Method for identifying low-altitude dynamic target unmanned aerial vehicle trajectory based on machine learning

Abstract

The application discloses a low-altitude dynamic target unmanned aerial vehicle track tracing identification method based on machine learning, which comprises the following steps: step one, constructing a known track state sequence set; step two, constructing an identity track neural graph; step three, updating synaptic connection weights by using an improved Hebbian learning rule, and generating an identity track memory graph through synaptic enhancement and attenuation mechanisms; step four, constructing a target track neural graph, updating synaptic connection weights by using the improved Hebbian learning rule, and obtaining a target track memory graph; step five, constructing a candidate identity set; step six, calculating a resonance identification score; step seven, identifying the candidate identity track memory graph with the highest resonance identification score as a target trace source, and outputting a corresponding identity label and synaptic path mapping result. The application realizes low-altitude dynamic target unmanned aerial vehicle track tracing identification by fusing the improved Hebbian learning rule and graph resonance identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of low-altitude airspace safety monitoring, and particularly relates to a low-altitude dynamic target unmanned aerial vehicle trajectory tracing identification method based on machine learning. BACKGROUND

[0002] With the continuous improvement of low-altitude airspace control requirements and the rapid growth of the number of civilian unmanned aerial vehicles, the behavior trajectory identification and identity tracing technology of low-altitude aircraft, especially unmanned aerial vehicles, has attracted widespread attention. The existing unmanned aerial vehicle identity identification method mainly relies on communication signaling analysis, telemetry device identification or flight control fingerprint matching to trace the identity, but in actual application, the following problems generally exist:

[0003] The identity identification means relying on communication signals is susceptible to interference, deception or encryption avoidance, and is difficult to restore the identity in the absence of signals or in disguise scenarios; the traditional behavior trajectory comparison method often only considers geometric features or trajectory statistical indicators, lacks deep modeling capability of the flight decision logic and behavior state evolution law of unmanned aerial vehicles, and is difficult to capture the identity-specific features behind the trajectory, resulting in low identification accuracy and weak generalization ability in similar trajectory or long time sequence scenarios. The existing methods generally lack a continuous learning mechanism, and cannot realize dynamic adaptation and memory update of new target behaviors, making it difficult to meet the identification requirements of diversified and intelligent unmanned aerial vehicle activities in low-altitude airspace.

[0004] Therefore, how to provide a low-altitude dynamic target unmanned aerial vehicle trajectory tracing identification method based on machine learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a low-altitude dynamic target unmanned aerial vehicle trajectory tracing identification method based on machine learning. The present application combines the improved Hebbian learning rule and the atlas resonance identification mechanism, constructs the structural mapping relationship between the identity trajectory memory atlas and the target trajectory neural atlas, and realizes the trajectory tracing identification of the low-altitude dynamic target unmanned aerial vehicle. By introducing the synaptic enhancement and attenuation mechanism, the ability to describe the temporal causal relationship of the model is improved; the synaptic connection coincidence degree and the behavior state similarity are combined to improve the accuracy of identity matching; finally, the atlas resonance score is used to complete the high-robustness identity identification, which has the advantages of high identification accuracy, strong self-learning ability, good adaptability and strong dynamic updating ability.

[0006] The low-altitude dynamic target unmanned aerial vehicle trajectory tracing identification method based on machine learning according to the embodiment of the present application comprises the following steps:

[0007] Step 1: Collect flight trajectory data of a plurality of low-altitude unmanned aerial vehicles with known identities, and construct a known trajectory state sequence set;

[0008] Step two: constructing an identity trajectory neural graph based on each known trajectory state sequence, the identity trajectory neural graph comprising a plurality of behavior state nodes and directed connection edges, and the initial synaptic connection weight of all synaptic connection edges being set to zero;

[0009] Step three: updating the synaptic connection weight in the identity trajectory neural graph by using an improved Hebbian learning rule, generating an identity trajectory memory graph through synaptic enhancement and attenuation mechanisms, and constructing an identity trajectory memory set in combination with a unique identity label;

[0010] Step four: collecting flight trajectory data of a target to be identified, constructing a target trajectory neural graph, and updating the synaptic connection weight in the target trajectory neural graph by using an improved Hebbian learning rule to obtain a target trajectory memory graph;

[0011] Step five: performing structural matching between the target trajectory memory graph and each identity trajectory memory graph in the identity trajectory memory set, and constructing a candidate identity set according to the synaptic connection coincidence degree and behavior state similarity;

[0012] Step six: performing graph resonance identification on all candidate identity trajectory memory graphs in the candidate identity set, and calculating a resonance identification score;

[0013] Step seven: identifying the candidate identity trajectory memory graph with the highest resonance identification score as the target trace source, and outputting the corresponding identity label and synaptic path mapping result.

[0014] Optionally, the step one is specifically:

[0015] Collecting flight trajectory data of a plurality of low-altitude unmanned aerial vehicles with known identities, the flight trajectory data specifically comprising a timestamp, spatial position information, speed, acceleration, and attitude angle;

[0016] Sorting the flight trajectory data in chronological order, and encapsulating the flight trajectory data at each time into a flight state vector, the flight state vector comprising spatial coordinate values, speed values, acceleration values, and attitude angle values at the corresponding time;

[0017] According to the continuous time sequence, the flight state vectors are combined into a known trajectory state sequence, and a known trajectory state sequence set is constructed in combination with the known trajectory state sequences of a plurality of low-altitude unmanned aerial vehicles.

[0018] Optionally, the step two is specifically:

[0019] Each flight state vector in the known trajectory state sequence is constructed into a behavior state node, the behavior state node carrying spatial coordinate values, speed values, acceleration values, and attitude angle values at the corresponding time;

[0020] A directed connection edge is constructed between two adjacent behavior state nodes, and the directed connection edge represents a time sequence transition relationship of the trajectory behavior;

[0021] A unique synapse connection edge identifier is assigned to each directed connection edge, and the synapse connection weight of all synapse connection edges is initialized to zero, which is used to represent the association strength between behavior states;

[0022] All behavior state nodes and directed connection edges corresponding to the same identity form a complete identity trajectory neural atlas, and the binding relationship with the unique identity label is preserved.

[0023] Optionally, the synapse connection weight in the identity trajectory neural atlas is updated by using the improved Hebbian learning rule, specifically:

[0024] In the identity trajectory neural atlas, any synapse connection edge corresponds to two behavior state nodes, wherein the former behavior state node is the first activated node, and the latter behavior state node is the second activated node;

[0025] When the two behavior state nodes on the synapse connection edge appear continuously in the known trajectory state sequence, the activation time difference of the first behavior state node activating the second behavior state node is calculated, and if the activation time difference is less than or equal to a preset maximum time difference threshold, the synapse is enhanced, and the current synapse connection weight of the two behavior state nodes is increased by the inverse of the activation time difference;

[0026] If the activation time difference is greater than the preset maximum time difference threshold or the activation order is reversed, the current synapse connection weight is not updated;

[0027] A cache queue recording the activation time stamp is set for all synapse connection edges in the identity trajectory neural atlas, and the latest activation time of each synapse connection edge in the known trajectory state sequence is counted;

[0028] For a synapse connection edge that has not been activated or has not been activated again since the last activation for more than a set time length, a decay mechanism is periodically executed; the decay mechanism includes: calculating the time interval between the current time and the latest activation time of the synapse connection edge, and calculating a decay value according to the time interval, multiplying the current synapse connection weight by the decay value to update the new synapse connection weight, and the decay function is a monotonically decreasing function with a value range of 0 to 1.

[0029] Optionally, the step four is specifically:

[0030] Collecting low-altitude unmanned aerial vehicle flight trajectory data of the target to be identified to construct a target trajectory state sequence;

[0031] construct a target trajectory neural graph based on the target trajectory state sequence, each behavior state node in the target trajectory neural graph corresponding to a flight state vector, and a synaptic connection edge being established between adjacent behavior state nodes, and synaptic connection weights of all synaptic connection edges being initialized as zero;

[0032] adopt the improved Hebbian learning rule to update synaptic connection weights in the target trajectory neural graph, obtain updated synaptic connection weights through synaptic strengthening and attenuation mechanisms, and obtain a target trajectory memory graph.

[0033] Optionally, the structure matching comprises comparing connection structures and synaptic connection weights of synaptic connection edges in the target trajectory memory graph and the identity trajectory memory graph, calculating synaptic connection coincidence degree and behavior state similarity;

[0034] The calculation method of the synaptic connection coincidence degree comprises: obtaining synaptic connection edges with same start and end behavior state nodes in the target trajectory memory graph and the identity trajectory memory graph, denoted as synaptic connection coincident edges, calculating a proportion of the number of the synaptic connection coincident edges to the number of all synaptic connection edges in the target trajectory memory graph, and obtaining the synaptic connection coincidence degree;

[0035] The calculation method of the behavior state similarity comprises: extracting flight state vectors corresponding to front and rear two behavior state nodes of the synaptic connection coincident edges, respectively, comparing flight state vectors at the same position in the target trajectory memory graph and the identity trajectory memory graph, adopting cosine similarity to measure the closeness of the included angle, and averaging cosine similarities of flight state vectors of all synaptic connection coincident edges to obtain the behavior state similarity;

[0036] The synaptic connection coincidence degree and the behavior state similarity are combined and calculated according to a preset weighting factor to obtain an overall matching similarity;

[0037] If the overall matching similarity of the identity trajectory memory graph and the target trajectory memory graph is higher than a preset similarity threshold, the identity trajectory memory graph is taken as a candidate identity trajectory memory graph, and a candidate identity set is constructed based on all candidate identity trajectory memory graphs.

[0038] Optionally, the step six specifically comprises:

[0039] performing graph resonance recognition on each candidate identity trajectory memory graph in the candidate identity set and the target trajectory memory graph, and performing node mapping and path comparison in the candidate identity trajectory memory graph in sequence according to a synaptic activation path in the target trajectory memory graph;

[0040] When the continuous synaptic connection path in the target trajectory memory graph has the same start and end behavior state nodes and consistent connection direction synaptic connection edges in the candidate identity trajectory memory graph, and the synaptic connection weight of the synaptic connection edge is not lower than the set minimum effective weight threshold, it is recorded as an effective resonance response;

[0041] The score of each effective resonance response is calculated, which is the sum of the path length score, the activation sequence consistency score and the synaptic connection weight score;

[0042] The path length score: when the number of continuously matched synaptic connection edges in the mapping path is greater than or equal to the preset minimum path length, 1 point is obtained;

[0043] The activation sequence consistency score: when the activation sequence of each behavior state node in the target trajectory memory graph is completely consistent with the order of the synaptic path in the candidate identity trajectory memory graph, and there is no reversal of the order of the behavior state nodes, 1 point is obtained;

[0044] The synaptic connection weight score: when the synaptic connection weight of more than half of the synaptic connection edges in the mapping path is higher than the preset high weight threshold, 1 point is obtained;

[0045] The score of each effective resonance response is calculated, which is the sum of the path length score, the activation sequence consistency score and the synaptic connection weight score;

[0046] Optionally, the step seven is specifically:

[0047] The candidate identity trajectory memory graph with the highest resonance recognition score is identified as the target trace source, and the unique identity label of the target trace source is output as the identity recognition result of the target to be identified, and the synaptic path mapping result between the target trajectory memory graph and the identified candidate identity trajectory memory graph is output, the synaptic path mapping result including the behavior state node pair information and the synaptic connection weight of all successfully matched synaptic connection edges between the target trajectory memory graph and the identified candidate identity trajectory memory graph.

[0048] The present application has the following advantages:

[0049] The application introduces an improved Hebbian learning rule and a graph resonance recognition mechanism by constructing an identity trajectory memory graph and a target trajectory memory graph, and adopts a time-causal driven synaptic enhancement and decay mechanism to model a flight behavior state in a deep layer, and explicitly retains a dynamic evolution law of a trajectory behavior, aiming at the problems of strong communication dependency, shallow trajectory modeling and lack of continuous learning ability in existing unmanned aerial vehicle identity recognition technologies; in an identity comparison stage, a graph level structure matching is performed based on a synaptic connection structure coincidence degree and a flight state vector similarity, so as to improve a discrimination ability of identity feature recognition; in an identification reasoning stage, a graph resonance response scoring strategy based on a path length, an activation sequence consistency and a synaptic connection weight significance is introduced, so as to significantly enhance a sensitivity and an identification precision of the system to a trajectory microstructure difference. Through a continuous updating mechanism of an identity trajectory memory body set, the application realizes dynamic self-learning of a low-altitude unmanned aerial vehicle behavior memory and long-term enhancement of an identity feature, and provides a high robustness and high expandability technical path for realizing unmanned aerial vehicle identity traceability recognition in a signal missing, trajectory similar or multi-target interference background. BRIEF DESCRIPTION OF DRAWINGS

[0050] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0051] Fig. 1 The overall flowchart of the low-altitude dynamic target unmanned aerial vehicle trajectory traceability recognition method based on machine learning proposed by the application;

[0052] Fig. 2 The flowchart of weight updating of the improved Hebbian learning rule of the low-altitude dynamic target unmanned aerial vehicle trajectory traceability recognition method based on machine learning proposed by the application in an identity trajectory neural graph;

[0053] Fig. 3 The structure matching flowchart between the identity trajectory memory graph and the target trajectory memory graph of the low-altitude dynamic target unmanned aerial vehicle trajectory traceability recognition method based on machine learning proposed by the application. DETAILED DESCRIPTION

[0054] The application will now be described in further detail, by way of example only, with reference to the accompanying drawings. These drawings are not to scale and are merely schematic representations used for the purpose of illustration only. They show embodiments of the application in which:

[0055] REFERENCE Figs. 1-3 The low-altitude dynamic target unmanned aerial vehicle trajectory traceability recognition method based on machine learning comprises the following steps:

[0056] Step one: collect flight trajectory data of multiple low-altitude unmanned aerial vehicles with known identities, and construct a known trajectory state sequence set;

[0057] Step two: construct an identity trajectory neural graph based on each known trajectory state sequence, the identity trajectory neural graph comprising a plurality of behavior state nodes and directed connection edges, and the initial synaptic connection weight of all synaptic connection edges being set to zero;

[0058] Step three: update the synaptic connection weight in the identity trajectory neural graph by using an improved Hebbian learning rule, generate an identity trajectory memory graph through synaptic enhancement and attenuation mechanisms, and construct an identity trajectory memory set in combination with a unique identity label;

[0059] Step four: collect flight trajectory data of a target to be identified, construct a target trajectory neural graph, update the synaptic connection weight in the target trajectory neural graph by using the improved Hebbian learning rule, and obtain a target trajectory memory graph;

[0060] Step five: perform structural matching between the target trajectory memory graph and each identity trajectory memory graph in the identity trajectory memory set, and construct a candidate identity set according to the synaptic connection coincidence degree and the behavior state similarity;

[0061] Step six: perform graph resonance identification on all candidate identity trajectory memory graphs in the candidate identity set, and calculate a resonance identification score;

[0062] Step seven: identify the candidate identity trajectory memory graph with the highest resonance identification score as the target trace source, and output the corresponding identity label and synaptic path mapping result.

[0063] In the embodiment, the step one is specifically:

[0064] Collect flight trajectory data of multiple low-altitude unmanned aerial vehicles with known identities, the flight trajectory data specifically comprising a timestamp, spatial position information, speed, acceleration, and attitude angle;

[0065] Sort the flight trajectory data in chronological order, encapsulate the flight trajectory data at each time into a flight state vector, and the flight state vector comprising spatial coordinate values, speed values, acceleration values, and attitude angle values at the corresponding time;

[0066] According to the continuous time sequence, the flight state vectors are combined into known trajectory state sequences, and the known trajectory state sequence set is constructed in combination with the known trajectory state sequences of multiple low-altitude unmanned aerial vehicles.

[0067] In the embodiment, the step two is specifically:

[0068] each flight state vector in the known trajectory state sequence is constructed as a behavior state node, the behavior state node carrying spatial coordinate values, velocity values, acceleration values and attitude angle values at the corresponding time;

[0069] a directed connection edge is constructed between two adjacent behavior state nodes, the directed connection edge representing a time sequence transfer relationship of trajectory behavior;

[0070] a unique synapse connection edge identifier is assigned to each directed connection edge, and the synapse connection weights of all synapse connection edges are initialized to zero, the synapse connection weights being used to represent the association strength between behavior states;

[0071] all behavior state nodes and directed connection edges corresponding to the same identity are constructed into a complete identity trajectory neural graph, and a binding relationship with a unique identity label is maintained.

[0072] In the embodiment, the synapse connection weights in the identity trajectory neural graph are updated using the improved Hebbian learning rule, specifically:

[0073] In the identity trajectory neural graph, any synapse connection edge corresponds to two behavior state nodes, wherein the former behavior state node is a pre-activation node, and the latter behavior state node is a post-activation node;

[0074] When the two behavior state nodes on the synapse connection edge appear continuously in the known trajectory state sequence, the activation time difference of the pre-activation node activating the post-activation node is calculated, and if the activation time difference is less than or equal to a preset maximum time difference threshold, synapse strengthening is performed, and the current synapse connection weight of the two behavior state nodes is increased by the inverse of the activation time difference;

[0075] If the activation time difference is greater than the preset maximum time difference threshold or the activation order is reversed, the current synapse connection weight is not updated;

[0076] A cache queue recording activation time stamps is set for all synapse connection edges in the identity trajectory neural graph, and the most recent activation time of each synapse connection edge in the known trajectory state sequence is counted;

[0077] For a synapse connection edge that has not been activated or has not been activated again since the last activation for more than a set time length, a decay mechanism is periodically executed; the decay mechanism includes: calculating the time interval between the current time and the most recent activation time of the synapse connection edge, and calculating a decay value according to the time interval, multiplying the current synapse connection weight by the decay value to update the new synapse connection weight, the decay function being a monotonically decreasing function with a value range of 0 to 1;

[0078] Decay function is defined according to time interval The difference between the current time and the latest activation time is defined as:

[0079] ;

[0080] Wherein, Indicates the decay value, satisfies , Indicates the decay rate coefficient;

[0081] ;

[0082] Wherein, Indicates the updated synaptic connection weight, Indicates the synaptic connection weight before updating;

[0083] The application is improved on the basis of the traditional Hebbian learning rule to adapt to the time sequence characteristics and neural graph structure demand in low-altitude unmanned aerial vehicle trajectory behavior recognition, and an improved Hebbian learning rule combining synaptic connection time sequence activation interval and dynamic weight updating mechanism is proposed.

[0084] The application strengthens the modeling of the time causal relationship of the synaptic connection edge in the neural graph by introducing the behavior state activation time difference mechanism. In the identity trajectory neural graph, each synaptic connection edge connects two flight behavior state nodes with physical meaning. When the two nodes are activated in continuous order in the trajectory sequence, if the activation time difference is within the preset threshold, it is determined that the synaptic connection edge has the condition of neural plasticity, and the weight enhancement operation is performed. The enhancement amplitude is no longer fixed or linear gain, but is calculated according to the reciprocal of the activation time difference, so that the connection weight of the node pair with closer time is increased more significantly, which reflects stronger time sequence correlation.

[0085] In order to avoid the accumulation of invalid connection redundancy in the network, the application introduces a decay mechanism with time memory function. Each synaptic connection edge is configured with a cache queue recording the latest activation time. If the synaptic connection edge has not been activated for a long time or no longer appears in the current trajectory, it is regarded as a natural phenomenon of behavior memory degradation. The system will periodically perform decay operation on the synaptic connection weight, and the decay degree is dynamically calculated by a monotonically decreasing function according to the inactivation time interval, so that the invalid path is gradually faded, realizing the dynamic adjustment and cleaning of the behavior memory strength, which helps to maintain the plasticity and accuracy of the identity graph structure.

[0086] Through the combination of the above enhancement and attenuation mechanisms, the improved Hebbian learning rule of the present application not only retains the biological heuristic linkage, i.e., the enhancement principle, but also integrates the modeling capability for high time resolution and dynamic characteristics in the behavior trajectory data, thereby improving the ability of the identity trajectory memory graph to depict the flight mode and the degree of differentiation, which is of key significance for realizing high-precision and low-misjudgment rate of unmanned aerial vehicle identity tracing recognition, and significantly enhances the application value of the present application in actual low-altitude security prevention scenarios.

[0087] In the present embodiment, the step four is specifically:

[0088] Collecting the low-altitude unmanned aerial vehicle flight trajectory data of the target to be identified, and constructing a target trajectory state sequence;

[0089] Based on the target trajectory state sequence, a target trajectory neural graph is constructed, each behavior state node in the target trajectory neural graph corresponds to a flight state vector, and a synaptic connection edge is established between adjacent behavior state nodes, and the synaptic connection weight of all synaptic connection edges is initialized to zero;

[0090] The improved Hebbian learning rule is used to update the synaptic connection weight in the target trajectory neural graph, and the updated synaptic connection weight is obtained through the synaptic enhancement and attenuation mechanism to obtain a target trajectory memory graph;

[0091] The step four and the steps one to three in the present application are consistent in execution logic, and are mainly used for structured modeling and neural graph construction of the flight trajectory data of the target to be identified. Specifically, first, the flight trajectory data of the low-altitude unmanned aerial vehicle to be identified in a specific task scenario is collected, and a target trajectory state sequence is constructed in chronological order. Then, the target trajectory neural graph is generated according to the target trajectory state sequence, in which each flight state is encapsulated as a behavior state node, and a directed synaptic connection edge is established between adjacent state nodes, and the initial synaptic connection weight is uniformly set to zero. Then, the improved Hebbian learning rule proposed by the present application is called, and the synaptic weight enhancement operation is performed according to the activation order and time difference between the behavior states, and the weight attenuation processing is performed on the connection which has not been activated for a long time, so as to obtain a target trajectory memory graph with structure characteristics and time memory capability. This step provides a complete structure input basis for the subsequent matching and resonance recognition with the known identity graph.

[0092] In the present embodiment, the structure matching includes comparing the connection structure and synaptic connection weight of the synaptic connection edge in the target trajectory memory graph and the identity trajectory memory graph, calculating the synaptic connection coincidence degree and the behavior state similarity;

[0093] The calculation method of the synapse connection coincidence degree comprises the following steps: acquiring synapse connection edges with same start and end behavior state nodes in the target trajectory memory graph and the identity trajectory memory graph, and recording the synapse connection edges as synapse connection coincidence edges; and calculating a proportion of the number of the synapse connection coincidence edges in the number of all synapse connection edges in the target trajectory memory graph, to obtain the synapse connection coincidence degree.

[0094] The calculation method of the behavior state similarity comprises the following steps: extracting flight state vectors corresponding to front and rear two behavior state nodes of the synapse connection coincidence edges respectively; comparing flight state vectors at the same position in the target trajectory memory graph and the identity trajectory memory graph; adopting cosine similarity to measure the closeness of the included angle; and taking an average value of cosine similarities of flight state vectors of all synapse connection coincidence edges, to obtain the behavior state similarity.

[0095] The synapse connection coincidence degree and the behavior state similarity are synthesized and calculated according to a preset weighting factor, to obtain an overall matching similarity.

[0096] If the overall matching similarity of the identity trajectory memory graph and the target trajectory memory graph is higher than a preset similarity threshold, the identity trajectory memory graph is taken as a candidate identity trajectory memory graph, and a candidate identity set is constructed based on all candidate identity trajectory memory graphs.

[0097] In the embodiment, the step six is specifically:

[0098] The graph resonance recognition is performed on each candidate identity trajectory memory graph in the candidate identity set and the target trajectory memory graph, and node mapping and path comparison are sequentially performed in the candidate identity trajectory memory graph according to a synapse activation path in the target trajectory memory graph.

[0099] When a continuous synapse connection path in the target trajectory memory graph has a synapse connection edge with same start and end behavior state nodes and consistent connection direction in the candidate identity trajectory memory graph, and a synapse connection weight of the synapse connection edge is not lower than a set minimum effective weight threshold, an effective resonance response is recorded once.

[0100] A score of each effective resonance response is calculated, and the score of each effective resonance response is a sum of a path length score, an activation sequence consistency score and a synapse connection weight score.

[0101] The path length score: when a number of continuously matched synapse connection edges in the mapping path is greater than or equal to a preset minimum path length, 1 point is obtained.

[0102] The activation sequence consistency score: when an activation sequence of each behavior state node in the target trajectory memory graph is completely consistent with an order of synapse paths in the candidate identity trajectory memory graph, and there is no behavior state node order reversal, 1 point is obtained.

[0103] the synaptic connection weight score: 1 point when more than half of the synaptic connection weights of the synaptic connection edges in the mapping path are higher than the preset high weight threshold;

[0104] The resonance recognition scores of the candidate identity trajectory memory graphs are obtained by aggregating and dividing the scores of each valid resonance response by the total number of valid resonance responses.

[0105] The resonance recognition score calculation method in the present application comprehensively reflects the structural resonance relationship between the target trajectory memory graph and the candidate identity trajectory memory graph by quantifying the matching degree of the path structure, the behavior sequence and the connection strength between the graphs. First, the path length score is used to ensure that the resonance response has sufficient structural integrity, avoiding interference caused by accidental matching of short paths. Second, the activation sequence consistency score emphasizes the time logic of the behavior evolution process, and only paths with strictly consistent node sequences can obtain points, thereby improving the accuracy of time sequence recognition. Third, the synaptic connection weight score is used to measure the similarity of connection strength, ensuring that only when there is a strong behavior memory response in the candidate graph is it considered to have relevance. After adding up the three scores of each valid resonance response and normalizing by the total number of responses, the influence of multiple low-quality responses and a small number of high-quality responses can be effectively balanced, improving the discrimination and stability of the score. Ultimately, a graph recognition basis with structural comparability and time sequence consistency weight characteristics is formed. The scoring mechanism is simple and clear, and has executability, which can stably support high-credibility target identity traceability judgment.

[0106] In the present embodiment, step seven is specifically:

[0107] The candidate identity trajectory memory graph with the highest resonance recognition score is identified as the target trace source, and the unique identity tag of the target trace source is output as the identity recognition result of the target to be identified. The synaptic path mapping result between the target trajectory memory graph and the identified candidate identity trajectory memory graph is also output, including the behavior state node pair information and synaptic connection weight of all successfully matched synaptic connection edges between the target trajectory memory graph and the identified candidate identity trajectory memory graph, for describing the structural coincidence relationship between the target trajectory behavior and the known identity trajectory.

[0108] Example 1

[0109] In order to verify the feasibility of the present application in implementation, multiple low-altitude unmanned aerial vehicles with known flight strategies and identity tags are selected for trajectory collection and identity graph construction training. Subsequently, a number of target unmanned aerial vehicles with unknown identities are randomly deployed for identity traceability recognition test. This embodiment comprehensively simulates the actual challenges of unmanned aerial vehicle identity recognition in a real low-altitude control environment, including trajectory deception, behavior disturbance and identity broadcast absence.

[0110] In the training phase, each known identity unmanned aerial vehicle continuously performs multiple preset flight tasks, and the system collects its flight state parameters at a fixed frequency, covering timestamp, spatial position, speed, acceleration and attitude angle information, and encapsulates them in time sequence as flight state vectors. Identity trajectory neural graph is constructed through these flight state vectors, and improved Hebbian learning rule is used to perform synaptic connection weight enhancement and attenuation training to form an identity trajectory memory graph with time causal dependence and memory capability.

[0111] In the recognition phase, several target unmanned aerial vehicles with unknown identities are introduced, and their flight trajectories are collected by the system in real time to construct target trajectory neural graphs, which also perform improved Hebbian rule learning process to obtain target trajectory memory graphs. The system compares the graphs with all graphs in the identity trajectory memory set, calculates the synaptic connection coincidence degree and behavior state similarity, screens the candidate identity set, and further calculates the resonance recognition score through the graph resonance recognition strategy to finally lock the recognition result. Table 1 below shows the recognition performance index statistics of the invention in the test phase.

[0112] Table 1 Target unmanned aerial vehicle identification traceability result statistics table

[0113] ;

[0114] According to the data in Table 1 above, it can be seen that the low-altitude dynamic target unmanned aerial vehicle trajectory traceability recognition method based on machine learning proposed by the invention shows high accuracy and stable performance in multiple recognition tasks. All five target numbers (T001 to T005) are successfully identified as their actual identities, and the system identification is completely consistent with the actual identity, with a recognition accuracy of 100%, verifying the effectiveness and reliability of the method in real application scenarios.

[0115] In terms of synaptic connection coincidence degree, the coincidence degree of the five groups of test data is above 0.76, with the highest reaching 0.845, indicating that the structure matching effect between the target trajectory memory graph and the identity trajectory memory graph is good, with high structure alignment accuracy; at the same time, the behavior state similarity is above 0.94, with the highest value being 0.972, which shows that even under the condition of certain dynamic disturbance of flight behavior, the system can still accurately extract flight state features and realize effective comparison, which reflects the robust adaptive ability of the invention to complex trajectory patterns.

[0116] In terms of recognition efficiency, the system's average recognition time is controlled between 2.2 to 2.6 seconds, showing good real-time response capability. The shortest recognition time is T004, only 2.2 seconds, which may be related to its more clear behavior trajectory atlas and higher matching degree; the longest recognition time is T002, 2.6 seconds, but still within the acceptable fast response range.

[0117] This embodiment fully demonstrates the significant beneficial effects of the present application in the field of low-altitude dynamic target unmanned aircraft trajectory tracing identification. By constructing the identity trajectory neural atlas and the target trajectory neural atlas, combined with the memory atlas formed by the improved Hebbian learning rule, the present application not only can effectively capture the time sequence characteristics of flight behavior, but also can accurately establish the synaptic connection relationship between behavior states, so as to realize high-quality identity recognition and trajectory tracing. In the atlas structure matching process, the system integrates the synaptic connection coincidence degree and the behavior state similarity to ensure the accuracy of the identification; in the atlas resonance identification stage, the path length, activation sequence consistency and synaptic weight matching degree evaluation index are introduced to realize the fine-grained selection and accurate discrimination of the candidate identity. In addition, the system has good real-time response capability and stability, and can adapt to the dynamic changes of various flight environments and different trajectory characteristics. The present application not only solves the problems of insufficient low-altitude target tracing identification accuracy and slow response in the prior art, but also provides an identification model with memory ability and self-adaptive adjustment ability, which has good application prospect and engineering promotion value.

[0118] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can make equivalent replacement or change according to the technical solution and inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A machine learning-based method for tracing and identifying the trajectory of low-altitude dynamic targets (UAVs), characterized in that, Includes the following steps: Step 1: Collect flight trajectory data of multiple low-altitude drones with known identities, and construct a set of known trajectory state sequences; Step 2: Construct an identity trajectory neural graph based on each known trajectory state sequence. The identity trajectory neural graph includes multiple behavioral state nodes and directed connection edges. The initial synaptic connection weights of all synaptic connection edges are set to zero. Step 3: The synaptic connection weights in the identity trajectory neural map are updated using an improved Hebbian learning rule. An identity trajectory memory map is generated through synaptic enhancement and decay mechanisms, and an identity trajectory memory set is constructed by combining it with a unique identity label. Step 4: Collect flight trajectory data of the target to be identified, construct a target trajectory neural map, and execute an improved Hebbian learning rule to update the synaptic connection weights in the target trajectory neural map to obtain a target trajectory memory map; Step 5: Perform structural matching between the target trajectory memory map and each identity trajectory memory map in the identity trajectory memory set, and construct a candidate identity set based on synaptic connection overlap and behavioral state similarity; Step 6: Perform graph resonance recognition on all candidate identity trajectory memory graphs in the candidate identity set and calculate the resonance recognition score; Step 7: Identify the candidate identity trajectory memory map with the highest resonance recognition score as the target source and output the corresponding identity label and synaptic path mapping result.

2. The machine learning-based trajectory tracing and identification method for low-altitude dynamic target UAVs according to claim 1, characterized in that, Step one specifically involves: Collect flight trajectory data of multiple low-altitude drones with known identities. The flight trajectory data specifically includes timestamps, spatial location information, speed, acceleration, and attitude angles. The flight trajectory data is sorted in chronological order, and the flight trajectory data at each moment is encapsulated into a flight state vector, which includes the spatial coordinate value, velocity value, acceleration value and attitude angle value at the corresponding moment. The flight state vectors are arranged in a continuous time sequence to form a known trajectory state sequence. Combined with the known trajectory state sequences of multiple low-altitude UAVs, a set of known trajectory state sequences is constructed.

3. The machine learning-based trajectory tracing and identification method for low-altitude dynamic target UAVs according to claim 1, characterized in that, Step two specifically involves: Each flight state vector in the known trajectory state sequence is constructed as a behavior state node, and the behavior state node carries the spatial coordinate value, velocity value, acceleration value and attitude angle value at the corresponding time. A directed connection edge is constructed between two adjacent behavior state nodes, and the directed connection edge represents the temporal sequence transition relationship of the trajectory behavior; Assign a unique synaptic connection edge identifier to each directed connection edge, and initialize the synaptic connection weight of all synaptic connection edges to zero. The synaptic connection weight is used to characterize the associative strength between behavioral states. All behavioral state nodes and directed connection edges corresponding to the same identity are used to form a complete identity trajectory neural graph, and the binding relationship with the unique identity label is preserved.

4. The machine learning-based trajectory tracing and identification method for low-altitude dynamic target UAVs according to claim 1, characterized in that, The method of updating the synaptic connection weights in the identity trajectory neural map using an improved Hebbian learning rule is as follows: In the identity trajectory neural map, any synaptic connection edge corresponds to two behavioral state nodes, where the preceding behavioral state node is the first activated node and the following behavioral state node is the second activated node. When two behavioral state nodes on the synaptic connection edge appear consecutively in a known trajectory state sequence, the activation time difference between the activation of the previous behavioral state node and the activation of the next behavioral state node is calculated. If the activation time difference is less than or equal to a preset maximum time difference threshold, synaptic enhancement is performed, and the current synaptic connection weight of the two behavioral state nodes is increased by the reciprocal of the activation time difference. If the activation time difference is greater than the preset maximum time difference threshold or the activation order is reversed, the current synaptic connection weight will not be updated. A cache queue for recording activation timestamps is set up for all synaptic connection edges in the identity trajectory neural map, and the most recent activation time of each synaptic connection edge in the known trajectory state sequence is counted; For synaptic connection edges that are not activated or have not been reactivated for a set period of time since the last activation, a decay mechanism is periodically executed. The attenuation mechanism includes: calculating the time interval between the current moment and the most recent activation moment of the synaptic connection edge, and calling the attenuation function to calculate the attenuation value based on the time interval, and updating the current synaptic connection weight by multiplying the attenuation value to obtain the new synaptic connection weight. The attenuation function is a monotonically decreasing function with a value range between 0 and 1.

5. The machine learning-based trajectory tracing and identification method for low-altitude dynamic target UAVs according to claim 1, characterized in that, Step four specifically involves: Collect low-altitude UAV flight trajectory data of the target to be identified and construct the target trajectory state sequence; Based on the target trajectory state sequence, a target trajectory neural map is constructed. Each behavioral state node in the target trajectory neural map corresponds to a flight state vector. Synaptic connection edges are established between adjacent behavioral state nodes, and the synaptic connection weights of all synaptic connection edges are initialized to zero. The improved Hebbian learning rule is used to update the synaptic connection weights in the target trajectory neural map. The updated synaptic connection weights are obtained through synaptic enhancement and decay mechanisms to obtain the target trajectory memory map.

6. The machine learning-based trajectory tracing and identification method for low-altitude dynamic target UAVs according to claim 1, characterized in that, The structural matching includes comparing the connection structure and synaptic connection weights of synaptic connection edges in the target trajectory memory map and the identity trajectory memory map, and calculating the synaptic connection overlap and behavioral state similarity. The calculation method of the synaptic connection overlap degree includes: obtaining synaptic connection edges with the same start and end behavior state nodes in the target trajectory memory map and the identity trajectory memory map, which are denoted as synaptic connection overlap edges, calculating the proportion of the number of synaptic connection overlap edges to the total number of synaptic connection edges in the target trajectory memory map, and obtaining the synaptic connection overlap degree. The method for calculating the behavioral state similarity includes: extracting the flight state vectors corresponding to the two preceding and following behavioral state nodes for the overlapping edges of the synaptic connections, comparing the flight state vectors at the same position in the target trajectory memory map and the identity trajectory memory map, using cosine similarity to measure the similarity of the included angle, and taking the average of the cosine similarity of the flight state vectors of all overlapping edges of the synaptic connections to obtain the behavioral state similarity. The synaptic connection overlap and the behavioral state similarity are combined and calculated according to a preset weighting factor to obtain the overall matching similarity. If the overall matching similarity between the identity trajectory memory map and the target trajectory memory map is higher than the preset similarity threshold, it is used as a candidate identity trajectory memory map, and a candidate identity set is constructed based on all candidate identity trajectory memory maps.

7. The machine learning-based trajectory tracing and identification method for low-altitude dynamic target UAVs according to claim 1, characterized in that, Step six specifically involves: For each candidate identity trajectory memory map in the candidate identity set, a graph resonance recognition is performed on the target trajectory memory map. Based on the synaptic activation path in the target trajectory memory map, node mapping and path comparison are performed sequentially in the candidate identity trajectory memory map. When a continuous synaptic connection path in the target trajectory memory map has a synaptic connection edge with the same start and end behavior state node and the same connection direction in the candidate identity trajectory memory map, and the synaptic connection weight of the synaptic connection edge is not lower than the set minimum effective weight threshold, it is recorded as an effective resonance response. Calculate the score for each effective resonance response, which is the sum of the path length score, the activation order consistency score, and the synaptic connection weight score; The path length score is as follows: 1 point is awarded when the number of consecutively matched synaptic connection edges in the mapped path is greater than or equal to the preset minimum path length. The activation order consistency score is as follows: 1 point is awarded when the activation order of each behavioral state node in the target trajectory memory map is completely consistent with the order of the synaptic paths in the candidate identity trajectory memory map, and there is no reversal of the behavioral state node order. The synaptic connection weight score is as follows: when more than half of the synaptic connection edges in the mapping path have a synaptic connection weight higher than a preset high weight threshold, 1 point is awarded. The resonance recognition score of the candidate identity trajectory memory map is obtained by summing the scores of each effective resonance response and dividing by the total number of effective resonance responses.

8. The machine learning-based trajectory tracing and identification method for low-altitude dynamic target UAVs according to claim 1, characterized in that, Step seven specifically involves: The candidate identity trajectory memory map with the highest resonance recognition score is identified as the target source. The unique identity label of the target source is output as the identity recognition result of the target to be identified. The synaptic path mapping result between the target trajectory memory map and the identified candidate identity trajectory memory map is also output. The synaptic path mapping result includes the behavior state node pair information and synaptic connection weight of all successfully matched synaptic connection edges between the target trajectory memory map and the identified candidate identity trajectory memory map.

Images