Multi-uav distributed cooperative modeling method based on lightweight convex input time convolution

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) system modeling technology, and discloses a multi-UAV distributed collaborative modeling method based on lightweight convex input temporal convolution. The method includes: each UAV node performs state estimation using a distributed observer based on its own measurement output and communication information from neighboring nodes during runtime; UAV nodes exchange state estimation information through an event-triggered communication mechanism, broadcasting only when preset trigger conditions are met; the inner network parameters of the identifier are trained offline, and the outer weight matrix of the identifier is designed with an online adaptive update law based on Lyapunov stability analysis and maintained bounded using a projection operator; this invention achieves a comprehensive technical effect—simultaneously lightweight airborne model computation, significantly reduced inter-node communication overhead, and model output that is friendly (fast) to subsequent optimization solutions—while ensuring consistent and eventually bounded distributed state estimation errors (i.e., system stability).

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) system modeling technology, specifically to a distributed collaborative modeling method for multiple UAVs based on lightweight convex input temporal convolution. Background Technology

[0002] Unmanned aerial vehicles (UAVs) have been widely used in emergency rescue, environmental monitoring, security patrols, and mapping due to their flexible deployment, high maneuverability, and relatively low cost. In these tasks, a single UAV typically needs to achieve stable flight, path planning, and payload execution in complex environments, thus placing high demands on the accuracy of the dynamic model and the reliability of online status information. However, UAV systems generally exhibit strong nonlinearity, strong coupling, and significant uncertainties, and are susceptible to factors such as wind disturbances, load variations, actuator nonlinearity, sensor noise, and unmodeled dynamics, making UAV dynamic modeling extremely challenging.

[0003] Existing UAV dynamics modeling methods mainly include mechanistic modeling and data-driven modeling. Mechanism modeling relies on physical priors and parameter identification, and is limited by environmental changes and parameter availability, often only yielding approximate models that fail to cover complex dynamics. Data-driven modeling utilizes deep neural networks to learn dynamic mappings from flight data, which can improve modeling accuracy to some extent, but still faces the following problems in practical applications: First, deep neural networks have complex structures and a large number of parameters, requiring high airborne computing resources and energy consumption, making it difficult to support high-frequency online computation; second, deep networks typically have non-convex mappings from input to output, limiting the speed of subsequent model-based online solutions or optimization calculations; third, the online training update laws of deep neural networks usually lack stability analysis support, posing a risk of insufficient reliability and making them unsuitable for applications in scenarios with high safety requirements.

[0004] As task scale and complexity increase, single-machine execution is gradually evolving into multi-machine collaborative operation. Multi-UAV systems, through distributed collaboration, can improve coverage and task parallelism, enhance system robustness, and reduce dependence on the performance of a single platform. In multi-UAV scenarios, each UAV node often only acquires local information or local measurement results, requiring neighborhood communication and consistency mechanisms to collaboratively reconstruct the system state or dynamic model. Compared to centralized solutions, distributed collaboration eliminates the need for a central fusion node, better adapts to constraints such as limited computing power and unstable communication links at each node, and offers better scalability and fault tolerance.

[0005] However, the key challenges of distributed collaborative modeling of multiple UAVs are as follows: First, unknown nonlinear dynamics and external disturbances are more difficult to accurately approximate in a multi-node distributed environment, and model errors may accumulate during the collaborative process and affect consistency. Second, wireless communication bandwidth and energy are limited. If periodic high-frequency communication is used to share model or state information, it will cause excessive communication overhead and network congestion, affecting system endurance and scalability. Third, if the parameters of deep neural networks are directly updated in online collaborative learning, stability risks are easily introduced, and there is a lack of verifiable bounded error guarantees.

[0006] Therefore, a distributed deep collaborative modeling method for multi-UAV systems is needed to improve the approximation accuracy of uncertain nonlinear dynamics while ensuring the stability of online updates of deep neural networks, reduce the communication burden through event-triggered communication, and take into account lightweight deployment and subsequent solution friendliness, so as to meet the comprehensive requirements of real-time performance, reliability and scalability in practical applications. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a distributed collaborative modeling method for multiple UAVs based on lightweight convex input temporal convolution.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A distributed collaborative modeling method for multiple UAVs based on lightweight convex input temporal convolution includes: Each drone node is constructed as an identifier, which includes a lightweight input module based on content-aware downsampling, as well as a temporal convolution module and a decoupled output module based on a convex input constraint strategy. During operation, each UAV node uses a distributed observer to estimate its state based on its own measurement output and communication information from neighboring nodes. The distributed observer includes a local dynamics approximation term implemented by the identifier and a consistency term based on neighborhood state estimation information. The drone nodes exchange state estimation information through an event-triggered communication mechanism, and broadcast only when preset triggering conditions are met. A multi-timescale learning strategy is adopted: the inner network parameters of the identifier are trained offline, and the outer weight matrix of the identifier is designed with an online adaptive update law based on Lyapunov stability analysis and the projection operator is used to keep it bounded, so that the global state estimation error is consistent and eventually bounded.

[0010] In one embodiment, the process of constructing the content-aware downsampling-based lightweight input module specifically includes: For the input state sequence of the drone node Perform fixed-interval sampling, and reduce the length of the state sequence from Reduce to The downsampled state subsequence is obtained. : ; Among them, intermediate variables ,Require able to be Divisible by integers, each state vector in the state sequence includes the location of the UAV node. ,speed Attitude rotation matrix angular velocity Control input and corresponding timestamp ; The length of the input state sequence; The length of the downsampled state subsequence; for The state vector at each time step; To ensure the impact of reduced input information after downsampling, calculation Forward difference of the state vector at each time step and backward difference To provide local content awareness information; Will , and By concatenating the sequences, a content-aware short sequence is obtained. : ; This is for splicing operations.

[0011] In one embodiment, the process of constructing the temporal convolution module and decoupled output module based on the convex input constraint strategy specifically includes: The temporal convolution module consists of It is composed of stacked convolutional layers that occur at the same time, the first one being... A temporal convolutional layer is represented as: ; ; in, For the first Layer intermediate features; For the first Layer output features; For the first Layer output features; For the first Layer-time convolution operator; when hour, For the input sequence Features obtained from the initial embedding layer; For batch normalization operators; For the first Layer convolution kernel; For the first Layer bias; This represents a one-dimensional convolution operation; These are residual connection coefficients or residual connection operators; The decoupled output module aggregates features and outputs them through skip connections: ; ; in, These are intermediate features aggregated through skip connections. This is the final output of the identifier. The activation function is a linear rectified function. For the first Convolution kernel with skip branches, For the first The bias of the jump branch. To decouple the convolution kernel or linear mapping kernel of the output module, For the corresponding bias; Both the temporal convolutional module and the decoupled output module employ neural networks. The weight parameters of both modules are constrained to be non-negative, and the activation function is chosen to be a convex and non-decreasing function, thus processing content-aware short sequences. Expand to To address the issue of negative weights.

[0012] In one embodiment, the identifier is trained offline using a multi-step loss function: ; in, The set of parameters for the identifier. For the training dataset; The length of the future state sequence of the drone nodes generated recursively; For the future of real drone nodes The state vector at each time step; For the predicted future drone node The state vector at each time step It is the F2 norm.

[0013] In one embodiment, the distributed observer includes a local dynamics approximation term implemented by the identifier and a consistency term based on neighborhood state estimation information, specifically including: The state equation of any modeled UAV node in a multi-UAV system: ; in, The state vector of the drone node; For unknown nonlinear dynamic mapping; External disturbance; State vector Regarding time The first derivative; No. Continuous measurement output of each drone node for: ; For the first The output matrix of each drone node; Approximation structure is used within compact sets: ; The ideal outer weight matrix; It is a vector function consisting of continuous bounded activation functions; This is an inner-layer feature mapping based on a distinguisher. Reconstruction error for bounded functions; For transpose; The state equations are equivalent to: ; For dynamic approximation terms; Definition of the first State estimation of each drone node Error in state estimation: ; This represents the state estimation error; Zero-order hold-the-state sampling estimation , ; For the first The drone node Next broadcast time; Define output estimate and output error: , ; For output estimation; For the output error; therefore: ; This is an equivalent representation of the output error; Define consistency terms: ; This is a consistent item; For the first The set of neighbors of each drone node; These are elements of the adjacency matrix; Define a distributed observer: ; ; in, For the estimated first The outer weight matrix of each drone node; For the first The estimated inner mapping of the i-th drone node is derived from the i-th The identification system is implemented by maintaining each drone node. The observer gain matrix is... express Rate of change over time.

[0014] In one embodiment, a global output matrix is ​​defined as follows: ; This represents the global output matrix of a sensor network composed of UAV nodes. This indicates the construction of a diagonal matrix; the gain matrix Choose a symmetric matrix that satisfies the following bilinear matrix inequality: ; in, The Laplace matrix of the communication graph; For Kronecker product; For design parameters; for An identity matrix of order 1. for An identity matrix of order 1. Indicates the number of drone nodes. The dimension of the state vector of a single drone node.

[0015] In one embodiment, the UAV nodes exchange state estimation information through an event-triggered communication mechanism, broadcasting only when preset triggering conditions are met, specifically including: A broadcast is triggered when the square norm of the sampling error of the i-th drone node satisfies a certain relationship with the square norm of the consistency term. Specifically, the event triggering mechanism is defined as follows: ; For trigger parameters; This refers to the number of drone nodes; It is a positive constant; Sampling error: ; The infimum; Trigger parameter selection method: , ; in, For design parameters; It is the induced norm of the matrix; The Laplace matrix of the communication graph; For Kronecker product; for An identity matrix of order 1.

[0016] In one embodiment, the inner network parameters of the identifier are trained offline, specifically including: The i-th drone node performs the following operations in each offline training time sequence: for the i-th... Estimated inner mapping of each drone node The parameters are trained and updated offline, and a switching operation is performed at the end of training to replace the original estimated inner layer mapping with the updated estimated inner layer mapping. It is a piecewise continuous function.

[0017] In one embodiment, the outer weight matrix of the identifier is designed with an online adaptive update law based on Lyapunov stability analysis and is kept bounded using a projection operator, specifically including: Define outer layer weight error ; This represents the ideal outer weight matrix of the identifier; The first estimated by the identifier The outer weight matrix of each drone node; The online update law for outer layer weight error is: ; ; ; in, This is the vectorized result of the outer weight matrix; The learning rate matrix is ​​positive definite. This is a projection operator used to constrain the outer weight matrix within a bounded set; For vectorization operators; To update the direction item, Let be a vector function consisting of continuous bounded activation functions. For the first Estimated inner layer mapping for each drone node, For output error, For the first The output matrix of each drone node express Regarding time The first derivative.

[0018] Compared with the prior art, the beneficial technical effects of the present invention are: 1. Existing methods typically employ deep neural networks directly for UAV modeling, failing to adequately consider the limitations of onboard computing resources, which can lead to slow model inference speeds or even deployment failures. This invention proposes a lightweight input module based on content-aware downsampling, reducing input sequence length and computational load while maintaining information sufficiency.

[0019] 2. Existing methods often construct deep neural network models that are non-convex mappings, making it difficult to meet the requirements of rapid solution for subsequent control and optimization design. This invention proposes a convex input constraint strategy for temporal convolutional networks. By using non-negative weight constraints, convex and non-decreasing activation functions, and positive and negative input expansion, the model satisfies the input convexity requirement, thereby improving the efficiency of subsequent solution.

[0020] 3. Existing distributed collaborative modeling and estimation methods often employ periodic communication to share information, resulting in high communication frequency and large bandwidth consumption. This invention adopts an event-triggered communication mechanism, broadcasting state estimation information only when triggering conditions are met, thereby improving network resource utilization and reducing communication overhead.

[0021] 4. Existing online learning methods for deep networks lack stability guarantees. This invention employs a multi-timescale learning strategy: inner layer network parameters are trained offline, while outer layer weights are trained online using an update law based on Lyapunov stability analysis combined with a projection operator to ensure boundedness, thereby obtaining a consistent eventually bounded (UUB) error conclusion. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is an architectural diagram of the identifier of the present invention. Detailed Implementation

[0023] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figure 1 As shown, this invention provides a distributed collaborative modeling method for multiple UAVs based on lightweight convex input temporal convolution, comprising the following steps: S1, each drone node is constructed as an identifier, which includes a lightweight input module based on content-aware downsampling, as well as a temporal convolution module and a decoupled output module based on a convex input constraint strategy; S2, each UAV node performs state estimation using a distributed observer based on its own measurement output and communication information from neighboring nodes during operation. The distributed observer includes a local dynamics approximation term implemented by the identifier and a consistency term based on neighborhood state estimation information. S3: The UAV nodes exchange state estimation information through an event-triggered communication mechanism and broadcast only when preset triggering conditions are met. S4. A multi-timescale learning strategy is adopted: the inner network parameters of the identifier are trained offline, and the outer weight matrix of the identifier is designed with an online adaptive update law based on Lyapunov stability analysis and the projection operator is used to keep it bounded, so that the global state estimation error is consistent and eventually bounded.

[0025] Here, the inner network refers to the network used to generate inner mappings (such as...). The network part corresponds to a feature extractor consisting of a lightweight input module and a temporal convolution module; the decoupled output module maps the extracted features to the dynamic estimation output.

[0026] The following is a detailed introduction in several parts.

[0027] 1. Data preparation and preprocessing.

[0028] Flight data acquisition was performed using the Gym-pybullet-drones simulation platform. The UAV's geometric parameters, mass attributes, inertial parameters, and collision model were configured using URDF format files. Multiple flight missions were executed in the simulation environment, including 20 motion modes such as hovering, straight flight, circular trajectory, figure-eight trajectory, and spiral ascent. The acquisition frequency was set to 100Hz, and the recorded status included position. ,speed Attitude rotation matrix angular velocity Control input and corresponding timestamp .

[0029] To enhance the robustness of the model, a Gaussian-distributed random perturbation force was applied to the centroid in the simulation, and noise was superimposed at the propeller output to simulate actuator fluctuations.

[0030] The dataset was divided into training, validation, and test sets in a 6:2:2 ratio, and all features were Z-score standardized. ,in and These are the mean and standard deviation of the training set, respectively, to eliminate the influence of dimensions and accelerate convergence.

[0031] 2. Construct a lightweight convex temporal input convolutional identifier.

[0032] This step constructs a lightweight convex-input temporal convolution (LCTC) discriminator, which consists of three modules: a lightweight input module based on content-aware downsampling, a temporal convolution module based on a convex input constraint strategy, and a decoupled output module. The input convexity is satisfied through constraints such as non-negative weights and convex non-decreasing activation.

[0033] (1) Construct a lightweight input module based on content-aware downsampling: To reduce the model's computational resource requirements, the state sequence of the original input from the UAV nodes is processed. Perform fixed-interval sampling, and shorten the sequence length. Reduce to (Preferred settings) ), to obtain the downsampled subsequence : ; Among them, intermediate variables ,Require able to be Divisible The original input state sequence of the UAV node; The sequence length; The length of the downsampled sequence; The input vector for each time step.

[0034] To ensure the impact of reduced input information after downsampling, calculation Forward difference of the state vector at each time step and backward difference To provide local content awareness information; to avoid information leakage, the forward difference of the first time step is set to 0. Among them, It is a forward difference sequence; It is a backward difference sequence.

[0035] The downsampled state sequence With local content-aware information and By splicing, a short sequence is obtained. : ; in, For content-aware short sequences; This is for splicing operations.

[0036] (2) Construct a temporal convolution module and an output module based on a convex input constraint strategy: Temporal convolution module: Stacked convolutional layers at the same time (preferred) ), No. Layers are represented as: ; ; in, For the first Layer intermediate features; For the first Layer output features; For the first Layer output features; For the first Layer-time convolution operator; when hour, For the input sequence Features obtained from the initial embedding layer; For batch normalization operators; For the first Layer convolution kernel; For the first Layer bias; This represents a one-dimensional convolution operation; These are the residual connection coefficients or residual connection operators.

[0037] The temporal convolutional layer consists of three stacked one-dimensional causal convolutional layers. To obtain a sufficiently large receptive field, the number of output channels of the three causal convolutional layers is preferably set to 512, 256, and 256; the kernel size is set to 3; and the dilation factor is set to 2.

[0038] Decoupled output module: Aggregates features and outputs them through skip connections: ; ; in, This is an aggregation feature; For network output; It is a linear rectified function; For the first Convolution kernel with skipped branches; For the first Bias of the jump branch; For output module convolution kernels or linear mapping kernels; This corresponds to the bias.

[0039] To simplify learning and enhance modularity, the system dynamics are decoupled into velocity submodules and attitude submodules; each decoding output module contains three fully connected layers, with the number of neurons preferably set to 512, 256, and 256.

[0040] Convex input constraint strategy: The weight parameters of the temporal convolutional and output modules are constrained to be non-negative; the activation function is selected as a convex and non-decreasing function (ReLU is preferred); content-aware short sequences are used. Expand to To address the issue of negative weights.

[0041] Both the temporal convolution module and the decoupled output module belong to the neural network structure and contain trainable parameters. The weight parameters of the temporal convolution module are the convolutional kernels of each layer. With bias The weight parameters of the decoupled output module are the skip-connection branch convolution kernels. Bias and output mapping kernel With bias .

[0042] 3. Model training and optimization based on multi-step loss.

[0043] To reduce the cumulative error of long-term forecasts, a multi-step loss function is adopted: ; in, For network parameter set; For the entire dataset; The length of the future state sequence generated recursively; For the future of real drone nodes The state vector at each time step; For the predicted future drone node The state vector at each time step.

[0044] The Adam optimizer is used, with the learning rate preferably set to 0.001, the batch size preferably set to 64, the training period to 150, and an early stopping strategy (with a tolerance of preferably 25) set to prevent overfitting.

[0045] 4. Design of a distributed collaborative modeling and event-triggered communication framework for multiple UAVs.

[0046] (1) Multi-UAV system model and measurement model: Consider the state equation of any modeled UAV in a multi-UAV system: ; in, This is the UAV state vector; For unknown nonlinear dynamic mapping; External disturbances The dimension of the state vector of a single drone node.

[0047] No. Continuous measurement output from each drone node: ; in, For output measurement; For the output matrix, Represents the measurement output vector The dimension of.

[0048] (2) Dynamic approximation structure based on lightweight convex input temporal convolutional discriminator: Approximation structure is used within compact sets: ; in, The ideal outer weight matrix; It is a vector function consisting of continuous bounded activation functions; This is the inner layer feature mapping implemented based on the lightweight convex input temporal convolutional discriminator constructed in step two; For the reconstruction error of a bounded function, express dimensionality This represents the feature dimension of the inner mapping (feature mapping) output. A "compact set" refers to a closed and bounded set in Euclidean space. Assume the system state... Satisfy throughout the entire operation ,in The function is compact, thus reducing the function reconstruction error. And the related perturbation terms remain bounded.

[0049] Inner layer feature mapping The high-dimensional eigenvectors output by the LCTC network are used to capture the complex nonlinear characteristics of the system dynamics; the outer weight matrix... These are linear weighting coefficients used to map inner layer features to state space derivatives.

[0050] The system equations are equivalent to: ; in, This is the dynamic approximation term.

[0051] (3) State estimation, error definition and zero-order hold-and-sampling broadcast strategy: Definition of the first State estimation of each drone node Error in state estimation: ; in, This represents the state estimation error.

[0052] Define the event time sequence Zero-order hold-the-moment sampling state estimation is used: , ; in, For sampling state estimation; For the first The node Next broadcast time.

[0053] Define sampling error: ; in, This represents the sampling error.

[0054] Define output estimate and output error: , ; in, For output estimation; This is the output error; therefore, we have ; in, This is an equivalent representation of the output error.

[0055] (4) Design of distributed consensus terms based on communication topology: Define consistency terms: ; in, This is a consistent item; For the first The set of neighbors of each node; These are the elements of the adjacency matrix.

[0056] (5) Design of distributed observers based on deep neural network approximation: Define a distributed observer: ; ; in, To estimate the first The outer weight matrix of each drone node; For the first The estimated inner mapping of each node (by the nodes) (The LCTC network under maintenance is implemented); is the observer gain matrix.

[0057] Define the global output matrix: ; in, This is the global output matrix of the sensor network. A sensor network refers to a network composed of... A measurement / communication network consisting of several UAV nodes; each UAV node is both an actuator and a sensor node, capable of measuring outputs. It also communicates with neighbors to share estimation information.

[0058] Gain matrix Choose a symmetric matrix that satisfies the following bilinear matrix inequality: ; in, The Laplace matrix of the communication graph; For Kronecker product; For design parameters; for An identity matrix of order 1.

[0059] (6) Design of an adaptive online update law for outer weights based on Lyapunov theory Define outer layer weight error: ; in, This represents the outer layer weight error. The ideal outer weight matrix; This is the estimated outer weight matrix.

[0060] The online update law for outer weights is: ; ; ; in, This is the vectorized result of the outer layer weights; The learning rate matrix is ​​positive definite. This is a projection operator (used to constrain weights within a bounded set); For vectorization operators; To update the direction item.

[0061] (7) Design of offline training strategy and online switching mechanism for inner layer mapping: No. The sequence of nodes during offline training Perform the following operations: estimate the inner mapping. The parameters are trained and updated offline, and at the time of training completion... Perform a switch operation to replace the original mapping with the updated inner mapping, so that... It is a piecewise continuous function.

[0062] Inner mapping The weights and biases of the LCTC network are defined in the sequence of offline training time steps. Updated at the outer layer; outer weights It is a linearly weighted coefficient matrix, which is continuously updated online through projective gradient descent.

[0063] 5. Establishment of error dynamics and derivation of global stacking expression.

[0064] (1) Definition of combinatorial terms and derivation of error dynamics: Define a combinatorial item: ; in, It is a combination item; External disturbance; This represents the function reconstruction error.

[0065] Substituting the definition from step four, we obtain the equivalent expression for the consistency term: ; The above equation is the analytical equivalent form of the consistency term, which is used for subsequent derivations.

[0066] Observer state estimation Differentiate, and combine with system dynamics Based on the definition in Part 4, the error dynamics can be obtained: ; in, This is the derivative of the state estimation error.

[0067] (2) Definition of global stacked variables Define stacked variables: ; ; ; ; ; ; in, A stacked vector of global state estimation errors; This is a stacked vector of global sampling errors; This is a stacked vector of global output errors. Estimate the stacking vector for the global state; A stacked vector of combined terms; Stacked vectors for consistency terms.

[0068] Define a block diagonal matrix: ; ; ; in, This is the diagonal matrix of the outer layer weight error block; Estimate the block diagonal matrix for the outer weights; This is the learning rate block diagonal matrix.

[0069] Define stacking characteristics: ; ; in, For feature stacking vectors; For nodes eigenvectors.

[0070] (3) Derivation of the relationship between the global error dynamics expression and the consistency term Stacking the single-node error dynamics yields the global error dynamics: ; in, This is the derivative of the global error; for An identity matrix of order 1.

[0071] The consistency terms satisfy: ; in, for An identity matrix of order 1.

[0072] Taking the norm of both sides of the above equation and applying Young's inequality, we get: ; in, Let be the Euclidean norm of the vector.

[0073] 6. Design of an event-triggered communication mechanism based on sampling error and consistency terms.

[0074] (1) Event triggering condition design: To reduce communication overhead and ensure system stability, an event triggering mechanism is designed as follows: when a node... Broadcasting is triggered when the square norm of the sampling error and the square norm of the consistency term satisfy a certain relationship. Specifically, the event triggering mechanism is defined as follows: ; in, For nodes The The next event moment; For trigger parameters; The number of nodes; It is a positive constant (used to avoid triggering infinitely many times). The term "infimum" is used in event triggering mechanisms to define the next trigger time as the earliest time point when the triggering condition is met. That is, the search starts after the current trigger time, and a broadcast is triggered once the inequality is true for the first time.

[0075] (2) Trigger parameter design and description Trigger parameters: ; ; in, For design parameters; It is the induced norm of the matrix; The Laplace matrix of the communication graph; For Kronecker product; for An identity matrix of order 1.

[0076] Trigger parameters By balancing the impact of sampling error with the strength of the consistency term, communication is ensured to be triggered only when necessary. Parameters It can be derived through Lyapunov stability analysis to ensure that the global error is consistent and eventually bounded.

[0077] 7. Stability analysis and consistent final bounded conclusion based on Lyapunov theory.

[0078] (1) Construction of Lyapunov functions: Define Lyapunov function candidates: ; ; in, It is a Lyapunov function; The trace of the matrix; for The inverse matrix; This is a combined state vector.

[0079] And there exists a constant. Make: ; in, This is the upper bound constant for the outer layer weight error term, which can be guaranteed by the projection operator.

[0080] (2) Lyapunov function derivative analysis and uniform final bounded conclusion: Along error dynamics Taking the derivative and combining the projection gradient update law with the event triggering mechanism, we can derive the following: The communication graph is connected, the perturbation is bounded, and there exists a constraint satisfying the bilinear matrix inequality. Under the condition of operating according to the outer weight update law and the event triggering mechanism, there exists a constant. , , Make: ; in, This is the convergence rate-related constant; This is a constant related to the perturbation and approximation error.

[0081] The above equation shows that the global state estimation error Uniformly eventually bounded (UUB) means that after the initial transient, the error eventually converges to a bounded state. Within the defined boundaries.

[0082] (3) Conclusion on the minimum trigger interval to avoid Zeno triggering: Furthermore, the time difference between adjacent events at any node has a consistent positive lower bound: ; in, For nodes The upper bound constant of the rate of change of the sampling error; It is a positive constant.

[0083] This inequality guarantees that there is a strict positive lower bound on the interval between adjacent triggering times, thereby avoiding the Zeno phenomenon (i.e., infinitely many triggering times occurring within a finite time).

[0084] This invention proposes a lightweight convex input temporal convolutional structure based on content-aware downsampling to achieve deployable collaborative modeling of UAV dynamics.

[0085] This invention proposes a convex input constraint strategy for temporal convolutional networks, which enables the model to meet the input convexity requirement, facilitating subsequent rapid solution and control / optimization design.

[0086] This invention proposes a distributed event-triggered collaborative modeling framework for multi-UAV systems, which combines model learning with neighborhood consistency terms to achieve efficient collaborative modeling.

[0087] This invention proposes a multi-timescale learning method that combines online update of outer layer weights with offline training of inner layer networks, and provides conclusions on error consistency and eventual boundedness and avoidance of Zeno triggering.

[0088] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0089] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0090] Based on the description of the above method embodiments, the present invention also provides an apparatus. The apparatus may be a system (including a distributed system), software (application), module, component, server, client, etc., that uses the methods described in the embodiments of this specification, combined with necessary implementation hardware. Based on the same innovative concept, the apparatuses in one or more embodiments provided in this disclosure are as described in the following embodiments. Since the implementation schemes and methods for solving the problem by the apparatus are similar, the implementation of the specific apparatus in the embodiments of this specification can be referred to the implementation of the foregoing methods, and repeated details will not be repeated. As used below, the terms "module" or "module group" refer to a combination of software and / or hardware capable of implementing a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0091] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory including instructions that can be executed by a processor to perform the above-described method. The storage medium may be a computer-readable storage medium, for example, a ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0092] In one embodiment, a computer device is provided, which may be a server. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data used in the methods described above. The network interface of the computer device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the methods described above.

[0093] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and not restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0095] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A distributed collaborative modeling method for multiple UAVs based on lightweight convex input temporal convolution, characterized in that, include: Each drone node is constructed as an identifier, which includes a lightweight input module based on content-aware downsampling, as well as a temporal convolution module and a decoupled output module based on a convex input constraint strategy. During operation, each UAV node uses a distributed observer to estimate its state based on its own measurement output and communication information from neighboring nodes. The distributed observer includes a local dynamics approximation term implemented by the identifier and a consistency term based on neighborhood state estimation information. The drone nodes exchange state estimation information through an event-triggered communication mechanism, and broadcast only when preset triggering conditions are met. A multi-timescale learning strategy is adopted: the inner network parameters of the identifier are trained offline, and the outer weight matrix of the identifier is designed with an online adaptive update law based on Lyapunov stability analysis and the projection operator is used to keep it bounded, so that the global state estimation error is consistent and eventually bounded.

2. The multi-UAV distributed collaborative modeling method based on lightweight convex input temporal convolution according to claim 1, characterized in that, The process of constructing the lightweight input module based on content-aware downsampling specifically includes: For the input state sequence of the drone node Perform fixed-interval sampling, and reduce the length of the state sequence from Reduce to The downsampled state subsequence is obtained. : ； Among them, intermediate variables ,Require able to be Divisible by integers, each state vector in the state sequence includes the location of the UAV node. ,speed Attitude rotation matrix angular velocity Control input and corresponding timestamp ; The length of the input state sequence; The length of the downsampled state subsequence; for The state vector at each time step; To ensure the impact of reduced input information after downsampling, calculation Forward difference of the state vector at each time step and backward difference To provide local content awareness information; Will , and By concatenating the sequences, a content-aware short sequence is obtained. : ; This is for splicing operations.

3. The multi-UAV distributed collaborative modeling method based on lightweight convex input temporal convolution according to claim 2, characterized in that, The process of constructing the temporal convolution module and decoupled output module based on the convex input constraint strategy specifically includes: The temporal convolution module consists of It is composed of stacked convolutional layers that occur at the same time, the first one being... A temporal convolutional layer is represented as: ； ； in, For the first Intermediate features of the layer; For the first Layer output features; For the first Layer output features; For the first Layer-time convolution operator; when hour, For the input sequence Features obtained from the initial embedding layer; For batch normalization operators; For the first Layer convolution kernel; For the first Layer bias; This represents a one-dimensional convolution operation; These are residual connection coefficients or residual connection operators; The decoupled output module aggregates features and outputs them through skip connections: ； ； in, These are intermediate features aggregated through skip connections. This is the final output of the identifier. The activation function is a linear rectified function. For the first Convolution kernel with skip branches, For the first The bias of the jump branch. To decouple the convolution kernel or linear mapping kernel of the output module, For the corresponding bias; Both the temporal convolutional module and the decoupled output module employ neural networks. The weight parameters of both modules are constrained to be non-negative, and the activation function is chosen to be a convex and non-decreasing function, thus processing content-aware short sequences. Expand to To address the issue of negative weights.

4. The multi-UAV distributed collaborative modeling method based on lightweight convex input temporal convolution according to claim 1, characterized in that, The identifier is trained offline using a multi-step loss function. ； in, The set of parameters for the identifier. For the training dataset; The length of the future state sequence of the drone nodes generated recursively; For the future of real drone nodes The state vector at each time step; For the predicted future drone node The state vector at each time step It is the F2 norm.

5. A multi-UAV distributed collaborative modeling method based on lightweight convex input temporal convolution as described in claim 1, characterized in that, The distributed observer includes a local dynamic approximation term implemented by the identifier and a consistency term based on neighborhood state estimation information, specifically including: The state equation of any modeled UAV node in a multi-UAV system: ; in, The state vector of the drone node; For unknown nonlinear dynamic mapping; External disturbance; Represents the state vector Regarding time The first derivative; No. Continuous measurement output of each drone node for: ; For the first The output matrix of each drone node; Approximation structure is used within compact sets: ; The ideal outer weight matrix; It is a vector function consisting of continuous bounded activation functions; This is an inner-layer feature mapping based on a distinguisher. Reconstruction error for bounded functions; For transpose; The state equations are equivalent to: ; For dynamic approximation terms; Definition of the first State estimation of each drone node Error in state estimation: ; This represents the state estimation error; Zero-order hold-the-state sampling estimation , ; For the first The drone node Next broadcast time; Define output estimate and output error: , ; For output estimation; For the output error; therefore: ; This is an equivalent representation of the output error; Define consistency terms: ; This is a consistent item; For the first The set of neighbors of each drone node; These are elements of the adjacency matrix; Define a distributed observer: ； ； in, For the estimated first The outer weight matrix of each drone node; For the first The estimated inner mapping of the i-th drone node is derived from the i-th The identification system is implemented by maintaining each drone node. The observer gain matrix is... express Rate of change over time.

6. A multi-UAV distributed collaborative modeling method based on lightweight convex input temporal convolution as described in claim 5, characterized in that, Define the global output matrix: ; This represents the global output matrix of a sensor network composed of UAV nodes. This indicates the construction of a diagonal matrix; the gain matrix Choose a symmetric matrix that satisfies the following bilinear matrix inequality: ； in, The Laplace matrix of the communication graph; For Kronecker product; For design parameters; for An identity matrix of order 1. for An identity matrix of order 1. Indicates the number of drone nodes. The dimension of the state vector of a single drone node.

7. A multi-UAV distributed collaborative modeling method based on lightweight convex input temporal convolution as described in claim 5, characterized in that, The UAV nodes exchange state estimation information through an event-triggered communication mechanism, broadcasting only when preset triggering conditions are met, specifically including: Broadcasting is triggered when the square norm of the sampling error of the i-th drone node satisfies a certain relationship with the square norm of the consistency term; specifically, the event triggering mechanism is defined as follows: ； For trigger parameters; This refers to the number of drone nodes; It is a positive constant; Sampling error: ; The infimum; Trigger parameter selection method: ， ； in, For design parameters; It is the induced norm of the matrix; The Laplace matrix of the communication graph; For Kronecker product; for An identity matrix of order 1.

8. A distributed collaborative modeling method for multiple UAVs based on lightweight convex input temporal convolution as described in claim 1, characterized in that, The inner network parameters of the identifier are trained offline, specifically including: The i-th drone node performs the following operations in each offline training time sequence: for the i-th... Estimated inner mapping of each drone node The parameters are trained and updated offline, and a switching operation is performed at the end of training to replace the original estimated inner layer mapping with the updated estimated inner layer mapping. It is a piecewise continuous function.

9. A distributed collaborative modeling method for multiple UAVs based on lightweight convex input temporal convolution as described in claim 1, characterized in that, The outer weight matrix of the identifier is designed with an online adaptive update law based on Lyapunov stability analysis and maintains boundedness using a projection operator, specifically including: Define outer layer weight error ; This represents the ideal outer weight matrix of the identifier; The first estimated by the identifier The outer weight matrix of each drone node; The online update law for outer layer weight error is: ； ； ； in, This is the vectorized result of the outer weight matrix; The learning rate matrix is ​​positive definite. This is a projection operator used to constrain the outer weight matrix within a bounded set; For vectorization operators; To update the direction item, Let be a vector function consisting of continuous bounded activation functions. For the first Estimated inner layer mapping for each drone node, For output error, For the first The output matrix of each drone node express Regarding time The first derivative.

Images