Unmanned aerial vehicle cluster formation operation mode adaptive identification method and device

Abstract

The unmanned aerial vehicle cluster formation operation mode adaptive identification method and device belong to the technical field of unmanned aerial vehicle cooperative control and mode identification, and particularly relate to an unmanned aerial vehicle cluster formation behavior analysis and operation mode adaptive identification. The method solves the problems of existing technologies, such as chaotic definition of formation operation modes, overlapping features, and lack of effective extraction of high-dimensional space-time coupling features of clusters, which leads to inaccurate identification and poor adaptability under complex working conditions. The method comprises the following steps: knowledge-driven hierarchical decomposition of global formation tasks of the unmanned aerial vehicle cluster; and mode identification of the formation data according to defined formation operation mode categories. The unmanned aerial vehicle cluster formation operation mode adaptive identification method and device are suitable for application scenarios such as unmanned aerial vehicle cluster cooperative control, formation performance intelligent evaluation, and task decision optimization.

Description

Technical Field

[0001] This invention relates to the field of UAV cooperative control and pattern recognition technology, and in particular to an adaptive recognition of UAV swarm formation behavior analysis and operation mode. Background Technology

[0002] Unmanned aerial vehicle (UAV) swarms perform various complex tasks through coordinated formation. Systematic evaluation of their formation performance provides crucial quantitative data for assessing mission success rates and optimizing subsequent swarm control strategies. Since the formation performance of UAV swarms often varies significantly under different operating modes (overall formation performance is composed of the combined contributions of each operating mode), accurate and adaptive identification of formation operating modes is a primary prerequisite for reliably evaluating formation performance and thus supporting improvements in control strategies.

[0003] Currently, formation operation modes are often represented intuitively by "formation phases." However, there is no unified definition of formation phases in this field, and the data characteristics of each phase overlap significantly. Common phases can be summarized as assembly, formation generation, formation maintenance, formation reconstruction, collision avoidance, turning (turning), and formation disbandment. Among these, some phases (such as formation reconstruction and collision avoidance) consist of multiple basic phases, and the characteristics between different phases are highly similar (with repetitive features). For example, formation generation, formation reconstruction, and collision avoidance may all exhibit disorder and inconsistency in formation structure. The lack of unified definitions and the problem of overlapping features (overlapping feature distributions may lead to potential misidentification) increase the difficulty of accurately identifying operation modes.

[0004] At the technical implementation level, existing recognition methods can be mainly divided into two categories: rule-based and data-driven, but both have obvious limitations: Rule-based methods typically rely on ground station commands, status words, and preset criteria for identification. For example, when the difference between the actual relative distance and the expected distance between the drones is less than a set threshold (e.g., 10 meters), it is determined to be in formation maintenance mode, and a corresponding status word is generated. This method faces several challenges in practice: (1) There is a delay between command execution, status generation and actual flight mode, resulting in asynchronous identification; (2) The judgment rules pre-set by experts are prone to failure under complex working conditions such as dense turns and tight formations; (3) Minor deviations caused by environmental disturbances or control fluctuations may conflict with fixed rules (expert-defined patterns), leading to misjudgments.

[0005] Data-driven unsupervised learning methods achieve pattern segmentation by clustering swarm operational data (classifying formation operation modes by analyzing the swarm's operational state). Commonly used algorithms include K-medoids, DBSCAN (Density-Based Spatial Clustering of Applications with Noise), K-means, and Gaussian Mixture Model (GMM). Among these, Gaussian Mixture Model has seen increasing improvements due to its advantages in distribution modeling. However, existing methods often extract features focused on single-machine parameter coupling or global distance / distribution metrics. When dealing with group-level data generated by UAV swarms, these methods struggle to fully characterize the complex spatial relationships and temporal dependencies within the swarm (the spatial and temporal characteristics of inter-machine coupling), leading to insufficient recognition accuracy in complex and dynamic formation behaviors.

[0006] In summary, current UAV swarm formation operation mode recognition technology mainly suffers from the following two problems: (1) the diverse definitions of formation stages and high feature overlap make it difficult to clearly define and distinguish operation modes; (2) existing feature extraction methods are insufficient in expressing the high-dimensional spatiotemporal coupling features of swarms, limiting the accuracy and adaptability of recognition under complex working conditions. Therefore, there is an urgent need for an adaptive recognition method for UAV swarm formation operation modes that can uniformly define operation modes and effectively extract group-level spatiotemporal features. Summary of the Invention

[0007] This invention proposes an adaptive identification method and device for UAV swarm formation operation modes, which solves the problems of chaotic formation operation mode definition, overlapping features, and lack of effective extraction of high-dimensional spatiotemporal coupling features of swarms in existing technologies, resulting in inaccurate identification and poor adaptability under complex working conditions.

[0008] The adaptive recognition method for UAV swarm formation operation mode according to the present invention includes the following steps: Step S1: Based on UAV formation control knowledge, the global formation task of the UAV swarm is decomposed into a knowledge-driven hierarchical structure, and multiple formation operation mode categories with unified and clear semantics are defined. The formation operation mode categories include: assembly mode, disband mode, consistency mode and inconsistency mode. Step S2: Obtain the formation data of the UAV cluster; the formation data includes formation command data issued by the ground station and flight status time sequence data of each UAV; Step S3: Based on the formation operation mode category defined in Step S1, perform pattern recognition on the formation data: For the assembly and disbandment modes, the first clustering model is used to identify them based on the altitude data in the flight status time series data to obtain the vertical feature identification results. For the consistent and inconsistent patterns, a spatiotemporal representation clustering model is used to identify them based on the relative position data between the local drone and the lead drone in the flight state time series data, and preliminary identification results are obtained. Step S4: Optimize the preliminary identification results and output the final formation operation mode sequence.

[0009] Furthermore, a preferred embodiment is provided, wherein the hierarchical decomposition in step S1 includes the following three-level decomposition: First-level decomposition: Based on the formation instructions in the formation instruction data, the global formation task is divided into multiple continuous formation intervals; The second level of decomposition: Within each formation interval, based on the dominant movement direction characteristics of the formation behavior, a vertical feature recognition stage and a horizontal feature recognition stage are divided; the vertical feature recognition stage corresponds to the assembly mode and the disbanding mode. The third level of decomposition: all formation coordination processes in the horizontal feature recognition stage are classified into two categories: consistent mode and inconsistent mode.

[0010] Furthermore, in a preferred embodiment, in step S3, the first clustering model is a Gaussian mixture model.

[0011] Furthermore, in a preferred embodiment, step S3, the identification using a spatiotemporal representation clustering model, includes: Step S3.1: For the formation type to which the current formation interval belongs, obtain a pre-trained Hidden Markov Model for that formation type; Step S3.2: Based on the relative position data within the current formation interval, calculate the Mahalanobis distance sequence; Step S3.3: Use the pre-trained Hidden Markov Model to decode the Markov distance sequence to obtain the corresponding hidden state sequence; Step S3.4: Based on the preset mapping relationship between hidden states and formation operation mode categories, convert the hidden state sequence into a sequence of consistent or inconsistent modes as the preliminary identification result of the current formation interval.

[0012] Furthermore, a preferred embodiment is provided, wherein the pre-trained Hidden Markov Model is obtained through the following steps: For each formation that appears in the task, execute: Step X1: Extract the relative position data from the training data corresponding to the formation and calculate the Mahalanobis distance sequence. The training data includes sample sequences known to be consistent patterns and known to be inconsistent patterns. Step X2: Construct a hidden Markov model with Gaussian emission probability; Step X3: Based on the Mahalanobis distance sequence, train the Hidden Markov Model using an iterative optimization algorithm. Calculate the log-likelihood value of the training data in each iteration. When the training reaches a preset number of iterations or the change in the log-likelihood value satisfies the convergence condition, stop training to obtain the pre-trained Hidden Markov Model.

[0013] Furthermore, a preferred embodiment is provided, wherein the optimization process in step S4 is density-driven recognition optimization, including: Step S4.1: Perform differential operations on the sequence data of the preliminary pattern recognition results to locate the mode switching point; Step S4.2: Set up a sliding window centered on the mode switching point, and calculate the switching density of the mode label in each sliding window; Step S4.3: Correct the mode labels in the sliding window with a switching density higher than the first threshold to inconsistent modes; Step S4.4: Smooth isolated pattern label segments with durations shorter than the second threshold into their adjacent mainstream pattern labels.

[0014] This invention also proposes an adaptive recognition device for UAV swarm formation operation modes, the device comprising the following modules: Module S1: Based on UAV formation control knowledge, the global formation task of the UAV swarm is decomposed into a knowledge-driven hierarchical structure, defining multiple formation operation mode categories that are unified and semantically clear. The formation operation mode categories include: assembly mode, disband mode, consistency mode and inconsistency mode. Module S2: Acquires the formation data of the UAV cluster; the formation data includes formation command data issued by the ground station and flight status time sequence data of each UAV; Module S3: Based on the formation operation mode category defined in module S1, perform pattern recognition on the formation data: For the assembly and disbandment modes, the first clustering model is used to identify them based on the altitude data in the flight status time series data to obtain the vertical feature identification results. For the consistent and inconsistent patterns, a spatiotemporal representation clustering model is used to identify them based on the relative position data between the local drone and the lead drone in the flight state time series data, and preliminary identification results are obtained. Module S4: Optimizes the preliminary identification results and outputs the final formation operation mode sequence.

[0015] The present invention also proposes a computer device comprising: a processor and a memory, the memory being used to store executable instructions of the processor, the processor being configured to execute the adaptive recognition method for UAV swarm formation operation mode described above by executing the executable instructions.

[0016] The present invention also proposes a computer storage medium storing a computer program, wherein when the computer program is executed, it performs the adaptive recognition method for UAV swarm formation operation mode described in any one of the above-mentioned methods.

[0017] The present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the adaptive recognition method for UAV swarm formation operation mode described in any one of the above-mentioned methods.

[0018] The present invention has the following beneficial effects: 1. The UAV swarm formation operation mode adaptive recognition method described in this invention introduces a knowledge-driven method based on hierarchical task networks to decompose the complex formation task into vertical and horizontal feature recognition sub-tasks. Based on the knowledge of consistency control theory, the horizontal features are uniformly mapped to "consistent mode" and "inconsistent mode". This can uniformly and clearly define the formation operation mode, making the feature boundaries between various modes clear, significantly improving the clarity of recognition, fundamentally solving the recognition difficulties caused by diverse definitions and overlapping features, and laying a reliable foundation for subsequent accurate recognition.

[0019] 2. The adaptive recognition method for UAV swarm formation operation mode described in this invention constructs a spatiotemporal representation clustering model, comprehensively utilizes Mahalanobis distance to extract inter-machine spatial structure features, and uses a Gaussian emission (probabilistic) Hidden Markov Model to jointly model parameter coupling and time-dependent features; further, through a density-driven recognition optimization method, it smooths and corrects high-frequency switching and short-pulse noise in the fuzzy transition segment, which can fully and deeply extract and utilize the high-dimensional spatiotemporal coupling features of the swarm, effectively suppressing misidentification, thereby achieving high-precision and high-robust recognition of operation modes under complex and dynamic formation conditions.

[0020] The adaptive identification method and device for UAV swarm formation operation mode described in this invention are applicable to application scenarios such as UAV swarm collaborative control, intelligent evaluation of formation performance, and optimization of mission decisions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating an adaptive identification method for unmanned aerial vehicle (UAV) swarm formation operation mode in one embodiment of the present invention. Figure 2 This is a schematic diagram of knowledge-driven task decomposition (knowledge-driven hierarchical decomposition) in one embodiment of the present invention. Figure 3 This is a schematic diagram of density-driven recognition optimization in one embodiment of the present invention; Figure 4 In one embodiment of the present invention, the formation trajectory (standardized) of "Cluster 1" in the simulation comparison experiment is shown; wherein, Figure 4 The middle (a) is a three-dimensional view of the "Cluster 1" formation flight process, which shows that the formation process always maintains the same altitude; Figure 4 (b) shows a portion of the "Cluster 1" formation trajectory in a two-dimensional view, from assembly to formation generation, completing the formation reconstruction from a rectangular formation to a triangular formation. The red and blue triangular arrows represent the directions of movement, respectively. Figure 5 In one embodiment of the present invention, in a simulation comparison experiment, three UAVs (numbered 1, 7, and 16) were randomly selected, and their relative longitude and relative latitude relative to the lead UAV (numbered 3) were calculated, resulting in a time series curve; wherein, Figure 5 (a) shows the relative longitude curves between the lead drone and the follower drone; Figure 5 (b) shows the relative latitude curves between the lead drone and the follow drone; Figure 6 In one embodiment of the present invention, a schematic diagram of the recognition results of the vertical feature pattern recognition subtask of the "Cluster 1" formation in a simulation comparison experiment is shown; wherein, Figure 6 (a) is a graph showing the recognition results mapped to altitude. The yellow line represents the assembly mode at the highest point, the green line represents the cruise mode, and the blue line represents the disbandment mode as the altitude gradually decreases. Figure 6 (b) shows a comparison between the pattern recognition labels (recognition patterns) and the actual patterns. Label "2" represents the assembly mode, label "3" represents the disbanding mode, and label "4" represents the cruise mode. Figure 7In one embodiment of the present invention, a schematic diagram of the recognition results of the consistency feature pattern recognition subtask of the "cluster 1" formation in a simulation comparison experiment is shown; wherein, the red triangle represents 20 UAVs and their movement directions; the green trajectory represents the consistent pattern, and other trajectories represent the inconsistent pattern; Figure 8 In one embodiment of the present invention, a schematic diagram of the overall identification results of the "Cluster 1" formation in a simulation comparison experiment is shown, where "0" represents the inconsistency mode and "1" represents the consistency mode. Detailed Implementation

[0023] To make the technical solutions and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail and completely below with reference to the accompanying drawings. The various embodiments described below are only some preferred embodiments of the present invention, and not all of them; the various embodiments described below are intended to explain the present invention and should not be construed as limiting the present invention; reasonable combinations of the technical features defined in the various embodiments of the present invention, as well as all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort, are all within the scope of protection of the present invention.

[0024] Implementation Method 1: This implementation method provides an adaptive recognition method for UAV swarm formation operation modes, aiming to achieve accurate and stable recognition of formation operation modes in complex scenarios through unified mode definition and enhanced spatiotemporal feature extraction. The method mainly includes the following steps: Step S1: Based on UAV formation control knowledge, the global formation task of the UAV swarm is decomposed into a knowledge-driven hierarchical structure, and multiple formation operation mode categories with unified and clear semantics are defined. The formation operation mode categories include: assembly mode, disband mode, consistency mode and inconsistency mode. Step S2: Obtain the formation data of the UAV cluster; the formation data includes formation command data issued by the ground station and flight status time sequence data of each UAV; Step S3: Based on the formation operation mode category defined in Step S1, perform pattern recognition on the formation data: For the assembly and disbandment modes, the first clustering model is used to identify them based on the altitude data in the flight status time series data to obtain the vertical feature identification results. For the consistent and inconsistent patterns, a spatiotemporal representation clustering model is used to identify them based on the relative position data between the local drone and the lead drone in the flight state time series data, and preliminary identification results are obtained. Step S4: Optimize the preliminary identification results and output the final formation operation mode sequence.

[0025] In this embodiment, the relative position data between the local machine and the lead drone includes the relative longitude and relative latitude between the local machine and the lead drone.

[0026] In this embodiment, the formation data is a multi-dimensional time-series data set, which should include the following types of data: Command information (i.e. formation command data, or command status data): formation change commands received from ground stations, etc., used for knowledge-driven task decomposition, and as the basis for dividing top-level tasks (such as different formations); Flight status time-series data: the position (e.g., longitude, latitude, altitude), velocity (three-axis velocity), and attitude data of each UAV at every moment; among which: The relative position data obtained from the flight status time series data is used to characterize the cooperative features, referring to the relative longitude difference and relative latitude difference between each "machine" and the designated "leader drone". This is the basis for subsequent calculation of Mahalanobis distance and extraction of inter-machine spatial features.

[0027] Altitude data obtained from flight status time-series data: the absolute or relative altitude of each drone is a key feature for identifying the two vertical characteristic patterns of "assembly" and "dispersal".

[0028] In this embodiment, the technical problem to be solved is how to provide a unified and clear definition of the unmanned aerial vehicle (UAV) swarm formation operation mode, and a formation pattern recognition method based on spatiotemporal representation.

[0029] To address the aforementioned technical challenges, an adaptive identification method for UAV swarm formation operation modes is proposed. This method aims to define complex formation phases as unified formation operation modes through a knowledge-driven task decomposition approach. Simultaneously, it fully extracts the spatiotemporal features between formations using a spatiotemporal representation-based discovery method to achieve formation mode identification, ultimately realizing accurate and stable identification of formation operation modes.

[0030] Specific technical solution: First, the knowledge-driven task decomposition module aims to define a unified operating mode by decomposing the complex stage tasks of the formation. The defined mode has a clear distribution and boundaries. Then, based on the spatiotemporal representation of the UAV swarm formation operation pattern recognition, according to the above clear sub-tasks and operation patterns, a pattern recognition model is constructed by representing the spatial features between UAVs, the temporal features within the UAVs, and the features between parameters. At the same time, in view of the inaccuracy of the pattern recognition model in the recognition of transition segments, the recognition results are optimized to finally obtain high-precision recognition results.

[0031] Implementation Method 2: The hierarchical decomposition in step S1 includes the following three levels of decomposition: First-level decomposition: Based on the formation instructions in the formation instruction data, the global formation task is divided into multiple continuous formation intervals; The second level of decomposition: Within each formation interval, based on the dominant movement direction characteristics of the formation behavior, a vertical feature recognition stage and a horizontal feature recognition stage are divided; the vertical feature recognition stage corresponds to the assembly mode and the disbanding mode. The third level of decomposition: all formation coordination processes in the horizontal feature recognition stage are classified into two categories: consistent mode and inconsistent mode.

[0032] In this embodiment, the division of the vertical feature recognition stage and the horizontal feature recognition stage within each formation interval based on the dominant movement direction characteristics of the formation behavior can also be described as: Within each formation interval, based on whether there is a significant change in the group consistency of the formation behavior, a vertical feature recognition stage and a horizontal feature recognition stage are distinguished: If a significant change in group consistency occurs, it is the vertical feature recognition stage; Otherwise, it is the horizontal feature recognition stage.

[0033] In this embodiment, the two categories of consistency mode and inconsistency mode are a simplified summary of the horizontal feature recognition stages such as "formation generation, formation maintenance, turning, and formation reconstruction".

[0034] It should be noted that, to address the difficulty in recognizing formation operation patterns caused by diverse definitions, overlapping features, and complexity in the formation phase, a knowledge-driven task decomposition method (or module) is proposed. This method (or module) draws on the idea of ​​Hierarchical Task Network (HTN), utilizing prior knowledge in the field of UAV formation control to structurally decompose the global formation operation pattern recognition task from top to bottom (layer-by-layer decomposition), defining sub-tasks (or sub-recognition tasks) with clear boundaries, easy distinction, and common features.

[0035] The specific decomposition process includes the following three levels: (1) First layer: Top-level decomposition based on formation instructions (formation instruction data): The system receives or parses formation control commands (also known as formation command data, which are roughly categorized by shape and include commands for assembly and formation changes to different formations, such as "Form Formation 1", "Form Formation 2", and "Form Formation F") sent by ground stations. Based on these commands, the entire mission timeline is divided into several consecutive formation intervals (e.g., Formation 1, Formation 2... Formation F). Within each interval, the cluster aims to form or maintain a specific geometric configuration (i.e., the target geometry / formation of the formation is the same or similar).

[0036] (2) Second layer: Subtask division based on motion direction characteristics (drone swarm flight motion direction): Within each defined formation interval, based on the dominant direction of movement in the formation behavior (i.e., distinguishing between vertical and horizontal movement dominance), two parallel recognition subtasks are further decomposed: Vertical Feature Pattern Recognition Subtask (Focusing on Stages of Significant Altitude Changes): This subtask targets stages where a swarm of drones exhibits significant, coordinated altitude changes in the vertical direction. Based on task knowledge, this type of subtask mainly corresponds to two modes: assembly (initial climb and rendezvous) and dispersal (final descent and separation).

[0037] The horizontal feature pattern recognition subtask (focusing on stages occurring within the same altitude plane) addresses the formation process of a swarm of drones primarily operating within the same approximate horizontal plane (same altitude layer). This encompasses a series of complex stages, including formation generation, formation maintenance, formation turning (turning), and formation reconfiguration (adjusting formation).

[0038] It should be noted that for the vertical feature pattern recognition subtask: "Begin assembly" is typically a specific command event issued by the ground station. Once this command is received, the start of the mission can be directly marked as the "assembly phase." This is a judgment made using "knowledge," without needing data analysis to "identify" whether it is an assembly. However, data (altitude) can still be used later to verify or refine the start and end points of this phase.

[0039] The "dispersal" instruction may not be obvious, or the dispersal behavior may be mixed with the final formation maintenance / flight behavior in the data, without a clear instruction boundary. Therefore, it is necessary to rely on data features (such as altitude descent, formation dispersal) to identify when the "dispersal" mode begins.

[0040] It should be noted that for the horizontal feature pattern recognition subtask: The behavioral processes covered by the "Horizontal Feature Pattern Recognition Subtask"—namely, formation generation, formation maintenance, formation turning (turning), and formation reconstruction—primarily involve the adjustment and coordination of the relative positions of UAVs in the horizontal plane, where their altitude is typically kept stable. Therefore, when recognizing these processes, attention should be paid to features in the horizontal plane (such as relative positions).

[0041] Furthermore, formation generation, formation maintenance, turning, and formation reconstruction exhibit multi-peak and overlapping distributions, and direct segmentation can easily lead to blurred boundaries. To address this inherent limitation, a unified and interpretable criterion is needed to capture the essential characteristics of the formation process.

[0042] (3) Third layer: Schema reduction based on consistency theory: To address the problem that multiple stages in the aforementioned "horizontal feature pattern recognition subtask" are difficult to distinguish directly, we introduce "consistency theory" as a unified criterion for reduction by drawing on "consistency-based control algorithms".

[0043] This "consistency theory" states that the core of formation control is to achieve consensus in the cluster state; therefore, regardless of whether the cluster is in the process of generation, maintenance, or reconstruction, it can be reduced to two modes based on its essential characteristic of whether it reaches and maintains a stable consistent state: Consistency mode: The cluster maintains a stable and orderly formation, and the motion status (speed, relative position) of each drone is coordinated and consistent.

[0044] Inconsistent mode: The cluster formation is in a state of adjustment, change or disorder, and is converging towards a consistent state or temporarily deviating.

[0045] In summary, for the "horizontal feature pattern recognition subtask", the core knowledge of "consistency" theory is applied to further decompose it into two final pattern labels to be identified: "formation consistency" (i.e., consistent pattern) and "formation inconsistency" (i.e., inconsistent pattern).

[0046] In summary, through this three-level decomposition and reduction, the previously mixed-definition formation stages (such as assembly, disbanding, generation, maintenance, turning, and reconfiguration) are clearly mapped into a few basic pattern labels with distinct characteristics (assembly, disbanding, consistent pattern, and inconsistent pattern), thus laying a solid logical foundation for subsequent accurate data-driven identification. The entire decomposition process is driven by domain knowledge, ensuring its interpretability and rationality.

[0047] In this embodiment, the hierarchical task network is an artificial intelligence planning method. Its core idea is to recursively decompose a complex top-level task (objective) into simpler and more specific sub-tasks by applying a series of "decomposition methods" (such as prior knowledge in the field of UAV formation control) until it is decomposed into basic executable actions.

[0048] In this embodiment, the knowledge-driven task decomposition method (or module) draws on the "idea" of Hierarchical Task Network (HTN) and creatively applies the "methodology" of HTN to the field of motion pattern recognition, rather than directly adopting or implementing a complete and general HTN planning system.

[0049] Specifically: The core idea of ​​HTN is to "decompose complex tasks into simpler subtasks layer by layer using domain knowledge until they are executable." This implementation adopts this logic to solve the problem of confusing formation pattern definitions.

[0050] Differences between the actual implementation and the HTN system: The general-purpose HTN system is a complete artificial intelligence framework that includes a task library, a method library, and a planner. It can handle a variety of uncertain tasks and perform dynamic planning and reasoning.

[0051] The knowledge-driven task decomposition method (or module) is a specialized decomposition process for UAV formation pattern recognition. It consistently applies pre-defined three-layer decomposition rules derived from domain knowledge (formation instructions → vertical / horizontal → consistent / inconsistent). This is a deterministic, goal-oriented decomposition scheme and is not a concept related to the HTN system.

[0052] In this embodiment, prior knowledge in the field of UAV formation control, also known as UAV swarm formation control knowledge, refers to the generally accepted concepts, rules, and logic used to describe and guide formation behavior within this field; it is a type of domain knowledge. Examples include prior experiences and theories such as "formation missions typically begin with assembly," "dispersal is the end of a mission," "UAVs should maintain relative positional stability during the formation maintenance phase," and "formation geometry changes during turning."

[0053] In this embodiment, the global formation operation mode identification task is to identify the operation mode of all UAVs throughout the entire mission.

[0054] In this embodiment, the sub-tasks share common characteristics: The common features refer to the significant attributes shared by a group of behaviors or states at a certain decomposition level, which can be used for classification and differentiation. For example, in the second-level decomposition, the common feature of "assembly" and "disbandment" is "organized and significant changes in group height in the vertical direction"; while the common feature of formation generation, maintenance, and other stages is "mainly carried out in the horizontal plane, involving coordination of relative positions between machines".

[0055] The subtask refers to a more specific and focused identification target obtained through decomposition than the original task. It is not a specific action instruction, but rather a time period or task segment in which its internal patterns need to be identified. For example, "identifying the vertical feature pattern under formation 1" is a subtask, and "identifying the horizontal feature pattern under formation 2" is another subtask. The global formation operation pattern identification task can be divided into multiple subtasks.

[0056] In this embodiment, the "consistency-based control algorithm" is widely used in UAV swarm formation control. The core advantage of this algorithm lies in its strong adaptability to environmental changes and task requirements, as well as its high robustness and stability. Its basic principle is that individuals in the swarm coordinate locally based solely on neighbor information. Through iterative interaction, the entire group gradually converges from an initial disordered or deviated state to a predefined ordered and consistent state.

[0057] Based on the "consistency-based control algorithm," the core and fundamental goal or prerequisite of formation control is to ensure that all UAVs within the cluster achieve and maintain consistency in their speed, heading, relative position, and other states. All specific formation behaviors (generation, maintenance, reconfiguration, etc.) are different processes or scenarios for achieving this goal.

[0058] It should be noted that consensus-based control algorithms (as a control method) are a type of active control algorithm. They design specific control laws, calculate and output control commands (such as acceleration and angular velocity) for each UAV in real time, driving the swarm to actively converge from a disordered state to a consistent state. This implementation addresses the "pattern recognition" problem, not the "control" problem, and therefore does not use a consensus-based control algorithm. Instead, this implementation abstracts a "consistency theory / idea" from consensus-based control algorithms to serve as a theoretical lens for observing and analyzing swarm behavior. This "consistency theory" states that the dynamic process (identification of horizontal features) of a controlled UAV swarm can essentially be abstracted into two macroscopic states: "tending towards consistency" (inconsistency mode) and "maintaining consistency" (consistency mode).

[0059] In this implementation, the "consistency theory" is creatively used to highly summarize and simplify complex formation behaviors (generation, maintenance, reconstruction, etc.), reducing them to two more essential and easier-to-model and identify pattern labels: "consistency" and "inconsistency".

[0060] In this implementation, the entire decomposition process is driven by domain knowledge: the logic, rules, and judgment criteria for the entire task decomposition, such as the division logic and definition of vertical subtasks (assembly / disbandment) and consistency subtasks (consistency / inconsistency), do not come from statistical analysis of data or machine learning, but directly from domain knowledge (prior knowledge) of UAV formation control summarized by human experts. For example, based on this knowledge, we can "know" that the task begins with assembly and ends with disbandment (generating vertical subtasks); we can "know" that the core of formation is to achieve a consistent state (generating a consistency identification subtask). Identification targets and classification criteria are established based on this knowledge, rather than these categories being "discovered" unsupervised from data.

[0061] In this implementation, a unified and interpretable criterion is used to capture the essential characteristics of the formation process: Inherent limitations: The four behaviors of formation generation, maintenance, turning, and reconstruction are complexly distributed and overlapping in the traditional feature space, making it very difficult to directly distinguish them.

[0062] Essential characteristic: "Whether the formation is in a consistent state".

[0063] A unified and interpretable criterion: Drawing on "consistency-based control algorithms," a "consistency theory" is proposed. This theory posits that regardless of whether a drone swarm is generating, maintaining, or reconstructing formations, its core control objective is to converge the swarm from an "inconsistent" (disorderly, adjustment) state to a "consistent" (orderly, stable) state. Therefore, all the aforementioned complex stages can be uniformly described and simplified using the more essential and easily distinguishable characteristics of "consistency" and "inconsistency."

[0064] One of the core innovations in this implementation is that it no longer directly and separately identifies specific stages such as "formation generation" and "maintaining," but simplifies it to the identification of "consistency." Specifically, the entire process of this solution is as follows: First, the formation is divided into large phases according to the formation instructions (Formation 1, Formation 2, etc.). Then, within each formation phase, vertical feature recognition (assembly / dispersal) and horizontal feature recognition are performed separately. Horizontal feature recognition is further simplified to identifying whether the formation belongs to a "consistent mode" or a "disconsistent mode". This is a hierarchical recognition process of "formation -> (vertical / horizontal) -> (consistent / disconsistent)".

[0065] Implementation method 3: In step S3, the first clustering model is a Gaussian mixture model.

[0066] In this embodiment, the first clustering model is a traditional probabilistic clustering (algorithm) model (a probability density function composed of a weighted sum of multiple Gaussian distributions).

[0067] The traditional probabilistic clustering (algorithm) models include Gaussian Mixture Model (GMM), K-Means (which can be regarded as a special case of GMM), DBSCAN (density-based), and hierarchical clustering, etc.

[0068] Among them, the Gaussian Mixture Model (GMM) needs to be trained, such as by estimating the weights, mean vectors and covariance matrices of each Gaussian component from the data.

[0069] It should be noted that the key to drone swarm formation pattern recognition lies in accurately characterizing the spatial coordination between drones and the temporal evolution within each drone, thereby achieving accurate recognition.

[0070] UAV swarm formation pattern recognition based on spatiotemporal representation identifies two types of data features: One approach is to use traditional methods (first clustering model, i.e., traditional probabilistic clustering) to identify patterns that do not have collaborative features (vertical feature pattern recognition). Another approach is to use the Spatiotemporal Representation Clustering (STReC) model to identify patterns with co-existing features (i.e., horizontal feature pattern recognition).

[0071] It should be noted that the spatiotemporal representation clustering (pattern recognition) model is a hybrid model based on a probabilistic graphical model (Hidden Markov Model, HMM) and a distance metric (Mahaviron distance). This model requires labeled data (i.e., data segments known to exhibit "consistent" or "inconsistent" patterns) to estimate the model's parameters. Specifically, the parameters that need to be trained include: the initial probabilities of the HMM, the state transition probability matrix, and the mean vector and covariance matrix of the Gaussian emission probability.

[0072] In this embodiment, feature selection: For the vertical feature pattern recognition subtask, based on knowledge of flight control strategies, the altitude changes significantly between set and disassembled patterns; therefore, altitude is chosen as the feature for recognizing vertical patterns.

[0073] It should be noted that traditional probabilistic clustering is widely used to identify vertical patterns that do not have collaborative features. Therefore, the GMM method is used to identify clustering and disbanding patterns.

[0074] Implementation Method 4: In step S3, the identification using a spatiotemporal representation clustering model includes: Step S3.1: For the formation type to which the current formation interval belongs, obtain a pre-trained Hidden Markov Model for that formation type; Step S3.2: Based on the relative position data within the current formation interval, calculate the Mahalanobis distance sequence; Step S3.3: Use the pre-trained Hidden Markov Model to decode the Markov distance sequence to obtain the corresponding hidden state sequence; Step S3.4: Based on the preset mapping relationship between hidden states and formation operation mode categories, convert the hidden state sequence into a sequence of consistent or inconsistent modes as the preliminary identification result of the current formation interval.

[0075] In this embodiment, the formation types include the formation shapes of drone clusters arranged in triangles, rhombuses, etc.

[0076] In this embodiment, the covariance matrix used in step S3.2 to calculate the Mahalanobis distance is obtained by statistically analyzing the relative position data of all UAVs under the corresponding formation type.

[0077] In this embodiment, feature selection: For the consistency feature pattern recognition subtask, the relative longitude and latitude between the local drone and the lead drone are used as the basis for identifying consistent and inconsistent patterns. The relative longitude and latitude between the local drone and the lead drone have strong spatial correlation, encompassing inter-drone cooperative features.

[0078] Implementation Method 5: The pre-trained Hidden Markov Model is obtained through the following steps: For each formation that appears in the task, execute: Step X1: Extract the relative position data from the training data corresponding to the formation and calculate the Mahalanobis distance sequence. The training data includes sample sequences known to be consistent patterns and known to be inconsistent patterns. Step X2: Construct a hidden Markov model with Gaussian emission probability; Step X3: Based on the Mahalanobis distance sequence, train the Hidden Markov Model using an iterative optimization algorithm. Calculate the log-likelihood value of the training data in each iteration. When the training reaches a preset number of iterations or the change in the log-likelihood value satisfies the convergence condition, stop training to obtain the pre-trained Hidden Markov Model.

[0079] In this embodiment, in step X3, the iterative optimization algorithm is the Baum-Welch algorithm, and the learned model parameters include: the initial state probability distribution, the state transition probability matrix, and the weights, mean vectors, and covariance matrices of each Gaussian component in the Gaussian emission probability.

[0080] In this embodiment, the hidden Markov model with Gaussian emission probability represents the time dependence of the mode state through its state transition probability matrix, and represents the coupling relationship of the feature parameters through its covariance matrix of Gaussian emission probability.

[0081] In this embodiment, a pattern recognition model of Spatiotemporal Representation Clustering (STReC) is constructed to identify consistent and inconsistent patterns.

[0082] Let the formation data of the drone swarm be... ,Include One sortie of drones, No. The observation data (flight status time series data) of each UAV sortie are represented as follows: It is N × M The matrix, The number of (time) sampling points. The parameter dimension (for each sampling point).

[0083] Formation data should include: Command information (formation command data): such as formation commands (formation change commands) and task phase commands (such as "start assembly"). This is mainly used for knowledge-driven task decomposition.

[0084] Unmanned aerial vehicle (UAV) flight status information (or flight status time-series data): This is precisely... The information typically includes: position (longitude, latitude, altitude), speed (three axes), attitude (pitch, roll, yaw angle), etc. M It is the total number of these state variables.

[0085] The relative longitude and relative latitude between the main drone and the lead drone are expressed as follows: ,in Relative longitude This refers to relative latitude.

[0086] Then, pattern recognition is performed on the formations within each formation interval.

[0087] Assume there is Each formation interval is a separate interval, and all subsequent calculations (such as Mahalanobis distance and training different STReC models) are performed independently within each formation interval.

[0088] Calculate the first The Mahalanobis distance (MD) of each formation interval characterizes the spatial features between machines. (1); in: , In formation Next, at the moment t drones i and j The relative position vector (2-dimensional); V: Formation The inverse of the covariance matrix of all UAV relative position vectors encodes the formation. Inherent spatial structural characteristics (correlation and dispersion between dimensions).

[0089] In formation Below, two drones i and j The position difference vector.

[0090] Calculation and significance of Mahalanobis distance: Input: For the first Given a formation interval, the input is the relative position vector of all drones at each time t within that interval. That is, [relative longitude, relative latitude].

[0091] Output: scalar value , indicating a specific formation Next, the i The drone and the first j A drone at all times t The degree to which the relative position deviates from the typical distribution of this formation.

[0092] Meaning of the calculation: Mahalanobis distance is an Euclidean distance that considers the internal correlation of the data. It measures how "abnormal" the relative position of any two machines is at the current moment relative to the expected relative position distribution under that formation. The smaller the value, the closer the formation spatial structure is to the ideal stable state of the formation (high consistency); the larger the value, the greater the deviation (high inconsistency). It extracts the quantitative characteristics of the spatial coordination relationship between machines.

[0093] Traditional clustering methods (such as GMM) only consider distribution and ignore dynamic processes, making them unsuitable for identifying consistent and inconsistent patterns.

[0094] Hidden Markov models (HMMs) explicitly extract temporal dependencies by imposing time constraints on state sequences through state transition matrices. However, since MD (Markov distance) features constitute continuous-valued observations, discrete-emission HMMs are not applicable and are prone to misclassification.

[0095] In this embodiment, a Gaussian emission HMM suitable for continuous value features is adopted, that is, using a Gaussian distribution (or Gaussian mixture model) to describe the probability of observing a certain continuous Markov distance value in a given hidden state. A hidden Markov model is constructed using the Gaussian emission probability to extract in-machine time information and parameter coupling information, thereby preserving the spatial coupling relationship represented by the covariance matrix and the temporal dependence represented by the state transition.

[0096] The Gaussian emission probability is expressed as: (2); in: At any moment n， The HMM model is in a hidden state. (or rather, the first) k Under the condition of (one hidden state), it was observed that The probability of; Hidden state The corresponding number of Gaussian components means that the observation distribution in a (hidden) state can be composed of a mixture of multiple Gaussian components to characterize a complex feature distribution. In hidden state Next, the g The weights of each Gaussian component are given, and the sum of the weights of all components is 1. and : These are in the hidden state respectively Next, the g The mean vector and covariance matrix of each Gaussian component are used to capture the parameter coupling relationship within the feature; The mean is The covariance matrix is The probability density function of a multivariate Gaussian distribution.

[0097] The calculation and significance of the Gaussian emission probability formula: Input: at time n From a certain hidden state of the HMMk (For example, in a "consistent" or "inconsistent" state) the observed feature vector ; It is a vector composed of the calculated Mahalanobis distance values.

[0098] Output: A probability value representing the probability of being in the hidden state. Under the premise of observing the current situation How likely is it?

[0099] Computational significance: This formula defines the observation model of the constructed single HMM model. The model is trained to include a subset of its hidden states (e.g., ...). , The emission probability of ) corresponds to a small Mahalanobis distance distribution (characterizing a "consistent" pattern), while another part of the hidden states (e.g.) , The emission probability of corresponds to a large and dispersed Mahalanobis distance distribution (characterizing an "inconsistent" pattern). Covariance matrix Used to characterize the coupling relationship between features (i.e., the dimensional correlation within a feature).

[0100] Based on the principles of Hidden Markov Models (HMMs), three main problems need to be solved: likelihood assessment, decoding, and model training, in order to obtain and optimize the model parameters. ,in, Indicates the initial probability. Represents the state transition matrix. It is the set of emission probability matrices.

[0101] ① Model training problem (learning): The aim is to find parameters that maximize the overall likelihood probability of the training data, starting from the initial model. .

[0102] The Baum-Welch algorithm is used for iterative optimization, with parameters re-estimated in each iteration (Equation 5), and the process stops when the likelihood converges or a specified number of iterations is reached. This process utilizes mixed data containing both "consistent" and "inconsistent" patterns to jointly train a single Hidden Mirror Model (HMM).

[0103] The parameters updated in each iteration are represented as follows: (5); in, It represents the number of valid samples for each component.

[0104] Objective: To "teach" the model to learn. Given a set of observation sequences (Mahanobis distance sequences) with known labels ("consistent" or "inconsistent"), the optimal model parameters are estimated using the Baum-Welch algorithm (Equation 5). These parameters define the initial distribution of states, the transition patterns, and the observation probabilities. This is the process of "training the STReC model".

[0105] Model Training: For each formation appearing in the task, a corresponding Hidden Markov Model (HMM) is constructed and trained. This model is jointly trained using sample data from all "consistent" and "inconsistent" patterns under that formation, and its parameters... The combined spatiotemporal features of the two modes under this formation are encoded. Training the model separately for different formations can effectively reduce error accumulation.

[0106] ② Likelihood assessment issues (assessment): Likelihood evaluation is a problem that involves given fixed initial parameters. and observation sequence This method efficiently and numerically stably calculates sequence probabilities. The posterior probabilities of each component can be obtained as follows: (3); Objective: To "test" the model. Given a trained model and an observed sequence, calculate the probability (likelihood) that the sequence was generated by the model. This can be used to monitor convergence during model training or to evaluate the overall fit between the sequence and the model in certain scenarios.

[0107] ③ Decoding problem (reasoning): The decoding problem is to recover the most probable chain of hidden states, thereby generating interpretable time segments and pattern labels. The Viterbi algorithm performs a maximum product path search on the grid, avoiding exhaustive search, while providing a backtracking pointer for optimality verification.

[0108] (4); Objective: To perform recognition using a model. Given a trained model and an observation sequence, to recover the most likely hidden state sequence. This implementation uses the Viterbi algorithm (Equation 4) for decoding. The decoded hidden state sequence is the direct basis for subsequent mapping to "consistent" or "inconsistent" pattern labels.

[0109] Pattern recognition (decoding and mapping): For the current formation interval to be identified, obtain the trained Hidden Markov Model (HMM) corresponding to its formation type for decoding and mapping: Decoding the most probable state sequence: the Mahalanobis distance sequence to be identified Input the trained HMM and decode it using the Viterbi algorithm to obtain the most probable sequence of hidden states. , means as follows: (6); in, These are the training parameters for the HMM corresponding to this formation; The operation is used to find the joint probability. Maximum hidden state path .

[0110] State-to-pattern mapping: Based on a predefined mapping relationship, the decoded hidden state sequence is... This is transformed into the final formation running pattern sequence. During training, the different hidden states of this HMM are automatically learned and correspond to the spatiotemporal features of "consistent" and "inconsistent" patterns, respectively. Mapping rules include, for example, mapping the state subset {... , } is mapped to a "consistency pattern", which maps the state subset { , } is mapped to "inconsistency mode".

[0111] Assuming the preliminary identification results are as follows, Indicates consistency mode, Indicates an inconsistent pattern.

[0112] The decision-making process in pattern recognition: Viterbi decoding and state mapping. Decoding: The Mahalanobis distance sequence to be identified is input into a pre-trained single Hidden Markov Model (HMM) with the corresponding formation. This model has been jointly trained with data containing both "consistent" and "inconsistent" samples, and its different hidden states naturally represent the dynamic characteristics of the two modes.

[0113] State Path Search: Using the Viterbi algorithm and dynamic programming, find an optimal hidden state path. This maximizes the joint probability of the path and the observed sequence (i.e., Equation (6)). This path is the most likely state evolution sequence.

[0114] Pattern determination: Based on the preset "hidden state - pattern category" mapping table, the optimal state path is determined. Each state in Real-time conversion to the corresponding pattern label ( or This outputs a preliminary pattern recognition sequence for the entire time period.

[0115] The core and construction of the STReC model: The core of the STReC model is a Hidden Markov Model (HMM) with Gaussian emission probabilities.

[0116] Constructing STReC: i.e., training the HMM (solving problem ①). Using the Baum-Welch algorithm, the model is iteratively trained using Mahalanobis distance sequences that mix "consistent" and "inconsistent" pattern data, allowing the model parameters to... It can best represent the combined statistical characteristics of the two models.

[0117] Recognition is performed using STReC: this involves applying a trained Hidden MM for decoding and mapping. First, the optimal hidden state sequence is decoded using the Viterbi algorithm (solving problem ③); then, the final pattern category sequence is obtained through state mapping.

[0118] In summary, the overall process of using the STReC model to perform pattern recognition on the formation data and obtain preliminary recognition results is roughly as follows: Step 1: Obtain data: Input the raw formation data (including formation instructions and flight status timings for each aircraft).

[0119] Step 2: Spatial Feature Extraction Operation: Calculate the Mahalanobis distance between relative positions according to the formation interval; Input: Relative latitude / longitude from the raw data; Output: Mahalanobis distance sequence. This sequence is a one-dimensional time series that incorporates inter-machine spatial relationships; Function: To transform high-dimensional, correlated relative position data into a scalar time-series signal that can characterize the "compactness of the formation spatial structure".

[0120] Step 3: Spatiotemporal Joint Modeling and Feature Extraction (Internal to HMM): Operation: Input the Mahalanobis distance sequence into the Gaussian emission HMM trained for the current formation.

[0121] Input: Mahalanobis distance sequence.

[0122] The internal function of HMM: Temporal feature extraction: The state transition matrix T of the HMM is used to implicitly learn the evolution of the pattern state over time.

[0123] Parameter coupling and state feature modeling: The distribution of Mahalanobis distance values ​​is modeled using Gaussian emission probabilities (their mean and covariance matrices). During training, the different hidden states of this HMM automatically learn and represent different feature distributions under "consistency" and "inconsistency" modes, respectively.

[0124] Output (after training): A joint probabilistic model that can simultaneously represent spatiotemporal features (i.e., a single HMM after training).

[0125] Step 4: Decoding and Pattern Mapping Operation: Within each formation interval, input the new Mahalanobis distance sequence into the corresponding formation's trained HMM, use the Viterbi algorithm to decode the most likely hidden state sequence, and then convert it into a pattern label according to a preset mapping rule.

[0126] Input: A new Mahalanobis distance sequence, corresponding to a trained HMM model of the formation.

[0127] Output: The most likely hidden state sequence is obtained through decoding, and the category label sequence of "consistent pattern" or "inconsistent pattern" is mapped.

[0128] It should be noted that the STReC model is trained for each type of formation: Different formations (such as rhombuses and straight lines) have completely different spatial structures (e.g., different covariance matrices V for relative position distribution). Training a dedicated Hidden Markov Model (HMM) for each formation enables the model to learn the spatiotemporal characteristics unique to the "consistent" and "inconsistent" patterns under that formation, thereby significantly reducing recognition errors caused by formation differences and improving overall accuracy.

[0129] The STReC model here specifically refers to an Hidden Markov Model (HMM) that integrates spatiotemporal representation capabilities. A STReC model is trained for each type of formation, rather than for each formation interval: the model learns the essential features of the "formation." As long as the formations are the same, their ideal spatial structure and pattern characteristics are similar; therefore, the same model can be used to identify data in all intervals under that formation.

[0130] Data Merging and Model Training: Assume there are three formations: "rectangle," "triangle," and "rhombus." When preparing the training data, the Mahalanobis distance sequences extracted from all intervals labeled "rectangle" (part of which are "consistent" segments, and the other part are "inconsistent") are merged together to train a Hidden Model (HMM) specifically for the "rectangle" formation. During training, this model automatically learns from the mixed data, making different hidden states correspond to the features of the "consistent" and "inconsistent" patterns, respectively. Similarly, separate HMMs are trained using the "triangle" and "rhombus" data.

[0131] Correspondence during recognition: During the recognition phase, when the system processes a certain pattern interval, it first determines which pattern the interval belongs to (e.g., "diamond"), and then calls the HMM model pre-trained for that pattern. Through the decoding and state mapping functions of this model, it simultaneously outputs "consistent" and "inconsistent" pattern recognition results.

[0132] Implementation Method 6: The optimization process in step S4 is density-driven recognition optimization, including: Step S4.1: Perform differential operations on the sequence data of the preliminary pattern recognition results to locate the mode switching point; Step S4.2: Set up a sliding window centered on the mode switching point, and calculate the switching density of the mode label in each sliding window; Step S4.3: Correct the mode labels in the sliding window with a switching density higher than the first threshold to inconsistent modes; Step S4.4: Smooth isolated pattern label segments with durations shorter than the second threshold into their adjacent mainstream pattern labels.

[0133] In this implementation, the pattern label refers to the result identified by the STReC model for each data point at each time step (e.g., the nth sampling point). (Inconsistency pattern) or (Consistency Pattern). The entire recognition result L is a sequence of these labels arranged in chronological order.

[0134] In this embodiment, the mode switching point refers to the moment when the mode label changes. For example, when the label changes from... (Inconsistency mode) becomes (Consistency mode), or vice versa. Because and They are usually represented by numbers (such as 0 and 1), and their absolute difference is 1 at the switching point and 0 when there is no switching. Therefore, the result of the difference operation is... It is 1 only when a mode switch occurs (i.e., when switching between inconsistent and consistent modes).

[0135] The purpose of locating mode switching points is to accurately pinpoint the boundary locations where all modes change in the initial identification results. This forms the basis for subsequent analysis of switching frequency (density) and optimization. Only by locating these points can we identify areas of instability.

[0136] In this embodiment, the difference operation: In mathematics and signal processing, "difference" refers to calculating the difference between two adjacent elements in a sequence. Here, it involves performing a "first-order difference" operation on the preliminary identification result (pattern label sequence L). Formula The calculation is the first n Labels of each moment tag with the previous moment The absolute difference.

[0137] In this embodiment, switching density refers to the frequency of mode label switching within a local time period (sliding window). It is calculated by measuring all difference values ​​within the sliding window. It is obtained by averaging the values ​​of (either 0 or 1). The closer the value is to 1, the more frequently the window switches.

[0138] The purpose of using a sliding window to calculate switching density is that it's difficult to quantify the degree of "instability" by directly observing the entire sequence. By setting a window of fixed time length (w sampling points) and sliding this window along the time axis, the switching density of each local region can be calculated. This is equivalent to transforming the time series into a density series that reflects "local instability." This allows for quantitative analysis and identification of which segments fluctuate dramatically (high density) and which segments remain stable (low density).

[0139] In this embodiment, the mode labels within a high-density sliding window are corrected to be inconsistent modes: In transitional phases where pattern features are ambiguous (e.g., the transition from "inconsistency" to "consistency"), the model may struggle to make accurate judgments due to overlapping data features, leading to inconsistent recognition results. (Inconsistency pattern) and The frequent and rapid switching between (consistent modes) is unreasonable in real-world physical scenarios (formation states do not repeatedly jump between states in milliseconds). This reflects a low confidence level in the model in this region.

[0140] Function: To uniformly and forcibly mark these frequently switching, low-confidence regions as " The "(Inconsistency Mode)" is based on a reasonable physical assumption: the transition process itself is an unstable and inconsistent state. This approach can eliminate unreasonable jitter, making the recognition results smoother and more realistic, and also enhancing the algorithm's robustness to fuzzy areas.

[0141] In this embodiment, the first threshold ( ): It is a threshold that is dynamically determined based on the data. Calculated according to the Raida criterion: .

[0142] Here and It is the whole The mean and standard deviation of a sequence (or a representative segment). This is the upper limit for outlier detection commonly used in statistics, meaning the upper limit of outlier detection for a given window. When the frequency is significantly higher than (more than 3 times the standard deviation) than the overall average, it is judged as "abnormally high frequency" and needs to be corrected.

[0143] In this implementation, an isolated pattern tag segment refers to a short segment of tags that is surrounded by another pattern tag before and after it, with a very short duration. For example: ... , , , , ..., among which the individual It is a "short spike" or "isolated segment".

[0144] Adjacent main pattern tags: These refer to the pattern tags surrounding this isolated segment. In the example above, These are the mainstream mode tags.

[0145] In this embodiment, isolated pattern tag segments with durations shorter than the second threshold are smoothed into their adjacent mainstream pattern tags: These brief "spiking" (isolated mode label segments) are more likely caused by noise, data disturbances, or accidental model misjudgment than by a genuine mode switch. Genuine mode transitions usually have a minimum duration.

[0146] Function: To filter out noise and prevent isolated or erroneous recognition segments caused by accidental fluctuations, so that the final pattern segmentation block is more complete and cleaner.

[0147] In this embodiment, the second threshold ( ): Calculated according to the Raida criterion as well: .

[0148] This is the lower limit for outlier detection. If the switching density within a window... If the density is extremely low (far below average) and the labels within it are isolated short segments, then this low density precisely indicates that this short segment is "out of place" with its surrounding area, and is likely noise that needs to be smoothed out.

[0149] It should be noted that due to the inherent ambiguity and significant overlap in the feature distribution of the transition zone, the (pattern) labels decoded by the STREC model using the Viterbi algorithm are often unstable, with frequent state switching within these intervals. That is, in the transition region, the values ​​of features representing "consistency" and "inconsistency" (such as Mahalanobis distance) are very close, and their distributions overlap. When decoding (using the Viterbi algorithm), the Hidden Markov Model (HMM) selects the most probable path based on minute probability differences, leading to labels at adjacent time points being in different states. or The data jumps back and forth between these points. Therefore, it is necessary to identify and optimize these unstable transition segments to suppress random fluctuations in the decoding results.

[0150] Furthermore, pattern recognition involves pattern segmentation, which divides continuous time-series data into different time periods or "segments" according to their respective (formation) operation modes. The ultimate goal is to assign continuous and accurate pattern labels to the entire task timeline, such as: [0s-10s: assembly], [10s-50s: inconsistency], [50s-100s: consistency], [100s-110s: disbandment]. During pattern segmentation, isolated or erroneous recognition segments may occur due to occasional fluctuations. Therefore, it is necessary to smooth these frequently switching transition intervals, making the boundaries of these segmented segments clearer and the internal segments more consistent, thereby enhancing the overall continuity and robustness of pattern segmentation.

[0151] The above two functions can be achieved by using density-driven recognition optimization.

[0152] Assuming the initial identification results generated by the STReC model are .

[0153] Density-driven recognition optimization methods primarily use first-order difference to index the switching boundary of the initial recognition results (the point in time when the pattern changes, which is the position of the "boundary" on the time axis) to find all The corresponding time n, these n are the positions of the switching boundary).

[0154] (7); if Then there are two modes ( and The switching begins (because the difference results are only available at the moment the switching occurs). Only then does it equal 1. Therefore, search for... Finding the point is equivalent to finding all the switching boundaries.

[0155] For each position n on the time axis (usually centered at each sampling point), take a length of... sliding window Calculate the switching density (mode label) within this sliding window (equivalent to finding the inconsistency of each window under each formation): that is, calculate all the switching densities within this sliding window. The average of (neither 0 nor 1) is obtained. .

[0156] (8); if This indicates the presence of frequently switching modes within the window; the boundaries of these modes are ambiguous and should be defined as inconsistent modes. ; if If , it indicates the existence of short spikes that need to be smoothed, i.e., pseudo-feature segments.

[0157] in: Density-driven recognition optimization aims to address two types of unreasonable scenarios in the initial recognition results: high-frequency oscillations in the transition region and isolated short-term misjudgments caused by noise. Its core principle is to utilize local switching density. Quantify the instability and make corrections accordingly.

[0158] (1) Calculation and significance of local handover density First, by calculating the switching density within the local window. This quantifies the instability of the identified sequence. It quantitatively describes the frequency of pattern label changes over a short period of time around time n. The higher the value, the more frequently the pattern label switches in that local area, and the more uncertain the result. Extremely low values ​​with a switching point suggest the possible presence of isolated noise points. It quantifies "instability" and "isolated noise," allowing the algorithm to operate based on explicit numerical rules (thresholds). and ) Perform automatic correction.

[0159] (2) Correction of high-density windows (handling ambiguity in the transition region): Why are high-density windows defined as inconsistency patterns? If a region's pattern label switches frequently, it indicates that the model cannot be certain whether the current state is "consistent" or "inconsistent" in that region; the probability of either is almost 50 / 50. From the perspective of formation behavior logic, this period of hesitation and ambiguity corresponds to the formation being in a transitional process of adjustment, change, and instability. The definition of an "inconsistency pattern" is precisely when the cluster is in a state of adjustment, change, or disorder. Therefore, these high-uncertainty regions are uniformly classified as "inconsistency patterns." This approach is consistent with physical facts and is a reasonable and robust decision that avoids the result from irrationally and rapidly oscillating between two certain states.

[0160] Specifically, in transitional regions where pattern features are ambiguous (e.g., when transitioning from "inconsistent" to "consistent"), the model decoding result will... and High-frequency oscillations form a high-density window. This oscillation does not conform to the continuity of physical processes. Based on the logic that "transient processes are inherently non-steady-state," this method uniformly modifies the labels within such high-density windows to... (Inconsistency mode). This is equivalent to robustly deciding a "uncertain, rapidly oscillating" identification result into a definite, cognitively consistent state (unstable state).

[0161] (3) Smoothing of short spikes (pseudo-feature segments) (handling random noise): Short spikes (pseudo-feature segments): These refer to isolated, erroneous pattern label fragments that do not reflect true pattern changes but are generated by noise, data perturbations, or random model fluctuations. They are extremely short-lived (e.g., only 1-2 sampling points). They are characterized by occurring in a very stable background and exhibiting extremely low switching density. Isolated switching points appear within the window. This method smooths these brief segments into the adjacent mainstream mode labels, thereby filtering out such obvious random noise and ensuring the continuity and rationality of mode segmentation blocks.

[0162] and It was obtained according to the Raida criteria.

[0163] (9); The Raida criterion is a method for detecting outliers based on the statistical 3σ principle. It assumes that the data follows a normal distribution, and approximately 99.7% of the data will fall within the interval [μ-3σ, μ+3σ] of the mean μ plus or minus three standard deviations σ. Data points falling outside this interval are considered low-probability events, i.e., outliers.

[0164] The Raida criterion is used to automatically determine the threshold for distinguishing between "high-frequency switching windows" and "low-frequency switching windows". and It does not require manual setting of fixed thresholds based on experience, but rather relies on the current sequence of recognition results. The algorithm dynamically calculates the mean and standard deviation based on the inherent statistical properties of the data. This allows the optimization algorithm to adapt to the fluctuating characteristics of different tasks and data, making it more versatile and robust.

[0165] In summary, the density-driven recognition optimization method can correct high-frequency switching regions into inconsistency patterns and smooth out short spikes, thereby improving the accuracy of pattern recognition.

[0166] In summary, based on the above theoretical research, this invention develops an adaptive recognition method for UAV swarm formation operation modes. This method unifies the definition of UAV swarm formation operation modes, which makes the distribution between modes clear and lays the foundation for subsequent recognition. At the same time, this method has the ability to fully extract inter-UAV spatial features, intra-UAV temporal features, and multi-parameter coupling features, achieving accurate formation mode recognition and providing different operating condition information for subsequent intelligent evaluation and decision-making.

[0167] Implementation Method 7: Provide a simulation comparison experiment, the specific implementation method is as follows: (1) Data settings: Three sets of experimental datasets were generated using a drone swarm formation simulation platform. These datasets covered different scales and formation changes. Sensor parameters such as attitude and position were collected during the experiments. The experimental data, shown in Table 1, includes the number of drones and formations in each sortie, and details the formation change process. Formations are represented by their first letters: "Rec" for rectangle, "Tri" for triangle, and "Dia" for rhombus. In addition to different formation combinations, for comprehensive validation, the three flights also simulated the complete formation phase, including assembly, formation generation, formation maintenance, formation reconfiguration, turning, and dispersal.

[0168] Taking "Cluster 1" as an example, its formation trajectory is as follows: Figure 4 As shown.

[0169] Figure 4 (a) is a three-dimensional view of the formation flight process, illustrating that the formation process always maintains the same altitude.

[0170] Figure 4 (b) A portion of the formation trajectory is shown in a two-dimensional view, from assembly to formation generation, completing the formation reconstruction from a rectangular formation to a triangular formation. The red and blue triangular arrows represent the directions of movement, respectively.

[0171] Three drones were randomly selected (numbered 1, 7, and 16), and their relative longitude and latitude relative to the lead drone (numbered 3) were calculated. The resulting time series curves are shown below. Figure 5 As shown.

[0172] Table 1. Cluster Formation Experiment Dataset Information

[0173] (2) Evaluation indicators: accuracy and F A score quantifies model performance by evaluating the difference between the identified results and the actual results.

[0174] The recognition accuracy is given by the following formula: (10); in, It is the number of correctly recognized patterns. That is the total number of samples.

[0175] F1. The score is a metric that balances precision and recall using a weighted average method.

[0176] Precision represents the number of samples that are identified as positive but are actually positive, while recall represents the proportion of samples that are identified as positive but are actually positive out of the total number of positive samples.

[0177] Precision and recall are respectively used as and Indicate, then F 1. A fraction is represented as: (11).

[0178] (3) Formation pattern recognition based on the proposed method: This section will describe the identification results of the method proposed in this invention.

[0179] First, GMM is used to identify vertical feature pattern recognition subtasks that include assembly and disassembly.

[0180] Vertical feature pattern recognition results are as follows Figure 6 As shown, Figure 6 The visualization results from parameter mapping and pattern labeling are shown. For example... Figure 6 As shown in (a), the recognition results are mapped to altitude. The yellow line represents the assembly mode at the highest point, the green line represents the cruise mode, and the blue line represents the disbandment mode with the altitude gradually decreasing. Figure 6 (b) demonstrates that the pattern recognition labels closely approximate the actual patterns, indicating the accurate recognition by the GMM in the vertical feature pattern recognition subtask. Label "2" represents the assembly pattern; label "3" represents the disbanding pattern; and label "4" represents the cruise pattern. Quantitative results show that the accuracy of the vertical feature pattern recognition subtask is 99.89%. F The score was 0.997, which verifies the effectiveness of GMM in vertical feature pattern recognition.

[0181] Then, the proposed STReC model is used to identify the consistent feature pattern recognition subtask, and a density-driven recognition optimization method is employed to optimize the results. The results are as follows: Figure 7 As shown, it is mapped onto the formation trajectory; to illustrate this more clearly, the entire formation process is divided into four parts to reduce the impact of trajectory overlap on the interpretation of the results; the red triangles in the figure represent 20 drones and their directions of movement; from rectangles to triangles to rhombuses, the changes in formation shape are clearly visible; the green trajectory represents the consistent pattern, and the other trajectories represent the inconsistent pattern.

[0182] Figure 8 The results of the overall identification are clearly shown, with "0" representing an inconsistent pattern and "1" representing a consistent pattern.

[0183] from Figure 7 and Figure 8 It can be seen that both consistent and inconsistent patterns were almost accurately identified.

[0184] However, due to the fuzzy features of the transition segment, the boundaries are unclear, thus the figure shows a pattern of incorrect identification. Quantization results show that the accuracy rates of the three cluster formations are 95.34%, 94.41%, and 94.88%, respectively. F The scores were 0.970, 0.965, and 0.960 respectively. Accuracy and F The average scores were 94.88% and 0.965, respectively.

[0185] Although the method proposed in this invention can identify consistent patterns to a large extent, small errors still occur because it cannot fully represent the transition between consistent and inconsistent patterns.

[0186] (4) Comparative experiments: In order to verify the effectiveness and performance of the proposed method, it will be further compared with other traditional and advanced methods.

[0187] This invention compares the state words generated by K-means, GMM, Improved Bayesian GMM (IBGMM), Hydride K-means and hierarchical (HyKH) clustering methods with those generated by UAV swarm formation rules, and adopts... and F 1. Score evaluation indicator.

[0188] To ensure a fair comparison, this invention applies MD, local clustering for each formation, and identification optimization methods to the comparison method.

[0189] As can be seen from Table 2, the present invention achieves high accuracy and... F It performed well in terms of scores, and the IBGMM and HyKH methods also performed well.

[0190] However, compared with the present invention, the HyKH method has an accuracy that is 9.39% lower. F The IBGMM method scored 0.061 lower than HyKH, indicating it performed worse. This is because IBGMM's uncertainty representation often dilutes the probability density in noisy transition regions, leading to misclassification. HyKH, on the other hand, excessively smooths inconsistencies and uneven distributions in transition segments, resulting in feature blurring and reduced accuracy in identifying inconsistencies. Furthermore, both methods are geared towards static distributions, leading to insufficient representation of key spatiotemporal coupling dynamics, thus degrading recognition performance.

[0191] This invention solves the above problems by fully representing the spatiotemporal characteristics of cluster formation data.

[0192] The results show that the accuracy of the comparison method is higher in clusters 2 and 3. F The scores of all other models showed significant instability. In contrast, the present invention consistently maintained a performance of over 94%, demonstrating excellent robustness across various scenarios.

[0193] Table 2 Performance evaluation results of the comparison methods for cluster formation operation mode recognition

[0194] The above description of several specific embodiments further details the technical solution provided by the present invention in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, reasonable combinations of embodiments, and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive recognition method for UAV swarm formation operation modes, characterized in that, The method includes the following steps: Step S1: Based on UAV formation control knowledge, the global formation task of the UAV swarm is decomposed into a knowledge-driven hierarchical structure, and multiple formation operation mode categories with unified and clear semantics are defined. The formation operation mode categories include: assembly mode, disband mode, consistency mode and inconsistency mode. Step S2: Obtain the formation data of the UAV cluster; the formation data includes formation command data issued by the ground station and flight status time sequence data of each UAV; Step S3: Based on the formation operation mode category defined in Step S1, perform pattern recognition on the formation data: For the assembly and disbandment modes, the first clustering model is used to identify them based on the altitude data in the flight status time series data to obtain the vertical feature identification results. For the consistent and inconsistent patterns, a spatiotemporal representation clustering model is used to identify them based on the relative position data between the local drone and the lead drone in the flight state time series data, and a preliminary identification result is obtained. Step S4: Optimize the preliminary identification results and output the final formation operation mode sequence.

2. The adaptive recognition method for UAV swarm formation operation mode according to claim 1, characterized in that, The hierarchical decomposition in step S1 includes the following three levels of decomposition: First-level decomposition: Based on the formation instructions in the formation instruction data, the global formation task is divided into multiple continuous formation intervals; The second layer of decomposition: Within each formation interval, based on the dominant movement direction characteristics of the formation behavior, the vertical feature recognition stage and the horizontal feature recognition stage are divided. The vertical feature recognition stage corresponds to the assembly mode and the disbanding mode; The third level of decomposition: all formation coordination processes in the horizontal feature recognition stage are classified into two categories: consistent mode and inconsistent mode.

3. The adaptive recognition method for UAV swarm formation operation mode according to claim 1, characterized in that, In step S3, the first clustering model is a Gaussian mixture model.

4. The adaptive recognition method for UAV swarm formation operation mode according to claim 1, characterized in that, In step S3, the identification using a spatiotemporal representation clustering model includes: Step S3.1: For the formation type to which the current formation interval belongs, obtain a pre-trained Hidden Markov Model for that formation type; Step S3.2: Based on the relative position data within the current formation interval, calculate the Mahalanobis distance sequence; Step S3.3: Use the pre-trained Hidden Markov Model to decode the Markov distance sequence to obtain the corresponding hidden state sequence; Step S3.4: Based on the preset mapping relationship between hidden states and formation operation mode categories, convert the hidden state sequence into a sequence of consistent or inconsistent modes as the preliminary identification result of the current formation interval.

5. The adaptive recognition method for UAV swarm formation operation mode according to claim 4, characterized in that, The pre-trained Hidden Markov Model is obtained through the following steps: For each formation that appears in the task, execute: Step X1: Extract the relative position data from the training data corresponding to the formation and calculate the Mahalanobis distance sequence. The training data includes sample sequences known to be consistent patterns and known to be inconsistent patterns. Step X2: Construct a hidden Markov model with Gaussian emission probability; Step X3: Based on the Mahalanobis distance sequence, train the Hidden Markov Model using an iterative optimization algorithm. Calculate the log-likelihood value of the training data in each iteration. When the training reaches a preset number of iterations or the change in the log-likelihood value satisfies the convergence condition, stop training to obtain the pre-trained Hidden Markov Model.

6. The adaptive recognition method for UAV swarm formation operation mode according to claim 1, characterized in that, The optimization process in step S4 is a density-driven recognition optimization, including: Step S4.1: Perform differential operations on the sequence data of the preliminary pattern recognition results to locate the mode switching point; Step S4.2: Set up a sliding window centered on the mode switching point, and calculate the switching density of the mode label in each sliding window; Step S4.3: Correct the mode labels in the sliding window with a switching density higher than the first threshold to inconsistent modes; Step S4.4: Smooth isolated pattern label segments with durations shorter than the second threshold into their adjacent mainstream pattern labels.

7. An adaptive recognition device for unmanned aerial vehicle (UAV) swarm formation operation mode, characterized in that, The device includes the following modules: Module S1: Based on UAV formation control knowledge, the global formation task of the UAV swarm is decomposed into a knowledge-driven hierarchical structure, defining multiple formation operation mode categories that are unified and semantically clear. The formation operation mode categories include: assembly mode, disband mode, consistency mode and inconsistency mode. Module S2: Acquires the formation data of the UAV cluster; the formation data includes formation command data issued by the ground station and flight status time sequence data of each UAV; Module S3: Based on the formation operation mode category defined in module S1, perform pattern recognition on the formation data: For the assembly and disbandment modes, the first clustering model is used to identify them based on the altitude data in the flight status time series data to obtain the vertical feature identification results. For the consistent and inconsistent patterns, a spatiotemporal representation clustering model is used to identify them based on the relative position data between the local drone and the lead drone in the flight state time series data, and a preliminary identification result is obtained. Module S4: Optimizes the preliminary identification results and outputs the final formation operation mode sequence.

8. A computer device, comprising: The processor and memory are characterized in that the memory is used to store executable instructions of the processor, the processor being configured to execute the adaptive recognition method for unmanned aerial vehicle swarm formation operation mode according to any one of claims 1-6 by executing the executable instructions.

9. A computer storage medium, characterized in that, The storage medium stores a computer program, which, when executed, performs the adaptive recognition method for UAV swarm formation operation mode as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the adaptive recognition method for drone swarm formation operation mode as described in any one of claims 1-6.

Images