the group to obtain the group uncertainty value; establish a group prior parameter table and then write the initial values ​​of the prior distribution parameters;

[0043] For each group identifier in the group posterior parameter buffer, read the expected group payoff value and the group uncertainty value, write the expected group payoff value and the group uncertainty value into the sampling parameter vector, and perform random number generation operation according to the Thompson sampling rule to generate the group sampling payoff value; after all group sampling payoff values ​​have been generated, perform sorting operation according to the group sampling payoff values ​​to generate the group priority sequence.

[0044] Within the group priority sequence, read the group identifiers sequentially, enter the observation unit level within the group, read the corresponding expected return component and uncertainty component for each observation unit identifier within the group, construct the unit sampling parameter vector, and perform Thompson sampling to generate the unit sampling return value; perform sorting operation on the sampling return values ​​of the observation units within the group to generate the unit priority sequence.

[0045] The group priority sequence and the corresponding unit priority sequence are concatenated hierarchically to generate the observation path decision sequence, and then the observation path decision sequence is written into the path decision cache.

[0046] Optionally, step 6, performing Thompson sampling, specifically includes:

[0047] Read the expected benefit component and uncertainty component corresponding to the target observation unit identifier in the long-term ecological benefit posterior distribution parameter table, and construct the unit sampling parameter vector, which contains the expected benefit value and the uncertainty value; read the random number seed and then generate a uniform random number u in the interval (0,1); perform standard normal distribution and inverse cumulative distribution calculation on the uniform random number u to obtain the standard sampling value z; perform scaling and translation operations on the standard sampling value z to generate the unit sampling benefit value, with the scaling factor being the uncertainty value and the translation term being the expected benefit value; write the unit sampling benefit value and the target observation unit identifier into the unit sampling benefit value table.

[0048] Optionally, step 7 specifically includes:

[0049] Read the observation path decision sequence in the path decision buffer, and read the observation unit identifier one by one in the order of the observation path decision sequence;

[0050] Locate the current time index position within the unified time index sequence, and call the observation unit time index mapping table to obtain the corresponding image frame sequence and acoustic frame sequence; perform convolutional neural network forward inference on the image frame sequence to generate category label vectors, and perform spectral transformation and recurrent neural network forward inference on the acoustic frame sequence to generate behavior label vectors;

[0051] Perform time index writing operations on category identifier vectors and behavior identifier vectors to form observation result record items; write the observation result record items into the observation event sequence table in time index order; perform topic identifier update operations and benefit statistics writing operations on newly added records in the observation event sequence table, and write the update results into the topic count table and the long-term ecological benefit posterior distribution parameter table; after completing the traversal of the observation path decision sequence, output the bird identification result table.

[0052] The beneficial effects of this invention are:

[0053] (1) This invention introduces a graph-coupled hierarchical Dirichlet process parameter update structure, incorporates the adjacency matrix information of the observation unit into the topic mixing ratio vector calculation process, so that the topic distribution between adjacent observation units can be updated collaboratively, avoids statistical bias caused by isolated modeling, and improves the structural consistency and statistical stability of multi-unit data joint modeling.

[0054] (2) This invention constructs a long-term ecological benefit posterior distribution parameter table, incorporates the expected benefit component and the uncertainty component into a unified statistical framework, introduces a posterior distribution sampling mechanism in the observation path decision-making process, and makes the path generation process consider both the benefit value and the fluctuation range, so as to achieve a balance control of benefits and risks.

[0055] (3) The present invention uses the group-level Thompson sampling mechanism to perform sampling and sorting operations at the group level and the observation unit level respectively, forming a hierarchical decision structure. It can achieve adaptive path update in multi-observation unit scenarios, improve the rationality of observation resource allocation and the overall scheduling consistency. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of a bird intelligent observation and recognition method based on reinforcement learning proposed in this invention.

[0058] Figure 2 This is a schematic diagram of the graph-coupled HDP update process of a bird intelligent observation and recognition method based on reinforcement learning proposed in this invention.

[0059] Figure 3 This is a schematic diagram of the group-level Thompson sampling decision process of a bird intelligent observation and recognition method based on reinforcement learning proposed in this invention. Detailed Implementation

[0060] 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.

[0061] refer to Figures 1-3 A bird intelligent observation and recognition method based on reinforcement learning includes the following steps:

[0062] Step 1: Acquire the observation image stream and acoustic stream, generate the observation unit identifier set, read the spatial coordinates of the observation units, and calculate the adjacency matrix of the observation units;

[0063] Step 2: Generate category identifier vectors and behavior identifier vectors within the time index window and write them into the observation event sequence table; Step 3: Establish the hierarchical Dirichlet process (HDP) parameter table;

[0064] Step 4: Select adjacent observation unit pairs according to the adjacency matrix, read the topic mixing ratio vector and topic counting vector, perform component coupling calculation to generate coupling ratio vector, combine topic matching count to generate topic sampling score vector, perform random sampling after normalization to generate topic identifier, update the topic counting table and global topic parameter table, and then write back to the topic mixing ratio vector table;

[0065] Step 5: Calculate the posterior distribution parameter table of long-term ecological benefits;

[0066] Step 6: Establish an observation unit group table, perform Thompson sampling on the group posterior distribution and the observation unit posterior distribution, generate group sampling payoff value and unit sampling payoff value, and generate observation path decision sequence according to the sampling payoff value;

[0067] Step 7: Execute the observations according to the observation path decision sequence and then write them into the observation event sequence table.

[0068] In this embodiment, step 1 specifically includes:

[0069] Acquire observation image streams and acoustic streams; perform frame extraction on the observation image streams to generate image frame sequences; and perform frame segmentation on the acoustic streams to generate acoustic frame sequences.

[0070] Write timestamps to the image frame sequence and the acoustic frame sequence to generate a unified time index sequence; read the spatial coordinate table of the observation unit and perform Euclidean distance calculation on the spatial coordinates to generate the distance matrix of the observation unit;

[0071] A threshold comparison is performed on the distance matrix of the observation units to generate an adjacency marker matrix. A symmetric verification is performed on the adjacency marker matrix to generate an adjacency matrix of the observation units. An observation unit time index mapping table is constructed according to the observation unit identifier set and the unified time index sequence, and then written to the data buffer.

[0072] In this embodiment, the execution threshold comparison in step 1 specifically includes:

[0073] Read all distance values ​​from the observation unit distance matrix, then set a preset distance threshold parameter, which is a fixed value. Subsequently, perform a comparison operation on each distance value in the observation unit distance matrix in turn. When the distance value between two observation units is less than or equal to the preset distance threshold parameter, write the value 1 at the corresponding position in the adjacency marker matrix; when the distance value is greater than the preset distance threshold parameter, write the value 0 at the corresponding position in the adjacency marker matrix.

[0074] After comparing all distance values ​​item by item, a self-correlation correction is performed on the adjacency marker matrix, uniformly writing the value of zero to the position corresponding to each observation unit, thus avoiding the formation of self-adjacency relationships. Subsequently, a symmetric consistency check is performed on the adjacency marker matrix; when a pair of observation units has inconsistent values ​​at two corresponding positions in the matrix, the larger marker value is uniformly written.

[0075] In this embodiment, step 1, performing the symmetry verification, specifically includes:

[0076] The label values ​​in the adjacency label matrix are read item by item in the matrix row and column order. For any two symmetrical positions corresponding to two different observation units, the label values ​​at these two positions are read respectively, and then a consistency comparison operation is performed on the two label values.

[0077] When the marked values ​​at two symmetrical positions are the same, the original values ​​remain unchanged; when the marked values ​​at two symmetrical positions differ, a symmetry correction process is performed on the observation unit, and the two positions are uniformly written with the value 1, so as to ensure that as long as either side meets the adjacency condition, the observation unit is considered to be adjacent to the whole.

[0078] After completing the symmetry consistency correction for all observation unit pairs, the data is rewritten into the matrix structure to form a symmetric observation unit adjacency matrix.

[0079] In this embodiment, step 2 specifically includes:

[0080] Select a time index window within a unified time index sequence, and read the image frame sequence and acoustic frame sequence corresponding to the time index window;

[0081] Perform forward inference operations on the image frame sequence using a convolutional neural network to generate bird category probability vectors. Then, perform maximum probability index reading operations on the bird category probability vectors to generate bird category identification vectors.

[0082] Perform Mel-frequency transform on the acoustic frame sequence to generate a spectrum matrix, perform forward inference on the spectrum matrix using a recurrent neural network to generate a behavior probability vector, and perform maximum probability index reading on the behavior probability vector to generate a behavior identifier vector;

[0083] Perform time index alignment operation on the bird category identifier vector and behavior identifier vector to generate category identifier vector table and behavior identifier vector table, and write the category identifier vector table and behavior identifier vector table into the observation event sequence table in time index order.

[0084] In this embodiment, step 2, the execution time index alignment operation, specifically includes:

[0085] Read the unified time index sequence, and then build a time position table according to the time index order. Then, read the timestamp information corresponding to the category identifier vector and the timestamp information corresponding to the behavior identifier vector respectively, and map the two types of timestamp information to the corresponding positions in the unified time index sequence.

[0086] For a given time index position, if both the category identifier vector and the behavior identifier vector have corresponding data, then the time index position is directly written to the category identifier vector and the behavior identifier vector; if only the category identifier vector or only the behavior identifier vector exists, then a null value is written at the missing position, and the time position remains unchanged in subsequent processing; if neither of the two vectors has data at a given time index position, then the time index position is skipped and no record is written.

[0087] After completing the traversal of all time index positions, the category identifier vectors are arranged in a uniform time index order to generate a category identifier vector table, and the behavior identifier vectors are arranged to generate a behavior identifier vector table.

[0088] In this embodiment, step 3 specifically includes:

[0089] Read the set of observation unit identifiers and the sequence table of observation events, perform grouping and slicing operations on the sequence table of observation events according to the observation unit identifiers, and generate a set of sub-tables of observation unit events.

[0090] Extract the category identifier vector table and behavior identifier vector table from the event sub-table set of the observation unit, and perform concatenation operation on the category identifier vector and behavior identifier vector in time index order to generate event identifier sequence;

[0091] Create a topic index table and then write the initial topic identifiers. Create a global topic parameter table and then write the category component count vector and behavior component count vector for each topic identifier.

[0092] Establish a subject-mixing ratio vector table for observation units, and then write initial values ​​for the subject-mixing ratio vector for each observation unit identifier; establish a subject counting table, and then write initial values ​​for the subject counting vector for each observation unit identifier. The subject counting vector contains the counting positions consistent with the subject index table.

[0093] The event identifier sequence is counted and accumulated according to the topic index table. The count accumulation result is written into the global topic parameter table and the topic count table. The initial value of the topic mixing ratio vector is normalized and then written into the topic mixing ratio vector table of the observation unit to form the hierarchical Dirichlet process HDP parameter table.

[0094] In this embodiment, step 3, establishing the topic index table, specifically includes:

[0095] The number of event identifier vectors in the observed event sequence table is counted, and different combinations of category identifiers and behavior identifiers are considered as different event types. Then, deduplication is performed on all event types to generate a set of event types.

[0096] Based on the event type set, an initial set of topic identifiers is generated according to preset topic initialization rules. The preset topic initialization rules can use a fixed-quantity initialization method or a method that generates identifiers proportionally to the number of event types. When using the fixed-quantity initialization method, a corresponding number of topic identifiers are generated based on the preset number of topics; when using the proportional generation method, the number of topics is calculated based on the size of the event type set, and then a corresponding number of topic identifiers are generated.

[0097] After determining the number of topics, a unique topic identifier is written for each topic in the order of generation. Then, all topic identifiers are written sequentially into the topic index table. The topic index table is a sequential structure table, with each row recording the position of a topic identifier.

[0098] After the topic index table is established, each topic identifier in the topic index table is read, and a corresponding category component count vector and behavior component count vector are established for it in the global topic parameter table. Then, an initial count value is written for each count position.

[0099] In this embodiment, step 3, establishing the observation unit subject-specific hybrid ratio vector table, specifically includes:

[0100] Read the observation unit identifier set and the topic index table to determine the number of observation units and the number of topics. Then, create a topic mixing ratio record row for each observation unit identifier. The length of this record row is the same as the number of topics, and each position corresponds to a topic identifier in the topic index table.

[0101] During the initialization phase, an initial ratio value is written to the topic mixing ratio record row for each observation unit. There are two ways to generate the initial ratio value: one is uniform initialization, which writes the same value to all topic positions; the other is a ratio initialization based on the statistical results of the observation unit event sub-table, which counts the number of counts corresponding to each topic in the event identifier sequence of the observation unit, and then writes the counts to the corresponding topic positions after proportionally converting them according to the total count.

[0102] After writing the proportions of all observation units, normalization is performed on each subject mixed proportion record line to keep the sum of all proportion values ​​in each record line a fixed value.

[0103] After normalization, the subject mixing ratio record rows of all observation units are written into the subject mixing ratio vector table of observation units in the order of observation unit identifiers, forming a ratio structure that corresponds one-to-one with the subject index table.

[0104] In this embodiment, step 4 specifically includes:

[0105] Read the observation unit adjacency matrix, observation unit topic mixing ratio vector table, topic count table, global topic parameter table, and topic index table; generate a set of adjacent observation unit pairs according to the position of the adjacency marker value of 1 in the observation unit adjacency matrix;

[0106] Perform sequential traversal of the set of adjacent observation units, read the event sub-table set of the corresponding observation unit of the adjacent observation unit, extract the event identifier sequence, establish the event topic identifier cache table, and write the initial topic identifier into the event topic identifier cache table according to the time index position.

[0107] For each time index position within the event identifier sequence, read the event identifier vector. Traverse the topic identifier set according to the topic index table. Read the category component count vector and behavior component count vector corresponding to the topic identifier in the global topic parameter table. Read the count according to the event identifier vector position and then perform summation to generate the event matching count value vector. Read the topic mixing ratio vector of the first observation unit and the topic mixing ratio vector of the second observation unit in the adjacent observation unit pair. Perform coupling coefficient product and component addition on the two topic mixing ratio vectors to generate the coupling ratio vector.

[0108] The coupling ratio vector and the event matching count vector are multiplied by components to generate a topic sampling score vector. The topic sampling score vector is summed and normalized by components to generate a topic sampling probability vector. Random sampling is performed on the topic sampling probability vector to obtain the topic sampling identifier.

[0109] Read the original topic identifier at the current time index position from the event topic identifier cache table, perform deduction update and write update on the topic count table, perform deduction update and write update on the global topic parameter table, and write the topic sampling identifier to the event topic identifier cache table;

[0110] After the traversal is complete, normalization updates are performed on the topic count vectors of the first and second observation units, the topic mixing ratio vector table of the observation units is written back, and the topic mixing ratio vector table of spatial adjacency constraints is output.

[0111] In this embodiment, step 4, establishing the event topic identifier cache table, specifically includes:

[0112] The time index sequence of the event identifiers is read to determine the number of events in the current time window for that observation unit. Then, a cache record row is created for each event location according to the time index sequence. Each cache record row contains the event time index location and the topic identifier field.

[0113] During the initialization phase, initial topic identifiers are written to all event locations. These initial topic identifiers are generated by reading the topic identifier set from the topic index table and then assigning them to each event location according to preset initialization rules. Initialization rules include either a uniform rotation allocation method or an allocation method based on the event match count.

[0114] After writing the topic identifiers for all event locations, all record rows are written to the event topic identifier cache table in chronological order. The event topic identifier cache table uses a sequential structure, with each row corresponding to an event location and recording the current topic allocation result.

[0115] In the subsequent coupled sampling update process, the topic identifier at the corresponding time index position in the cache table is deducted and written to update the cache table, and then the time index order is kept unchanged.

[0116] In this embodiment, step 4, performing component multiplication, specifically includes:

[0117] When generating the topic sampling score vector, the coupling ratio vector and the event matching count vector are matched according to the topic index order. First, it is confirmed that the number of components in the two vectors is the same and the order of the components is consistent with the topic index table. Then, the two values ​​at the corresponding topic positions are read one by one according to the component positions. The coupling ratio value and the event matching count value at each topic position are multiplied item by item, and the product result is written into the same topic position to form a new numerical sequence.

[0118] The above-described item-by-item multiplication process does not perform cross-position operations or summation. Each topic position only has a calculation relationship with its corresponding position, ensuring that topics do not overlap.

[0119] After all topic locations have been calculated, the resulting set of products constitutes the topic sampling score vector. Each component in this vector simultaneously reflects the spatial adjacency coupling strength and the degree of matching with the current event, thus providing a basic numerical basis for subsequent probability normalization and random sampling.

[0120] In this embodiment, step 5 specifically includes:

[0121] Read the spatial adjacency constraint topic mixing ratio vector table and the global topic parameter table, and read the topic mixing ratio vector one by one according to the observation unit identifier;

[0122] Traverse the set of topic identifiers in the topic index table, read the corresponding category component count vector and behavior component count vector for each topic identifier, and calculate the topic statistical weight value.

[0123] The topic mixing ratio vector and the topic statistical weight value are multiplied together to generate the topic return component vector.

[0124] Perform component summation on the theme benefit component vector to generate the benefit statistics of the observation unit; perform variance calculation on the benefit statistics of the observation unit and the theme count vector to generate the uncertainty statistics; write the benefit statistics into the expected benefit component according to the observation unit identifier, write the uncertainty statistics into the uncertainty component, and generate the long-term ecological benefit posterior distribution parameter table of the observation unit.

[0125] In this embodiment, step 6 specifically includes:

[0126] Read the long-term ecological benefit posterior distribution parameter table and the observation unit identifier set, perform a partitioning operation on the observation unit identifier set according to the preset grouping rules, and generate an observation unit group table; establish a group posterior parameter cache for each group identifier, the group posterior parameter is obtained by summing the expected benefit components of the observation units in the group to obtain the group expected benefit value, and by summing the uncertainty components of the observation units in the group to obtain the group uncertainty value; establish a group prior parameter table and then write the initial values ​​of the prior distribution parameters;

[0127] For each group identifier in the group posterior parameter buffer, read the expected group payoff value and the group uncertainty value, write the expected group payoff value and the group uncertainty value into the sampling parameter vector, and perform random number generation operation according to the Thompson sampling rule to generate the group sampling payoff value; after all group sampling payoff values ​​have been generated, perform sorting operation according to the group sampling payoff values ​​to generate the group priority sequence.

[0128] Within the group priority sequence, read the group identifiers sequentially, enter the observation unit level within the group, read the corresponding expected return component and uncertainty component for each observation unit identifier within the group, construct the unit sampling parameter vector, and perform Thompson sampling to generate the unit sampling return value; perform sorting operation on the sampling return values ​​of the observation units within the group to generate the unit priority sequence.

[0129] The group priority sequence and the corresponding unit priority sequence are concatenated hierarchically to generate the observation path decision sequence, and then the observation path decision sequence is written into the path decision cache.

[0130] In this embodiment, step 6, establishing the group post-test buffer, specifically includes:

[0131] The system reads the set of group identifiers. For each group identifier, the system allocates a separate storage area in memory, defined as the group posterior parameter cache. This cache corresponds one-to-one with each group identifier, and the group identifier is then used as an index key for access and location.

[0132] When establishing the buffer, first, the expected return storage location and the uncertainty storage location are written for each group. These two storage locations are used to record the summative expected return value and the summative uncertainty value of the observation units within the group, respectively. Then, all observation unit identifiers contained in the group are traversed, and the expected return component and uncertainty component of each observation unit in the long-term ecological return posterior distribution parameter table are read one by one. The read expected return values ​​are accumulated to generate the group expected return value, and the read uncertainty values ​​are accumulated to generate the group uncertainty value. Finally, the accumulated result is written to the corresponding buffer location of the group.

[0133] After all groups have been traversed and written, a set of group posterior parameter buffers is formed, matching the number of group identifiers. This buffer serves as the direct data source for subsequent group-level Thompson sampling, and is then synchronously refreshed after each observation path update to maintain the real-time performance and consistency of the posterior parameters.

[0134] In this embodiment, step 6, establishing the group prior parameter table, specifically includes:

[0135] Read all group identifiers from the observation unit group table, and build a group prior parameter record structure for each group identifier. Each group prior parameter record structure contains two data fields: the group expected return prior component and the group uncertainty prior component. Then, allocate independent storage locations for the two fields respectively.

[0136] During the initialization phase, initial values ​​are written to the prior components of expected returns for each group based on a preset ecological benchmark return level. These initial values ​​can be uniform constant values ​​or benchmark averages generated from historical statistical results. Initial uncertainty values ​​are written to the prior components of group uncertainty; these initial uncertainty values ​​characterize the initial fluctuation range of group returns.

[0137] After the initial write is completed, the group prior parameter record structure is written into the group prior parameter table in the order of group identifier. Then, the index mapping relationship between the group identifier and the prior parameter record is established.

[0138] During subsequent observations, when the posterior parameters of the group are updated, the system writes the posterior statistical results back to the corresponding fields of the group prior parameter table according to the preset fusion rules, thereby realizing the rolling update of the prior parameters and forming a dynamic and adaptive hierarchical decision prior structure.

[0139] In this embodiment, step 6, performing Thompson sampling, specifically includes:

[0140] Read the expected benefit component and uncertainty component corresponding to the target observation unit identifier in the long-term ecological benefit posterior distribution parameter table, and construct the unit sampling parameter vector, which contains the expected benefit value and the uncertainty value; read the random number seed and then generate a uniform random number u in the interval (0,1); perform standard normal distribution and inverse cumulative distribution calculation on the uniform random number u to obtain the standard sampling value z; perform scaling and translation operations on the standard sampling value z to generate the unit sampling benefit value, with the scaling factor being the uncertainty value and the translation term being the expected benefit value; write the unit sampling benefit value and the target observation unit identifier into the unit sampling benefit value table.

[0141] In this embodiment, step 7 specifically includes:

[0142] Read the observation path decision sequence in the path decision buffer, and read the observation unit identifier one by one in the order of the observation path decision sequence;

[0143] Locate the current time index position within the unified time index sequence, and call the observation unit time index mapping table to obtain the corresponding image frame sequence and acoustic frame sequence; perform convolutional neural network forward inference on the image frame sequence to generate category label vectors, and perform spectral transformation and recurrent neural network forward inference on the acoustic frame sequence to generate behavior label vectors;

[0144] Perform time index writing operations on category identifier vectors and behavior identifier vectors to form observation result record items; write the observation result record items into the observation event sequence table in time index order; perform topic identifier update operations and benefit statistics writing operations on newly added records in the observation event sequence table, and write the update results into the topic count table and the long-term ecological benefit posterior distribution parameter table; after completing the traversal of the observation path decision sequence, output the bird identification result table.

[0145] Example 1:

[0146] To verify the feasibility of this invention in practice, it was applied to a multi-observation unit collaborative observation scenario in a natural ecological monitoring area. Several fixed observation units were deployed within this area, each possessing both image and acoustic acquisition capabilities. In traditional methods, each observation unit samples in turn according to a preset inspection sequence, with fixed paths. This approach cannot dynamically adjust to changes in the observation event sequence, easily leading to the same area being visited multiple times in a short period, while other areas are missed within the same cycle. This results in insufficient observation coverage, a high proportion of duplicate observations, and significant fluctuations in the number of effectively identified samples.

[0147] In this embodiment, image and acoustic streams are first acquired to generate a set of observation unit identifiers, followed by the calculation of the observation unit adjacency matrix. Subsequently, category and behavior identifier vectors are generated within the time index window and written into the observation event sequence table, establishing a hierarchical Dirichlet process (HDP) parameter table. Under the adjacency matrix constraints, adjacent observation units are coupled and sampled to update the execution graph, and the topic mixing ratio vector table is written back to form a spatial adjacency-constrained topic mixing ratio vector table. Then, the long-term ecological benefit posterior distribution parameter table is calculated. Finally, Thompson sampling is performed on the group posterior distribution and the observation unit posterior distribution to generate an observation path decision sequence, followed by observation execution.

[0148] After continuous cycle testing, the results of comparing the traditional fixed path method, the single-point greedy strategy and the method of this invention are shown in the table below.

[0149] Table 1: Comparison of the effects of collaborative observation by multiple observation units

[0150] Indicator Name Traditional fixed path method Single-point greedy strategy Method of the present invention Difference Description Average observation coverage (%) 63.4 71.2 89.6 An increase of 18.4 percentage points compared to the traditional level. Percentage of repeated observations (%) 27.8 21.5 9.3 A decrease of 18.5 percentage points compared to the traditional level. Number of validly identified samples (strips) 12840 15320 21475 8,635 more items than traditional ones Path switching stability index 0.61 0.73 0.92 Increased by 0.31 compared to the traditional method. Average decision time per cycle (milliseconds) 84 96 88 4 milliseconds longer than traditional

[0151] Table 1 presents the comparison results of the three methods under the same set of observation units. Regarding average observation coverage, the traditional fixed path method achieves 63.4%, the single-point greedy strategy achieves 71.2%, and the method of this invention reaches 89.6%, an increase of 18.4 percentage points compared to the traditional method, and a decrease of 7.8 percentage points compared to the single-point greedy strategy, representing a difference of 10.6 percentage points. This indicates that the path decision sequence has a more balanced coverage. Regarding the proportion of repeated observations, the traditional fixed path method achieves 27.8%, the single-point greedy strategy achieves 21.5%, and the method of this invention achieves 9.3%, a decrease of 18.5 percentage points compared to the traditional method, and a decrease of 12.2 percentage points compared to the single-point greedy strategy. This indicates that the graph coupling update of adjacent observation unit pairs reduces repeated visits to the same area. Regarding the number of effectively identified samples, the traditional method yielded 12,840 samples, the single-point greedy strategy yielded 15,320 samples, and the method of this invention reached 21,475 samples, an increase of 8,635 samples compared to the traditional method and 6,155 samples compared to the single-point greedy strategy. The increase in samples is consistent with the improvement in coverage, indicating that the effective records available for modeling in the observed event sequence table are more concentrated. Regarding the path switching stability index, the traditional method is 0.61, the single-point greedy strategy is 0.73, and the method of this invention is 0.92, an increase of 0.31 compared to the traditional method and 0.19 compared to the single-point greedy strategy. The improvement in the stability index means that the variation of the observed path decision sequence within adjacent periods is more controllable. Regarding the average decision time per period, the traditional method is 84 milliseconds, and the method of this invention is 88 milliseconds, an increase of only 4 milliseconds, and a reduction of 8 milliseconds compared to the single-point greedy strategy. This indicates that after adding graph-coupled sampling updates and hierarchical sampling decision calculations, the overall computational burden remains within an acceptable range. In summary, the present invention shows consistent numerical improvements in coverage, repetition rate, effective sample output, and path stability.

[0152] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A bird intelligent observation and recognition method based on reinforcement learning, characterized in that, Includes the following steps: Step 1: Acquire the observation image stream and acoustic stream, generate the observation unit identifier set, read the spatial coordinates of the observation units, and calculate the adjacency matrix of the observation units; Step 2: Generate category identifier vectors and behavior identifier vectors within the time index window and write them into the observation event sequence table; Step 3: Establish the Hierarchical Dirichlet Process (HDP) parameter table; Step 4: Select adjacent observation unit pairs according to the adjacency matrix, read the topic mixing ratio vector and topic counting vector, perform component coupling calculation to generate coupling ratio vector, combine topic matching count to generate topic sampling score vector, perform random sampling after normalization to generate topic identifier, update the topic counting table and global topic parameter table, and then write back to the topic mixing ratio vector table; Step 5: Calculate the posterior distribution parameter table of long-term ecological benefits; Step 6: Establish an observation unit group table, perform Thompson sampling on the group posterior distribution and the observation unit posterior distribution, generate group sampling payoff value and unit sampling payoff value, and generate observation path decision sequence according to the sampling payoff value; Step 7: Execute the observations according to the observation path decision sequence and then write them into the observation event sequence table.

2. The bird intelligent observation and recognition method based on reinforcement learning according to claim 1, characterized in that, Step 1 specifically includes: Acquire observation image streams and acoustic streams; perform frame extraction on the observation image streams to generate image frame sequences; and perform frame segmentation on the acoustic streams to generate acoustic frame sequences. Write timestamps to image frame sequences and acoustic frame sequences to generate a unified time index sequence; Read the spatial coordinate table of the observation unit, and perform Euclidean distance calculation on the spatial coordinates to generate the distance matrix of the observation unit; A threshold comparison is performed on the distance matrix of the observation units to generate an adjacency marker matrix, and a symmetric verification is performed on the adjacency marker matrix to generate an adjacency matrix of the observation units. An observation unit time index mapping table is constructed based on the observation unit identifier set and the unified time index sequence, and then written to the data cache area.

3. The bird intelligent observation and recognition method based on reinforcement learning according to claim 2, characterized in that, Step 2 specifically includes: Select a time index window within a unified time index sequence, and read the image frame sequence and acoustic frame sequence corresponding to the time index window; Perform forward inference operations on the image frame sequence using a convolutional neural network to generate bird category probability vectors. Then, perform maximum probability index reading operations on the bird category probability vectors to generate bird category identification vectors. Perform Mel-frequency transform on the acoustic frame sequence to generate a spectrum matrix, perform forward inference on the spectrum matrix using a recurrent neural network to generate a behavior probability vector, and perform maximum probability index reading on the behavior probability vector to generate a behavior identifier vector; Perform time index alignment operation on the bird category identifier vector and behavior identifier vector to generate category identifier vector table and behavior identifier vector table, and write the category identifier vector table and behavior identifier vector table into the observation event sequence table in time index order.

4. The bird intelligent observation and recognition method based on reinforcement learning according to claim 3, characterized in that, Step 3 specifically includes: Read the set of observation unit identifiers and the sequence table of observation events, perform grouping and slicing operations on the sequence table of observation events according to the observation unit identifiers, and generate a set of sub-tables of observation unit events. Extract the category identifier vector table and behavior identifier vector table from the event sub-table set of the observation unit, and perform concatenation operation on the category identifier vector and behavior identifier vector in time index order to generate event identifier sequence; Create a topic index table and then write the initial topic identifiers. Create a global topic parameter table and then write the category component count vector and behavior component count vector for each topic identifier. Establish a subject-mixing ratio vector table for each observation unit, and then write the initial value of the subject-mixing ratio vector for each observation unit identifier. After establishing a topic counting table, the initial value of the topic counting vector is written for each observation unit identifier. The topic counting vector contains the counting positions consistent with the topic index table. The event identifier sequence is counted and accumulated according to the topic index table. The count accumulation result is written into the global topic parameter table and the topic count table. The initial value of the topic mixing ratio vector is normalized and then written into the topic mixing ratio vector table of the observation unit to form the hierarchical Dirichlet process HDP parameter table.

5. The bird intelligent observation and recognition method based on reinforcement learning according to claim 4, characterized in that, Step 4 specifically includes: Read the adjacency matrix of observation units, the subject mixing ratio vector table of observation units, the subject count table, the global subject parameter table, and the subject index table; Generate a set of adjacent observation unit pairs based on the positions in the observation unit adjacency matrix where the adjacency marker value is 1; Perform sequential traversal of the set of adjacent observation units, read the event sub-table set of the corresponding observation unit of the adjacent observation unit, extract the event identifier sequence, establish the event topic identifier cache table, and write the initial topic identifier into the event topic identifier cache table according to the time index position. Read the event identifier vector at each time index position in the event identifier sequence, traverse the topic identifier set according to the topic index table, read the category component count vector and behavior component count vector corresponding to the topic identifier in the global topic parameter table, read the count according to the event identifier vector position, and then perform summation to generate the event matching count value vector. Read the subject mixing ratio vector of the first observation unit and the subject mixing ratio vector of the second observation unit in the adjacent observation unit pair, and perform coupling coefficient product and component addition on the two subject mixing ratio vectors to generate a coupling ratio vector; The coupling ratio vector and the event matching count vector are multiplied by components to generate a topic sampling score vector. The topic sampling score vector is summed and normalized by components to generate a topic sampling probability vector. Random sampling is performed on the topic sampling probability vector to obtain the topic sampling identifier. Read the original topic identifier at the current time index position from the event topic identifier cache table, perform deduction update and write update on the topic count table, perform deduction update and write update on the global topic parameter table, and write the topic sampling identifier to the event topic identifier cache table; After the traversal is complete, normalization updates are performed on the topic count vectors of the first and second observation units, the topic mixing ratio vector table of the observation units is written back, and the topic mixing ratio vector table of spatial adjacency constraints is output.

6. The bird intelligent observation and recognition method based on reinforcement learning according to claim 5, characterized in that, Step 5 specifically includes: Read the spatial adjacency constraint topic mixing ratio vector table and the global topic parameter table, and read the topic mixing ratio vector one by one according to the observation unit identifier; Traverse the set of topic identifiers in the topic index table, read the corresponding category component count vector and behavior component count vector for each topic identifier, and calculate the topic statistical weight value. The topic mixing ratio vector and the topic statistical weight value are multiplied together to generate the topic return component vector. Perform component summation on the subject benefit component vector to generate the benefit statistics of the observation unit; Variance calculation is performed on the observation unit's benefit statistics and the theme count vector to generate uncertainty statistics; the benefit statistics are written into the expected benefit component according to the observation unit identifier, and the uncertainty statistics are written into the uncertainty component to generate the long-term ecological benefit posterior distribution parameter table of the observation unit.

7. The bird intelligent observation and recognition method based on reinforcement learning according to claim 6, characterized in that, Step 6 specifically includes: Read the long-term ecological benefit posterior distribution parameter table and the observation unit identifier set, perform a partitioning operation on the observation unit identifier set according to the preset grouping rules, and generate an observation unit group table; A group posterior parameter buffer is established for each group identifier. The group posterior parameter is obtained by summing the expected revenue components of the observation units within the group to obtain the group expected revenue value, and by summing the uncertainty components of the observation units within the group to obtain the group uncertainty value. Create a group prior parameter table and then write in the initial values ​​of the prior distribution parameters; For each group identifier in the group posterior parameter buffer, read the expected group payoff value and the group uncertainty value, write the expected group payoff value and the group uncertainty value into the sampling parameter vector, and perform random number generation operation according to the Thompson sampling rule to generate the group sampling payoff value; After generating all group sampling benefit values, a sorting operation is performed based on the group sampling benefit values ​​to generate a group priority sequence; Read the group identifiers sequentially within the group priority sequence, enter the observation unit level within the group, read the corresponding expected return component and uncertainty component for each observation unit identifier within the group, construct the unit sampling parameter vector, and perform Thompson sampling to generate the unit sampling return value. Perform sorting operations on the sampled revenue values ​​of observation units within the group to generate a unit priority sequence; The group priority sequence and the corresponding unit priority sequence are concatenated hierarchically to generate the observation path decision sequence, and then the observation path decision sequence is written into the path decision cache.

8. The bird intelligent observation and recognition method based on reinforcement learning according to claim 7, characterized in that, Step 6, the execution of Thompson sampling, specifically includes: Read the expected benefit component and uncertainty component corresponding to the target observation unit identifier in the long-term ecological benefit posterior distribution parameter table, and construct the unit sampling parameter vector. The unit sampling parameter vector contains the expected benefit value and the uncertainty value. Read the random number seed and then generate a uniform random number u in the interval (0,1); The standard sample value z is obtained by performing standard normal distribution and inverse cumulative distribution calculation on a uniform random number u; Scaling and translation operations are performed on the standard sampled value z to generate the unit sampled revenue value. The scaling factor is the uncertainty value, and the translation term is the expected revenue value. Write the unit sampling benefit value and the target observation unit identifier into the unit sampling benefit value table.

9. A bird intelligent observation and recognition method based on reinforcement learning according to claim 8, characterized in that, Step 7 specifically includes: Read the observation path decision sequence in the path decision buffer, and read the observation unit identifier one by one in the order of the observation path decision sequence; Locate the current time index position within the unified time index sequence, and call the observation unit time index mapping table to obtain the corresponding image frame sequence and acoustic frame sequence; A convolutional neural network is used to perform forward inference on the image frame sequence to generate a category label vector, and a spectral transformation and a recurrent neural network are used to perform forward inference on the acoustic frame sequence to generate a behavior label vector. Perform time index writing operations on the category identifier vector and behavior identifier vector to form observation result record items; Write the observation results records into the observation event sequence table in time index order; Perform topic identifier update operation and benefit statistics writing operation on newly added records in the observation event sequence table, and write the update results into the topic count table and the long-term ecological benefit posterior distribution parameter table; After completing the traversal of the observation path decision sequence, output the bird identification result table.

Images