Unmanned aerial vehicle return detection method and system based on AI identification

Abstract

The invention provides an unmanned aerial vehicle return detection method and system based on AI recognition, and relates to the technical field of image recognition, and the method comprises the steps: extracting multi-scale features through an improved moving window hierarchical converter model, obtaining target attitude information through the combination of cross-modal comparative learning and a dynamic routing capsule network, and obtaining target attitude information; an alignment feature representation is obtained by using a feature enhancement module and an adversarial feature alignment method, the alignment feature representation is input into a graph attention diffusion model to obtain an accurate state representation, and a state evolution trajectory is obtained in combination with an adaptive causal convolutional network and an ordinary differential equation network; posterior distribution is determined through a variational inference network and an autoregression flow model, a state sequence is predicted by using a neural process network and a dynamic point process model, an expert strategy is learned by using a hierarchical generative adversarial imitation learning network and a distributed trust domain evolutionary strategy algorithm, and optimal parameter configuration is obtained; and accurate prediction of the return state of the unmanned aerial vehicle is realized.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a drone return detection method and system based on AI recognition. Background Art

[0002] The rapid development of drone technology has led to its increasingly widespread application in various fields, such as aerial photography, logistics distribution, agricultural monitoring, etc. The autonomous return function of drones is a key link to ensure their safe operation and efficient operation. A reliable drone return system needs to accurately perceive the surrounding environment, identify the return target, and perform precise trajectory control, which puts higher requirements on drone autonomous navigation, target recognition, and state estimation technologies;

[0003] Existing drone return detection methods mainly rely on GPS positioning, image recognition and sensor fusion technology, but still have problems such as low recognition accuracy in complex environments, susceptibility to signal interference, and inability to robustly estimate and compensate for environmental changes and system errors.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a drone return detection method and system based on AI recognition, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a drone return detection method based on AI recognition, comprising:

[0007] The multi-source image data in the drone return process is obtained and added to the improved window-shifting hierarchical transformer model. The multi-scale feature maps are extracted through multiple attention windows and added to the cross-modal contrastive learning module. The semantic consistency features are obtained by calculating the mutual information in different views and added to the dynamic routing capsule network. The routing weights are updated by calculating the coupling coefficients between the feature vectors and a high-dimensional feature vector containing the target posture information is output. The high-dimensional feature vector is added to the feature enhancement module including the channel recalibration unit and the spatial recalibration unit. The adaptive weights are updated by calculating the global statistical information to obtain the enhanced feature vector. The enhanced feature vector is aligned according to the adversarial feature alignment method to minimize the maximum mean difference between different data distributions to obtain the aligned feature representation.

[0008] The aligned feature representation is added to a pre-built graph attention diffusion model and represented as a spatiotemporal graph structure, the spatiotemporal graph structure is denoised through a diffusion process to obtain an accurate state representation, the state representation is input into an adaptive causal convolutional network, multi-scale temporal features are extracted through a dilated convolutional layer with a variable receptive field, the state evolution trajectory is obtained by combining with the continuous-time dynamic system in the neural ordinary differential equation network, the posterior distribution of the state evolution trajectory is determined through a variational inference network and the posterior distribution is mapped to a standard normal distribution in combination with an autoregressive flow model to obtain a probability density, the probability density is added to a conditional neural process through a neural process network, the data distribution is modeled in combination with a context encoder and a target decoder, the time type and time are predicted in combination with a dynamic point process model, and a state prediction sequence containing uncertainty is obtained;

[0009] The state prediction sequence is added to the first discriminator of the hierarchical generative adversarial imitation learning network, the implicit distribution of the expert strategy is determined through adversarial training and learning, and the expert strategy is decomposed into sub-tasks at multiple abstract levels. The state transition diagram of the return task is constructed in combination with the option chain, the state evolution gradient is calculated by the dual proximal strategy optimization algorithm and added to the second discriminator based on neural architecture search, the optimal feature extraction structure is determined, the optimal feature extraction structure and the corresponding network parameters are updated by multiple parallel evaluation units in the distributed trust domain evolution strategy algorithm to obtain the optimal parameter configuration, the optimal parameter configuration is mapped to a multidimensional feature vector corresponding to the return state, the state prediction is performed through the multidimensional evaluation network, the uncertainty of the state prediction is determined by the sequential Monte Carlo method, and the final judgment result is obtained.

[0010] In an optional embodiment,

[0011] The multi-source image data in the drone return process is obtained and added to the improved window-shifting hierarchical transformer model. The multi-scale feature map is extracted through multiple attention windows and added to the cross-modal contrastive learning module. The semantic consistency feature is obtained by calculating the mutual information in different views and added to the dynamic routing capsule network. The routing weight is updated by calculating the coupling coefficient between the feature vectors and a high-dimensional feature vector containing the target posture information is output. The high-dimensional feature vector is added to the feature enhancement module including the channel recalibration unit and the spatial recalibration unit. The adaptive weight is updated by calculating the global statistical information to obtain the enhanced feature vector. The enhanced feature vector is aligned according to the adversarial feature alignment method to minimize the maximum mean difference between different data distributions. The alignment feature representation includes:

[0012] Acquire infrared images, visible light images, and depth images during the drone return process as multi-source image data;

[0013] The multi-source image data is input into a shifting window hierarchical transformer model, wherein the shifting window hierarchical transformer model comprises three layers of attention windows, wherein the first layer of attention window extracts target contour features, the second layer of attention window extracts target region features, and the third layer of attention window extracts environmental background features, and the features extracted by the three layers of attention windows are fused to obtain a multi-scale feature map;

[0014] The multi-scale feature map is input into a cross-modal contrastive learning module, which calculates the cosine similarity between feature vectors of feature maps of different modalities at the same position, constructs a mutual information maximization objective function based on the cosine similarity, and optimizes the mutual information maximization objective function to obtain semantic consistency features;

[0015] Inputting the semantic consistency feature into a dynamic routing capsule network, wherein the dynamic routing capsule network comprises a plurality of capsule layers, each of the capsule layers comprises a plurality of capsules representing different posture attributes, adjacent capsule layers are connected via a transformation matrix, the parameters of the transformation matrix are obtained by calculating the coupling coefficient between capsules and iteratively updating the routing weights, and the outputs of the lower-layer capsules are weightedly fused based on the routing weights to obtain a high-dimensional feature vector;

[0016] The high-dimensional feature vector is input into a self-calibration feature enhancement module, wherein the self-calibration feature enhancement module comprises a channel recalibration unit and a spatial recalibration unit, wherein the channel recalibration unit generates a channel weight vector through global average pooling and a fully connected layer, and the spatial recalibration unit generates a spatial weight matrix through global maximum pooling, and the channel weight vector and the spatial weight matrix are respectively weightedly fused with the high-dimensional feature vector to obtain an enhanced feature vector;

[0017] The enhanced feature vector is input into an adversarial feature alignment module, which includes a feature extractor and a domain discriminator. According to the adversarial feature alignment method, the enhanced feature vector is mapped to a common feature space through the feature extractor, the data domain to which the feature belongs is discriminated through the domain discriminator, and the maximum mean difference between the source domain feature and the target domain feature in the reproducing kernel Hilbert space is calculated based on the Gaussian kernel function, and the alignment feature representation is obtained by minimizing the maximum mean difference.

[0018] In an optional embodiment,

[0019] The mutual information maximization objective function based on the cosine similarity is constructed as shown in the following formula:

[0020] ;

[0021] Among them, J represents the objective function value, α(i, j) represents the attention weight of position (i, j), sim(x i,y j ) represents the eigenvector x of the two modes at position (i, j) i and j The cosine similarity between them, λ represents the balance factor, d(x i ,y j ) represents the positive sample pair (x i ,y j ), d(x i ,y k ) represents the negative sample pair (x i ,y k ), M represents the boundary value in the contrast loss, which is used to control the distance difference between positive and negative samples.

[0022] In an optional embodiment,

[0023] The aligned feature representation is added to the pre-built graph attention diffusion model and represented as a spatiotemporal graph structure. The spatiotemporal graph structure is denoised through a diffusion process to obtain an accurate state representation. The state representation is input into an adaptive causal convolutional network, and multi-scale temporal features are extracted through a dilated convolutional layer with a variable receptive field. The continuous-time dynamic system in the neural ordinary differential equation network is combined for solution, and the state evolution trajectory is obtained, including:

[0024] The aligned feature representation is input into a pre-built graph attention diffusion model, a spatiotemporal graph structure is constructed by setting state feature nodes of multiple time steps and edges of temporal dependencies between states, attention weights between nodes are calculated according to the graph attention layer, and node features are aggregated to obtain context node representations containing long-term and short-term dependencies;

[0025] A diffusion process is performed on the spatiotemporal graph structure through multiple diffusion steps, and in each diffusion step, Gaussian noise is added to the node features and the features are updated through a multi-layer graph neural network. The feature update process of the graph neural network is calculated based on the node features of the previous diffusion step, the Gaussian noise intensity, and the time interval of adjacent time steps. The inverse diffusion process is performed according to the iteratively updated node features to remove the noise, so as to obtain an accurate state representation;

[0026] The precise state representation is input into an adaptive causal convolutional network, the adaptive causal convolutional network is constructed by setting a plurality of dilated convolutional layers with different dilated rates, time series features of multiple scales are extracted according to the dilated convolutional layers with different dilated rates, and the time series features of the multiple scales are feature spliced ​​to obtain multi-scale time series features;

[0027] The multi-scale time series features are input into the neural ordinary differential equation network, and the right function of the ordinary differential equation is parameterized by setting a fully connected network with multiple hidden layers. The high-order Runge-Kutta method is used to perform numerical integration at multiple time points to solve the ordinary differential equation and obtain the state evolution trajectory.

[0028] In an optional embodiment,

[0029] The posterior distribution of the state evolution trajectory is determined by the variational inference network and the posterior distribution is mapped to the standard normal distribution in combination with the autoregressive flow model to obtain the probability density. The probability density is added to the conditional neural process through the neural process network, the context encoder and the target decoder are combined to model the data distribution, and the dynamic point process model is combined to predict the time type and time, and the state prediction sequence containing uncertainty is obtained, including:

[0030] The state evolution trajectory is input into a variational inference network constructed by setting a fully connected network with multiple hidden layers, and the variational distribution parameters are calculated according to minimizing the reconstruction error and the Kuhler-Leibler divergence, and the variational distribution parameters are input into an autoregressive flow model composed of multiple coupling layers, and the posterior distribution is mapped to a standard normal distribution to obtain a probability density function of the state evolution;

[0031] Inputting the probability density function of the state evolution into a conditional neural process, constructing a context encoder of the conditional neural process by setting a multi-layer long short-term memory network, encoding the state sequence into a context vector, and constructing a target decoder of the conditional neural process by setting a multi-layer long short-term memory network according to the splicing result of the context vector and a random vector sampled from a standard normal distribution, to generate a target state sequence;

[0032] The target state sequence is input into the dynamic point process model, and a fully connected network parameterized intensity function is set up with multiple hidden layers. The time and type of the next state transfer are predicted according to the current time and state. The parameters of the intensity function are optimized by maximizing the log-likelihood function of the time series. The intensity function is sampled using the Monte Carlo method to obtain the time and type of state transfer. The time and type of state transfer are combined with the target state sequence to generate a state prediction sequence containing uncertainty.

[0033] In an optional embodiment,

[0034] The state prediction sequence is added to the first discriminator of the hierarchical generative adversarial imitation learning network, the implicit distribution of the expert strategy is determined through adversarial training learning and the expert strategy is decomposed into subtasks at multiple abstract levels, the state transition diagram of the return task is constructed in combination with the option chain, the state evolution gradient is calculated by the dual proximal strategy optimization algorithm and added to the second discriminator based on neural architecture search, the optimal feature extraction structure is determined, the optimal feature extraction structure and the corresponding network parameters are updated by multiple parallel evaluation units in the distributed trust region evolution strategy algorithm to obtain the optimal parameter configuration, the optimal parameter configuration is mapped to a multi-dimensional feature vector corresponding to the return state, the state prediction is performed through the multi-dimensional evaluation network, the uncertainty of the state prediction is determined by the sequential Monte Carlo method, and the final judgment result includes:

[0035] The image sequence, position information, and posture data of the drone's historical return process are collected as expert strategy data, and the state prediction sequence containing uncertainty is input into the first discriminator of the hierarchical generative adversarial imitation learning network. The input state sequence is encoded by the generator of the hierarchical generative adversarial imitation learning network and a state sequence that imitates the expert strategy is generated. The adversarial loss between the state sequence and the real state sequence of the expert strategy is calculated by the first discriminator, and the high-level features of the state sequence are extracted by the second discriminator. The parameters of the generator and the discriminator are alternately updated by back propagation and gradient descent to obtain the implicit distribution of the expert strategy.

[0036] Based on the option chain method, the expert strategy is decomposed into subtasks at multiple abstract levels, and a state transition diagram is constructed. By setting nodes representing subtasks in the state transition diagram, edges between nodes represent migration probabilities and migration costs between subtasks, and using a dynamic programming algorithm to perform path search and planning on the state transition diagram, the optimal option sequence for completing the original task is obtained;

[0037] Input the implicit distribution of the expert strategy into the dual proximal strategy optimization algorithm, set the objective function of the drone return state prediction to minimize the deviation between the predicted state sequence and the true state sequence as the original problem, transform the original problem into a dual problem by introducing a dual variable, alternately optimize the prediction deviation minimization problem and the corresponding dual problem to calculate the state evolution gradient, input the state evolution gradient into the second discriminator based on neural architecture search, construct a candidate network architecture by setting the convolution kernel size search space and the fully connected layer neuron number search space, evaluate different network architectures using the state evolution gradient as the optimization target, and select the network architecture with the smallest state evolution gradient as the optimal feature extraction structure;

[0038] The optimal feature extraction structure is input into the distributed trust domain evolution strategy algorithm, and a parallel evaluation unit is constructed by dividing the evolutionary population into multiple sub-populations, and fitness evaluation and parameter update operations are performed on each sub-population in parallel to obtain a local optimal solution. The individual migration strategy is determined by calculating the trust between sub-populations, and the individuals with the highest fitness in each sub-population are migrated to other sub-populations to update the global optimal solution, thereby obtaining the optimal network parameter configuration;

[0039] The optimal network parameter configuration is mapped to a multidimensional feature vector corresponding to the return state through a fully connected network, a multidimensional evaluation network is constructed by setting up multiple parallel evaluation subnetworks, the evaluation subnetwork is used to extract and evaluate state features from different dimensions, the state features are repeatedly sampled multiple times using the sequential Monte Carlo method, the sampling results are input into the multidimensional evaluation network to obtain multiple prediction results, the statistical distribution characteristics of the prediction results are calculated to obtain the expected value and uncertainty measure of the state prediction, and the final return state determination result is generated.

[0040] In an optional embodiment,

[0041] Fitness evaluation and parameter update operations are performed in parallel on each sub-population to obtain the local optimal solution. The individual migration strategy is determined by calculating the trust between sub-populations, including:

[0042] Construct a multi-objective fitness evaluation function to calculate the fitness score of individuals in each sub-population. The fitness evaluation function includes a prediction error term, a parameter complexity penalty term, and a network structure penalty term. The parameter distribution similarity and historical iteration effect between sub-populations are calculated to establish a trust matrix between sub-populations.

[0043] Arrange the trust matrix in descending order according to the trust value, take the two sub-populations with the largest trust value as the first group, and take the two sub-populations with the largest trust value in the remaining sub-populations as the second group, and repeat the grouping until all sub-populations are grouped. In each group, set the sub-population with a large trust value as the source population, and the sub-population with a small trust value as the target population;

[0044] Calculate the trust difference between the source population and the target population, use the product of the trust difference and the preset benchmark migration number as the actual number of migrated individuals, form an individual migration strategy based on the source population, the target population and the number of migrated individuals, and select the individual with the highest fitness score from the source population as the migration individual based on the individual migration strategy, and replace the individual with the lowest fitness score in the target population with the migration individual;

[0045] Perform local parameter adjustment on the migrated individuals, where the adjustment amplitude is equal to the difference in trust between the source population and the target population multiplied by the preset adjustment coefficient. Recalculate the trust between sub-populations based on the adjusted individuals, update the trust matrix, and execute the individual migration strategy repeatedly until the preset termination condition is reached.

[0046] A second aspect of an embodiment of the present invention provides a drone return detection system based on AI recognition, comprising:

[0047] The first unit is used to obtain multi-source image data in the process of returning the drone and add it to the improved window-shifting hierarchical transformer model, extract multi-scale feature maps through multiple attention windows and add them to the cross-modal contrastive learning module, obtain semantic consistency features by calculating the mutual information in different views and add them to the dynamic routing capsule network, update the routing weights by calculating the coupling coefficients between feature vectors and output a high-dimensional feature vector containing target posture information, add the high-dimensional feature vector to a feature enhancement module including a channel recalibration unit and a spatial recalibration unit, update the adaptive weights by calculating global statistical information to obtain an enhanced feature vector, align the enhanced feature vectors according to the adversarial feature alignment method to minimize the maximum mean difference between different data distributions, and obtain an aligned feature representation;

[0048] The second unit is used to add the alignment feature representation to a pre-built graph attention diffusion model and represent it as a spatiotemporal graph structure, denoise the spatiotemporal graph structure through a diffusion process to obtain an accurate state representation, input the state representation into an adaptive causal convolutional network, extract multi-scale time series features through a hole convolution layer with a variable receptive field, solve it in combination with the continuous-time dynamic system in the neural ordinary differential equation network to obtain a state evolution trajectory, determine the posterior distribution of the state evolution trajectory through a variational inference network and map the posterior distribution to a standard normal distribution in combination with an autoregressive flow model to obtain a probability density, add the probability density to a conditional neural process through a neural process network, model data distribution in combination with a context encoder and a target decoder, predict time type and time in combination with a dynamic point process model, and obtain a state prediction sequence containing uncertainty;

[0049] The third unit is used to add the state prediction sequence to the first discriminator of the hierarchical generative adversarial imitation learning network, determine the implicit distribution of the expert strategy through adversarial training learning and decompose the expert strategy into sub-tasks at multiple abstract levels, construct a state transition diagram of the return task in combination with the option chain, calculate the state evolution gradient through the dual proximal strategy optimization algorithm and add it to the second discriminator based on neural architecture search, determine the optimal feature extraction structure, update the optimal feature extraction structure and the corresponding network parameters through multiple parallel evaluation units in the distributed trust domain evolution strategy algorithm to obtain the optimal parameter configuration, map the optimal parameter configuration to a multi-dimensional feature vector corresponding to the return state, perform state prediction through a multi-dimensional evaluation network, determine the uncertainty of the state prediction through a sequential Monte Carlo method, and obtain a final judgment result.

[0050] According to a third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] A fourth aspect of the embodiments of the present invention is:

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0057] In the present invention, multi-source image data and an improved window-shifting hierarchical transformer model are used to extract multi-scale features, and cross-modal contrastive learning and dynamic routing capsule networks are combined to effectively capture target posture information and semantic consistency features, thereby improving detection accuracy. The feature enhancement module and adversarial feature alignment method further enhance the robustness of feature representation, making it more adaptable to different data distributions. The graph attention diffusion model is used to denoise the spatiotemporal graph structure to obtain an accurate state representation. By combining an adaptive causal convolutional network, a neural ordinary differential equation network, and a variational inference network, accurate prediction of the state evolution trajectory is achieved, and uncertainty factors are considered through a neural process network and a dynamic point process model to improve the reliability of the prediction. The expert strategy is learned using a hierarchical generative adversarial imitation learning network, and the state transition diagram of the return task is constructed in combination with the option chain. The optimal feature extraction structure and parameter configuration are determined through optimization algorithms such as the dual proximal strategy optimization algorithm and the distributed trust domain evolution strategy algorithm, and the uncertainty of state prediction is evaluated through the sequential Monte Carlo method, ultimately obtaining a better return strategy and judgment result. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the process of a drone return detection method based on AI recognition according to an embodiment of the present invention;

[0059] Figure 2 Schematic diagram of the structure of a drone return detection system based on AI recognition according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 FIG. 1 is a flow chart of a drone return detection method based on AI recognition according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] S1. Acquire multi-source image data during the return process of the drone and add it to the improved window-shifting hierarchical transformer model, extract multi-scale feature maps through multiple attention windows and add them to the cross-modal contrastive learning module, obtain semantic consistency features by calculating the mutual information in different views and add them to the dynamic routing capsule network, update the routing weights by calculating the coupling coefficients between feature vectors and output a high-dimensional feature vector containing target posture information, add the high-dimensional feature vector to a feature enhancement module including a channel recalibration unit and a spatial recalibration unit, update the adaptive weights by calculating global statistical information, obtain an enhanced feature vector, align the enhanced feature vectors according to the adversarial feature alignment method to minimize the maximum mean difference between different data distributions, and obtain an aligned feature representation;

[0064] The improved window-shifting hierarchical transformer model is a deep learning model that combines a window mechanism and a hierarchical structure. It can dynamically adjust the attention range and optimize the efficiency and accuracy of information transmission when processing sequence data. The dynamic routing capsule network is a neural network model that uses a capsule network structure. Information is transmitted between different capsules through a dynamic routing mechanism to improve the model's expression ability and generalization performance for complex features. The channel recalibration unit is a module used in a convolutional neural network. By re-adjusting the weights or activation values ​​of each channel, the representation of important features is strengthened and redundant features are suppressed, thereby improving the representation ability and performance of the model. The spatial recalibration unit is similar to the channel recalibration unit, but its function is to re-adjust the feature map in the spatial dimension. By changing the weight of the spatial position, the spatial representation of the feature map is optimized and the spatial sensitivity of the model is enhanced. The adversarial feature alignment method is a method that uses an adversarial training mechanism to make the feature distribution of the source domain and the target domain consistent through an adversarial network, thereby improving the migration performance and generalization ability of cross-domain tasks.

[0065] In an optional embodiment,

[0066] The multi-source image data in the drone return process is obtained and added to the improved window-shifting hierarchical transformer model. The multi-scale feature map is extracted through multiple attention windows and added to the cross-modal contrastive learning module. The semantic consistency feature is obtained by calculating the mutual information in different views and added to the dynamic routing capsule network. The routing weight is updated by calculating the coupling coefficient between the feature vectors and a high-dimensional feature vector containing the target posture information is output. The high-dimensional feature vector is added to the feature enhancement module including the channel recalibration unit and the spatial recalibration unit. The adaptive weight is updated by calculating the global statistical information to obtain the enhanced feature vector. The enhanced feature vector is aligned according to the adversarial feature alignment method to minimize the maximum mean difference between different data distributions. The alignment feature representation includes:

[0067] Acquire infrared images, visible light images, and depth images during the drone return process as multi-source image data;

[0068] The multi-source image data is input into a shifting window hierarchical transformer model, wherein the shifting window hierarchical transformer model comprises three layers of attention windows, wherein the first layer of attention window extracts target contour features, the second layer of attention window extracts target region features, and the third layer of attention window extracts environmental background features, and the features extracted by the three layers of attention windows are fused to obtain a multi-scale feature map;

[0069] The multi-scale feature map is input into a cross-modal contrastive learning module, which calculates the cosine similarity between feature vectors of feature maps of different modalities at the same position, constructs a mutual information maximization objective function based on the cosine similarity, and optimizes the mutual information maximization objective function to obtain semantic consistency features;

[0070] Inputting the semantic consistency feature into a dynamic routing capsule network, wherein the dynamic routing capsule network comprises a plurality of capsule layers, each of the capsule layers comprises a plurality of capsules representing different posture attributes, adjacent capsule layers are connected via a transformation matrix, the parameters of the transformation matrix are obtained by calculating the coupling coefficient between capsules and iteratively updating the routing weights, and the outputs of the lower-layer capsules are weightedly fused based on the routing weights to obtain a high-dimensional feature vector;

[0071] The high-dimensional feature vector is input into a self-calibration feature enhancement module, wherein the self-calibration feature enhancement module comprises a channel recalibration unit and a spatial recalibration unit, wherein the channel recalibration unit generates a channel weight vector through global average pooling and a fully connected layer, and the spatial recalibration unit generates a spatial weight matrix through global maximum pooling, and the channel weight vector and the spatial weight matrix are respectively weightedly fused with the high-dimensional feature vector to obtain an enhanced feature vector;

[0072] The enhanced feature vector is input into an adversarial feature alignment module, which includes a feature extractor and a domain discriminator. According to the adversarial feature alignment method, the enhanced feature vector is mapped to a common feature space through the feature extractor, the data domain to which the feature belongs is discriminated through the domain discriminator, and the maximum mean difference between the source domain feature and the target domain feature in the reproducing kernel Hilbert space is calculated based on the Gaussian kernel function, and the alignment feature representation is obtained by minimizing the maximum mean difference.

[0073] The domain discriminator is a model used to distinguish whether data comes from a source domain or a target domain in adversarial training. Its task is to guide the feature alignment process so that the features of the source domain and the target domain cannot be distinguished, thereby achieving adversarial learning between domains. The regenerating kernel Hilbert space is a mathematical concept commonly used in kernel methods. By mapping data to a higher-dimensional Hilbert space, the data is easier to separate in the new space, thereby optimizing the performance of the learning algorithm.

[0074] The sensor system carried by the drone is used to synchronously collect infrared images, visible light images, and depth images. For example, long-wave infrared thermal imagers such as FLIR Boson are used to collect infrared images to obtain temperature information; conventional cameras are used to collect visible light images to obtain rich texture and color information; depth cameras such as Intel Real Sense are used to collect depth images to obtain distance information of the target. These multi-source image data provide complementary information of the target and the environment, laying the foundation for subsequent processing.

[0075] The acquired multi-source image data is input into the improved window-shifting hierarchical transformer model for feature extraction. The improved window-shifting hierarchical transformer model contains three layers of attention windows of different sizes to simulate the multi-scale focusing mechanism of human eyes when observing objects. The first layer of attention windows is small, mainly focusing on the local details of the target, such as the edge and marking line of the drone landing pad, and extracting the target contour features. Assuming that the image size is 256x256 pixels, the first layer window size is set to 32x32 pixels. The second layer of attention window size is moderate, such as 64x64 pixels, focusing on the overall morphology of the target area, such as the shape and size of the landing pad, and extracting the target area features. The third layer of attention window is larger, such as 128x128 pixels, focusing on the environmental information around the target, such as surrounding buildings and trees, and extracting environmental background features. The feature maps extracted by these three layers of attention windows are fused to obtain a feature map containing multi-scale information. For example, after splicing the three feature maps, 1x1 convolution is used for feature fusion to obtain the final multi-scale feature map.

[0076] The multi-scale feature map is input into the cross-modal contrastive learning module to learn the semantic consistency between different modal data. The core idea is to maximize the similarity between the feature vectors of different modal feature maps at the same spatial position. For example, the cosine similarity of the feature vectors of the infrared image feature map and the visible light image feature map at the same position is calculated. By constructing and optimizing the mutual information maximization objective function based on cosine similarity, the feature representations of different modalities are made more consistent at the semantic level, and semantic consistency features are obtained.

[0077] The semantic consistency features are input into the dynamic routing capsule network to extract a high-dimensional feature vector containing the target posture information. The dynamic routing capsule network consists of multiple capsule layers, each of which contains multiple capsules. Each capsule represents a certain posture attribute of the target, such as position, angle, size, etc. Adjacent capsule layers are connected by transformation matrices. By calculating the coupling coefficient of the feature vectors between adjacent capsules and iteratively updating the routing weights, a high-dimensional feature vector is finally obtained. This vector incorporates the posture information of the target. For example, it can represent the three-dimensional posture of the target apron relative to the drone.

[0078] The high-dimensional feature vector is then input into the self-calibration feature enhancement module to further improve the expressiveness of the feature. The self-calibration feature enhancement module includes a channel recalibration unit and a spatial recalibration unit. The channel recalibration unit extracts global statistical information through global average pooling, and generates a channel weight vector through a fully connected layer to weight the channel dimension of the feature map. The spatial recalibration unit generates a spatial weight matrix through global maximum pooling to weight the spatial dimension of the feature map. The channel weight vector and the spatial weight matrix are multiplied by the high-dimensional feature vector respectively to obtain an enhanced feature vector. This process can adaptively adjust the weight of the feature, highlight important features, and suppress noise interference.

[0079] The enhanced feature vector is input into the adversarial feature alignment module to eliminate the data distribution differences between different data sources. This module contains a feature extractor and a domain discriminator. The feature extractor maps the features of different data sources to a common feature space. The domain discriminator attempts to distinguish which data source the feature comes from. By minimizing the maximum mean difference between the source domain features and the target domain features in the reproducing kernel Hilbert space, the feature distributions of different data sources are made more consistent, and an aligned feature representation is obtained. For example, infrared image data can be used as the source domain and visible light image data as the target domain. Through adversarial training, the feature representations of the two data are made closer in the common space.

[0080] In this embodiment, through multi-source data fusion and multi-scale feature extraction, the target and environmental information can be perceived more accurately, thereby improving the positioning accuracy and return success rate of the UAV in complex environments. Infrared images are not affected by lighting conditions, and depth images can provide distance information, thereby improving the return performance of the UAV at night. Through adaptive feature enhancement and adversarial feature alignment, the need for data annotation and manual intervention can be reduced, thereby simplifying the deployment and maintenance of the system.

[0081] In an optional embodiment,

[0082] The mutual information maximization objective function based on the cosine similarity is constructed as shown in the following formula:

[0083] ;

[0084] Among them, J represents the objective function value, α(i, j) represents the attention weight of position (i, j), sim(x i ,y j ) represents the eigenvector x of the two modes at position (i, j) i and j The cosine similarity between them, λ represents the balance factor, d(x i ,y j ) represents the positive sample pair (x i ,y j ), d(x i ,y k ) represents the negative sample pair (x i ,y k ), M represents the boundary value in the contrast loss, which is used to control the distance difference between positive and negative samples.

[0085] In this embodiment, the objective function can effectively learn the correlation between multimodal data, improve the performance of the multimodal model, further enhance the robustness and generalization ability of the model through the idea of ​​comparative learning, and can be applied to various multimodal tasks.

[0086] S2. Add the aligned feature representation to the pre-built graph attention diffusion model and represent it as a spatiotemporal graph structure, denoise the spatiotemporal graph structure through a diffusion process to obtain an accurate state representation, input the state representation into an adaptive causal convolutional network, extract multi-scale time series features through a hole convolution layer with a variable receptive field, solve it in combination with the continuous-time dynamic system in the neural ordinary differential equation network to obtain a state evolution trajectory, determine the posterior distribution of the state evolution trajectory through a variational inference network and map the posterior distribution to a standard normal distribution in combination with an autoregressive flow model to obtain a probability density, add the probability density to a conditional neural process through a neural process network, model the data distribution in combination with a context encoder and a target decoder, predict the time type and time in combination with a dynamic point process model, and obtain a state prediction sequence containing uncertainty;

[0087] The graph attention diffusion model is a model that combines graph structure and attention mechanism. It can effectively propagate information on graph data, use attention mechanism to strengthen the relationship between nodes, and thus improve the learning and reasoning ability of graph data. The adaptive causal convolutional network is a special convolutional neural network that adopts causal convolution structure to ensure that the output is only affected by current and past information, and dynamically adjusts the parameters of the convolution kernel through adaptive mechanism to better capture the causal relationship in time series data. The neural ordinary differential equation network is a model that approximates the solution of ordinary differential equations through neural network, represents the solution of ordinary differential equations as the output of neural network, and combines deep learning optimization method for training to achieve modeling of complex dynamic systems. The hole convolution layer with variable receptive field is a convolution layer with extended receptive field. Introducing holes in the convolution operation increases the range of the receptive field while keeping the number of parameters unchanged, so that the model can capture a wider range of contextual information. The autoregressive flow model is a generative model that models the probability distribution of data in an autoregressive manner, gradually generates samples, and optimizes each step in the generation process through back propagation. It can effectively process complex sequence data. The neural process network is a Bayesian inference model based on a neural network. It models the relationship between input and output as a process, and can perform fast inference and generation. It is particularly suitable for processing data with uncertainty and high noise. The dynamic point process model is a model based on point process theory, which can describe a sequence of events that changes over time. It uses dynamic adjustment of the relationship and occurrence probability between events, and is widely used in time series analysis in biology, finance and other fields.

[0088] In an optional embodiment,

[0089] The aligned feature representation is added to the pre-built graph attention diffusion model and represented as a spatiotemporal graph structure. The spatiotemporal graph structure is denoised through a diffusion process to obtain an accurate state representation. The state representation is input into an adaptive causal convolutional network, and multi-scale temporal features are extracted through a dilated convolutional layer with a variable receptive field. The continuous-time dynamic system in the neural ordinary differential equation network is combined for solution, and the state evolution trajectory is obtained, including:

[0090] The aligned feature representation is input into a pre-built graph attention diffusion model, a spatiotemporal graph structure is constructed by setting state feature nodes of multiple time steps and edges of temporal dependencies between states, attention weights between nodes are calculated according to the graph attention layer, and node features are aggregated to obtain context node representations containing long-term and short-term dependencies;

[0091] A diffusion process is performed on the spatiotemporal graph structure through multiple diffusion steps, and in each diffusion step, Gaussian noise is added to the node features and the features are updated through a multi-layer graph neural network. The feature update process of the graph neural network is calculated based on the node features of the previous diffusion step, the Gaussian noise intensity, and the time interval of adjacent time steps. The inverse diffusion process is performed according to the iteratively updated node features to remove the noise, so as to obtain an accurate state representation;

[0092] The precise state representation is input into an adaptive causal convolutional network, the adaptive causal convolutional network is constructed by setting a plurality of dilated convolutional layers with different dilated rates, time series features of multiple scales are extracted according to the dilated convolutional layers with different dilated rates, and the time series features of the multiple scales are feature spliced ​​to obtain multi-scale time series features;

[0093] The multi-scale time series features are input into the neural ordinary differential equation network, and the right function of the ordinary differential equation is parameterized by setting a fully connected network with multiple hidden layers. The high-order Runge-Kutta method is used to perform numerical integration at multiple time points to solve the ordinary differential equation and obtain the state evolution trajectory.

[0094] The Gaussian noise refers to random noise with a Gaussian distribution, which is characterized by a mean of 0 and a variance of a certain constant. It is widely used to simulate noise interference in the real world, especially in signal processing and statistical analysis. The right function of the ordinary differential equation is a part of the ordinary differential equation, usually representing the function on the right side of the equation, which determines the dynamic changes of the system and is often used to describe dynamic processes in physics, engineering and other fields. The high-order Runge-Kutta method is a numerical method for solving ordinary differential equations. It achieves higher accuracy through multiple estimates and weighted averaging, and is suitable for solving ordinary differential equations in complex or rigid problems.

[0095] Extract the state features of the system and convert them into aligned feature representations. For example, in a traffic flow prediction system, state features may include the number of vehicles on the road, vehicle speed, weather conditions, etc. These features need to be preprocessed and normalized to ensure that they are in the same scale range and aligned with the timestamps. Assuming that at a point in time, the number of vehicles is 100, the vehicle speed is 60 km / h, and the weather is sunny, these features can be represented as [100, 60, 1], where 1 represents sunny weather.

[0096] Construct a predefined graph attention diffusion model. The model contains state feature nodes of multiple time steps and edges of temporal dependencies between states, forming a spatiotemporal graph structure. For example, the state features of five consecutive time steps can be used as nodes, and the state transition relationships between adjacent time steps can be used as edges. The weights of the edges can be initialized according to the time interval or other prior knowledge.

[0097] The constructed spatiotemporal graph structure is subjected to a diffusion process, which includes multiple diffusion steps. In each diffusion step, Gaussian noise is added to the node features. The intensity of the noise gradually increases as the diffusion step proceeds. For example, in the first diffusion step, Gaussian noise with a standard deviation of 0.1 is added to the node feature [100, 60, 1] to obtain [100.05, 59.98, 1.02], and the features are updated through a multi-layer graph neural network. The feature update process of the graph neural network is calculated based on the node features of the previous diffusion step, the Gaussian noise intensity, and the time interval between adjacent time steps. For example, considering the node features of the previous and next time steps of the current node, and the time interval of 1 hour, the features of the current node can be updated, and the inverse diffusion process is performed according to the iteratively updated node features to remove the noise and obtain an accurate state representation.

[0098] The obtained precise state representation is input into the adaptive causal convolutional network, which is composed of multiple dilated convolutional layers with different dilation rates. Dilated convolutional layers with different dilation rates can extract temporal features of multiple scales. For example, a convolutional layer with a dilation rate of 1 can extract local features, and a convolutional layer with a dilation rate of 2 can extract features of a larger range. These features of different scales are spliced ​​to obtain multi-scale temporal features.

[0099] The multi-scale time series features are input into the neural ordinary differential equation network. By setting a fully connected network with multiple hidden layers to parameterize the right function of the ordinary differential equation, a high-order Runge-Kutta method is used to numerically integrate and solve the ordinary differential equation at multiple time points to obtain the state evolution trajectory. For example, the fourth-order Runge-Kutta method can be used to solve the ordinary differential equation with a time step of 0.1 to obtain the state prediction values ​​at multiple future time points.

[0100] In this embodiment, by combining the graph attention mechanism, diffusion process, adaptive causal convolution and neural ordinary differential equations, the spatiotemporal dependency and dynamic evolution law of the system state can be effectively captured, thereby improving the accuracy of state prediction. The diffusion process can effectively reduce the impact of noise on the model and improve the robustness of the model. Adaptive causal convolution can adaptively adjust the receptive field according to the characteristics of the data, further enhancing the robustness of the model. The graph attention mechanism can provide information on the attention weights between nodes, revealing the dependencies between system states, thereby improving the interpretability of the model. The neural ordinary differential equation network can describe the continuous time evolution process of the state, provide more fine-grained state evolution information, and enhance the interpretability of the model.

[0101] In an optional embodiment,

[0102] The posterior distribution of the state evolution trajectory is determined by the variational inference network and the posterior distribution is mapped to the standard normal distribution in combination with the autoregressive flow model to obtain the probability density. The probability density is added to the conditional neural process through the neural process network, the context encoder and the target decoder are combined to model the data distribution, and the dynamic point process model is combined to predict the time type and time, and the state prediction sequence containing uncertainty is obtained, including:

[0103] The state evolution trajectory is input into a variational inference network constructed by setting a fully connected network with multiple hidden layers, and the variational distribution parameters are calculated according to minimizing the reconstruction error and the Kuhler-Leibler divergence, and the variational distribution parameters are input into an autoregressive flow model composed of multiple coupling layers, and the posterior distribution is mapped to a standard normal distribution to obtain a probability density function of the state evolution;

[0104] Inputting the probability density function of the state evolution into a conditional neural process, constructing a context encoder of the conditional neural process by setting a multi-layer long short-term memory network, encoding the state sequence into a context vector, and constructing a target decoder of the conditional neural process by setting a multi-layer long short-term memory network according to the splicing result of the context vector and a random vector sampled from a standard normal distribution, to generate a target state sequence;

[0105] The target state sequence is input into the dynamic point process model, and a fully connected network parameterized intensity function is set up with multiple hidden layers. The time and type of the next state transfer are predicted according to the current time and state. The parameters of the intensity function are optimized by maximizing the log-likelihood function of the time series. The intensity function is sampled using the Monte Carlo method to obtain the time and type of state transfer. The time and type of state transfer are combined with the target state sequence to generate a state prediction sequence containing uncertainty.

[0106] The Monte Carlo method is a numerical calculation method based on random sampling. It estimates the solution of a complex problem by simulating a large number of random events. It is widely used in many fields such as statistical physics, financial modeling, and optimization problems.

[0107] The variational inference network is used to learn the posterior distribution of the state evolution trajectory. The historical state evolution trajectory is used as input, and the true posterior distribution is approximated through a fully connected network structure with multiple hidden layers. By minimizing the reconstruction error and the measure of the difference between the two distributions, the parameters of the variational distribution are continuously adjusted to make it close to the true posterior distribution. For example, assuming that the state evolution trajectory consists of a series of position coordinates, the variational inference network will learn the posterior distribution of these coordinates.

[0108] The learned posterior distribution is mapped to a standard normal distribution using an autoregressive flow model. The autoregressive flow model consists of multiple coupling layers, each of which performs a reversible transformation on the input variable. In this way, complex non-Gaussian distributions can be converted into simple standard normal distributions, which is convenient for subsequent processing. For example, the previously learned posterior distribution of the position coordinates is converted into a standard normal distribution through an autoregressive flow model to obtain the corresponding probability density function value.

[0109] The obtained probability density function value is input into the conditional neural process. The conditional neural process consists of a context encoder and a target decoder. The context encoder uses a multi-layer long short-term memory network to encode the historical state sequence and extract context information. The target decoder also uses a multi-layer long short-term memory network to combine the context information and a random vector sampled from a standard normal distribution to generate a target state sequence. For example, the probability density function value of the position coordinate is used as the input of the conditional neural process, combined with the historical position sequence, to generate multiple possible future position sequences.

[0110] The generated target state sequence is input into the dynamic point process model. The dynamic point process model models the time and type of state transitions through a parameterized intensity function. The intensity function consists of a fully connected network with multiple hidden layers, and its parameters are optimized by maximizing the log-likelihood function of the time series. Using the Monte Carlo method to sample the intensity function, multiple possible future state transition times and types can be obtained. For example, assuming that the state transition types are "acceleration", "deceleration" and "constant speed", the dynamic point process model can predict the probabilities of these state transitions occurring at different moments in the future, and combine them with the previously generated position sequence to obtain the final state prediction sequence containing uncertainty. For example, at a certain point in time, it is predicted that the object may accelerate with an 80% probability and maintain a constant speed with a 20% probability, and the corresponding speed and position are given.

[0111] In this embodiment, by combining the variational inference network and the autoregressive flow model, the posterior distribution of the state evolution trajectory can be captured more accurately, thereby improving the prediction accuracy. Through the dynamic point process model, the time and type of state transition can be probabilistically modeled to quantify the uncertainty of the prediction results. It combines a variety of advanced deep learning models and can effectively process complex state evolution trajectories, such as nonlinear, non-Gaussian, and time-irregular data.

[0112] S3. Add the state prediction sequence to the first discriminator of the hierarchical generative adversarial imitation learning network, determine the implicit distribution of the expert strategy through adversarial training and decompose the expert strategy into sub-tasks at multiple abstract levels, construct a state transition diagram of the return task in combination with the option chain, calculate the state evolution gradient through the dual proximal strategy optimization algorithm and add it to the second discriminator based on neural architecture search, determine the optimal feature extraction structure, update the optimal feature extraction structure and the corresponding network parameters through multiple parallel evaluation units in the distributed trust domain evolution strategy algorithm, obtain the optimal parameter configuration, map the optimal parameter configuration into a multi-dimensional feature vector corresponding to the return state, perform state prediction through a multi-dimensional evaluation network, determine the uncertainty of the state prediction through the sequential Monte Carlo method, and obtain the final judgment result.

[0113] The hierarchical generative adversarial imitation learning network is a framework that combines generative adversarial networks and imitation learning. It guides the learning of intelligent agents by simulating expert behaviors at different levels, and uses the generative adversarial mechanism to improve the generation quality and generalization ability of strategies. The implicit distribution refers to the potential probability distribution that cannot be directly observed in machine learning and statistical modeling. It is usually estimated or inferred through the hidden structure of the model to describe hidden variables or processes and help improve the expressive power of the model. The option chain is a method in reinforcement learning for dynamically selecting multiple action options according to the state of the environment and the goal of the task during the decision-making process, thereby optimizing the strategy by selecting different options in multi-tasks or complex tasks. The dual proximal policy optimization algorithm is an optimization algorithm in reinforcement learning. By introducing the dual problem and The proximal method balances learning efficiency and stability, and is usually used to optimize the estimation of policy gradients and improve algorithm convergence. The state evolution gradient is a gradient calculation method in reinforcement learning, which is used to describe the changes in system states during the time step evolution process. It relies on the gradient information of the current state and helps the model optimize the strategy during training. The distributed trust region evolution strategy algorithm is an evolution strategy optimization algorithm based on a distributed computing framework. It limits the size of the search space by introducing a trust region mechanism, thereby improving the stability and convergence speed of the algorithm in a large-scale environment. The sequential Monte Carlo method is a Monte Carlo method suitable for processing sequence data. It evaluates the behavior of the sequence through multiple simulations and random sampling, and then estimates the expected value of the entire process. It is widely used in dynamic programming and sequential decision-making problems.

[0114] In an optional embodiment,

[0115] The state prediction sequence is added to the first discriminator of the hierarchical generative adversarial imitation learning network, the implicit distribution of the expert strategy is determined through adversarial training learning and the expert strategy is decomposed into subtasks at multiple abstract levels, the state transition diagram of the return task is constructed in combination with the option chain, the state evolution gradient is calculated by the dual proximal strategy optimization algorithm and added to the second discriminator based on neural architecture search, the optimal feature extraction structure is determined, the optimal feature extraction structure and the corresponding network parameters are updated by multiple parallel evaluation units in the distributed trust region evolution strategy algorithm to obtain the optimal parameter configuration, the optimal parameter configuration is mapped to a multi-dimensional feature vector corresponding to the return state, the state prediction is performed through the multi-dimensional evaluation network, the uncertainty of the state prediction is determined by the sequential Monte Carlo method, and the final judgment result includes:

[0116] The image sequence, position information, and posture data of the drone's historical return process are collected as expert strategy data, and the state prediction sequence containing uncertainty is input into the first discriminator of the hierarchical generative adversarial imitation learning network. The input state sequence is encoded by the generator of the hierarchical generative adversarial imitation learning network and a state sequence that imitates the expert strategy is generated. The adversarial loss between the state sequence and the real state sequence of the expert strategy is calculated by the first discriminator, and the high-level features of the state sequence are extracted by the second discriminator. The parameters of the generator and the discriminator are alternately updated by back propagation and gradient descent to obtain the implicit distribution of the expert strategy.

[0117] Based on the option chain method, the expert strategy is decomposed into subtasks at multiple abstract levels, and a state transition diagram is constructed. By setting nodes representing subtasks in the state transition diagram, edges between nodes represent migration probabilities and migration costs between subtasks, and using a dynamic programming algorithm to perform path search and planning on the state transition diagram, the optimal option sequence for completing the original task is obtained;

[0118] Input the implicit distribution of the expert strategy into the dual proximal strategy optimization algorithm, set the objective function of the drone return state prediction to minimize the deviation between the predicted state sequence and the true state sequence as the original problem, transform the original problem into a dual problem by introducing a dual variable, alternately optimize the prediction deviation minimization problem and the corresponding dual problem to calculate the state evolution gradient, input the state evolution gradient into the second discriminator based on neural architecture search, construct a candidate network architecture by setting the convolution kernel size search space and the fully connected layer neuron number search space, evaluate different network architectures using the state evolution gradient as the optimization target, and select the network architecture with the smallest state evolution gradient as the optimal feature extraction structure;

[0119] The optimal feature extraction structure is input into the distributed trust domain evolution strategy algorithm, and a parallel evaluation unit is constructed by dividing the evolutionary population into multiple sub-populations, and fitness evaluation and parameter update operations are performed on each sub-population in parallel to obtain a local optimal solution. The individual migration strategy is determined by calculating the trust between sub-populations, and the individuals with the highest fitness in each sub-population are migrated to other sub-populations to update the global optimal solution, thereby obtaining the optimal network parameter configuration;

[0120] The optimal network parameter configuration is mapped to a multidimensional feature vector corresponding to the return state through a fully connected network, a multidimensional evaluation network is constructed by setting up multiple parallel evaluation subnetworks, the evaluation subnetwork is used to extract and evaluate state features from different dimensions, the state features are repeatedly sampled multiple times using the sequential Monte Carlo method, the sampling results are input into the multidimensional evaluation network to obtain multiple prediction results, the statistical distribution characteristics of the prediction results are calculated to obtain the expected value and uncertainty measure of the state prediction, and the final return state determination result is generated.

[0121] The migration probability is a parameter in a Markov process or a decision-making process, which indicates the probability of transitioning from one state to another. It is usually related to the dynamic evolution of the system and the decision-making rules, and is used to describe the transition law between system states. The migration cost refers to the cost or loss incurred in the process of transitioning from one state to another, and is usually used to optimize resource consumption in the decision-making process or reduce the burden on the system. In many optimization problems, the migration cost is one of the objectives that need to be minimized. The dual variable is an auxiliary variable introduced in the optimization problem, and is usually used to transform the original problem into its dual problem, helping to obtain the optimal solution to the original problem by solving the dual problem. It is often used in convex optimization and Lagrange multiplier method.

[0122] When the drone performs the return mission, its flight trajectory, attitude information and image sequence of the surrounding environment are recorded. The flight trajectory includes information such as longitude, latitude, altitude and speed, and the attitude information includes data such as pitch angle, roll angle and yaw angle. The image sequence is collected by the camera on the drone, recording the field of view of the drone during the return process. For example, the data of 100 successful drone returns are recorded, each containing 1000 frames of images, each frame of the image has a resolution of 1920x1080 pixels, and the corresponding longitude, latitude, altitude, speed, pitch angle, roll angle and yaw angle are recorded at the same time.

[0123] The collected image sequences, position information and attitude data are integrated into an expert strategy dataset. It contains the experience of successful return of drones in various situations. For example, the collected data of 100 successful return of drones are organized into a dataset, each of which contains 1,000 frames of images and their corresponding flight trajectory and attitude information.

[0124] The collected image data is preprocessed, such as image grayscale, normalization, and denoising, to improve the training efficiency and performance of the model. For example, the image is converted to a grayscale image and the pixel value is normalized to between 0 and 1 to reduce the amount of calculation and improve the robustness of the model.

[0125] Construct a hierarchical generative adversarial imitation learning network consisting of a generator and a discriminator. The generator is used to generate a state sequence that imitates the expert strategy, and the discriminator is used to distinguish the generated state sequence from the real state sequence of the expert strategy. For example, the generator uses a recurrent neural network structure, the input is a preprocessed image sequence, and the output is the predicted flight trajectory and posture information; the discriminator uses a convolutional neural network structure, the input is the real or generated flight trajectory and posture information, and the output is the true or false judgment result.

[0126] The preprocessed image sequence is input into the hierarchical generative adversarial imitation learning network for training. Through adversarial training, the generator can generate more realistic state sequences, and the discriminator can better distinguish between real and generated state sequences. For example, the network is trained using the Adam optimizer to update the parameters of the generator and discriminator by minimizing the adversarial loss and state evolution gradient.

[0127] Based on the option chain method, the expert strategy is decomposed into subtasks at multiple abstract levels to construct a state transition graph. The nodes in the state transition graph represent subtasks, and the edges represent the migration probability and migration cost between subtasks. For example, the drone return task is decomposed into subtasks such as takeoff, cruising, and landing, and the migration probability and migration cost between subtasks are calculated based on the expert strategy data to construct a state transition graph.

[0128] The dynamic programming algorithm is used to search and plan paths on the state transition graph to obtain the optimal sequence of options to complete the original task. For example, the Dijkstra algorithm is used to search for the shortest path on the state transition graph to obtain the optimal subtask sequence to complete the drone return task.

[0129] By setting the search space of convolution kernel size and the search space of the number of neurons in the fully connected layer, we construct candidate network architectures, use the state evolution gradient as the optimization target to evaluate different network architectures, and select the network architecture with the smallest state evolution gradient as the optimal feature extraction structure. For example, we construct 100 different network architectures, use the state evolution gradient as the evaluation indicator, and select the architecture with the smallest gradient.

[0130] The optimal feature extraction structure is input into the distributed trust region evolution strategy algorithm. By dividing the evolutionary population into multiple sub-populations to build a parallel evaluation unit, fitness evaluation and parameter update operations are performed in parallel on each sub-population. For example, the population is divided into 10 sub-populations and trained in parallel on 10 machines. The individual migration strategy is determined by calculating the trust between sub-populations, and the individuals with the highest fitness in each sub-population are migrated to other sub-populations to update the global optimal solution and obtain the optimal network parameter configuration.

[0131] The optimal network parameter configuration is mapped to a multidimensional feature vector corresponding to the return state through a fully connected network. A multidimensional evaluation network is constructed by setting up multiple parallel evaluation subnetworks, and the evaluation subnetworks are used to extract and evaluate state features from different dimensions. The sequential Monte Carlo method is used to perform multiple repeated sampling of state features, and the sampling results are input into the multidimensional evaluation network to obtain multiple prediction results. The statistical distribution characteristics of the prediction results are calculated to obtain the expected value and uncertainty measure of the state prediction. Combined with the expected value and uncertainty measure of the state prediction, the final return state determination result is generated.

[0132] In this embodiment, through imitation learning and state prediction, the UAV can be returned autonomously and safely in various complex environments. Through state transition diagrams and path planning, the return path of the UAV can be optimized and the return efficiency can be improved. Through distributed training and uncertainty assessment, the robustness of the model can be improved so that it can still maintain high performance in the face of noise and interference.

[0133] In an optional embodiment,

[0134] Fitness evaluation and parameter update operations are performed in parallel on each sub-population to obtain the local optimal solution. The individual migration strategy is determined by calculating the trust between sub-populations, including:

[0135] Construct a multi-objective fitness evaluation function to calculate the fitness score of individuals in each sub-population. The fitness evaluation function includes a prediction error term, a parameter complexity penalty term, and a network structure penalty term. The parameter distribution similarity and historical iteration effect between sub-populations are calculated to establish a trust matrix between sub-populations.

[0136] Arrange the trust matrix in descending order according to the trust value, take the two sub-populations with the largest trust value as the first group, and take the two sub-populations with the largest trust value in the remaining sub-populations as the second group, and repeat the grouping until all sub-populations are grouped. In each group, set the sub-population with a large trust value as the source population, and the sub-population with a small trust value as the target population;

[0137] Calculate the trust difference between the source population and the target population, use the product of the trust difference and the preset benchmark migration number as the actual number of migrated individuals, form an individual migration strategy based on the source population, the target population and the number of migrated individuals, and select the individual with the highest fitness score from the source population as the migration individual based on the individual migration strategy, and replace the individual with the lowest fitness score in the target population with the migration individual;

[0138] Perform local parameter adjustment on the migrated individuals, where the adjustment amplitude is equal to the difference in trust between the source population and the target population multiplied by the preset adjustment coefficient. Recalculate the trust between sub-populations based on the adjusted individuals, update the trust matrix, and execute the individual migration strategy repeatedly until the preset termination condition is reached.

[0139] The trust matrix is ​​a matrix structure used to represent the trust relationship between parties, in which each element represents the degree of trust between different entities. It is usually used in scenarios such as multi-party collaboration, decision support and information exchange to help optimize trust-related decision-making processes.

[0140] Initialize multiple subpopulations. Each subpopulation contains a certain number of individuals, each of which represents a potential solution to the problem and is represented by a set of parameters. For example, suppose the problem is to find the minimum value of a function, and each individual can be represented by a set of function parameters. Initialize 10 subpopulations, each of which contains 50 individuals, and the parameters of each individual are randomly generated.

[0141] Fitness evaluation and parameter update operations are performed in parallel for each subpopulation. The fitness evaluation function is used to measure the quality of each individual's solution to the problem, and is composed of a prediction error term, a parameter complexity penalty term, and a network structure penalty term. The prediction error term measures the gap between the individual's predicted value and the true value, the parameter complexity penalty term measures the complexity of the individual's parameters, and the network structure penalty term measures the complexity of the individual's network structure. For example, if an individual can fit the training data well, but the parameters are very complex, its fitness score will be affected by the parameter complexity penalty term and reduced. The parameter update operation uses evolutionary algorithms such as genetic algorithms and particle swarm algorithms to update the individual's parameters according to the individual's fitness score, so that it evolves towards a better solution.

[0142] Calculate the trust between subpopulations. The trust is calculated based on the parameter distribution similarity and historical iteration effect between subpopulations. Parameter distribution similarity measures the similarity of individual parameter distributions in two subpopulations, and historical iteration effect measures the performance improvement brought about by migrating individuals from two subpopulations to each other in past iterations. For example, if the parameter distributions of two subpopulations are very similar and migrating individuals from each other can significantly improve each other's performance, the trust between them will be high. Store the calculation results in a trust matrix. Assuming that the parameter distribution similarity of subpopulation 1 and subpopulation 2 is 0.8 and the historical iteration effect is 0.9, the trust between them can be set to 0.8×0.9=0.72.

[0143] Determine the individual migration strategy based on the trust matrix. Arrange the trust matrix in descending order by trust value, take the two subpopulations with the largest trust values ​​as the first group, take the two subpopulations with the largest trust values ​​among the remaining subpopulations as the second group, and so on, until all subpopulations are grouped. In each group, the subpopulation with a large trust value is set as the source population, and the subpopulation with a small trust value is set as the target population. For example, assuming that subpopulation 1 and subpopulation 2 have the highest trust, subpopulation 1 is used as the source population and subpopulation 2 is used as the target population.

[0144] Calculate the trust difference between the source population and the target population, and take the product of the trust difference and the preset benchmark migration number as the actual number of migrated individuals. For example, assuming that the trust difference between the source population and the target population is 0.2, and the preset benchmark migration number is 5, then the actual number of migrated individuals is 0.2×5=1.

[0145] Select the individual with the highest fitness score from the source population as the migration individual, replace the individual with the lowest fitness score in the target population with the migration individual, and perform local parameter adjustment on the migrated individual. The adjustment range is equal to the difference in trust between the source population and the target population multiplied by the preset adjustment coefficient. For example, assuming that the difference in trust between the source population and the target population is 0.2, and the preset adjustment coefficient is 0.1, the adjustment range is 0.2×0.1=0.02.

[0146] Based on the adjusted individuals, the trust between sub-populations is recalculated and the trust matrix is ​​updated. The individual migration strategy is executed cyclically until the preset termination condition is reached, such as reaching the maximum number of iterations or finding a solution that meets the requirements.

[0147] In this embodiment, through parallel evaluation and trust-guided migration strategy, the algorithm can find the local optimal solution more quickly. The multi-subpopulation mechanism and migration strategy help maintain population diversity and prevent the algorithm from falling into the local optimum. The multi-objective fitness evaluation function and local parameter adjustment can help the algorithm find a higher quality solution.

[0148] Figure 2FIG. 1 is a schematic diagram of the structure of a drone return detection system based on AI recognition according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0149] The first unit is used to obtain multi-source image data in the process of returning the drone and add it to the improved window-shifting hierarchical transformer model, extract multi-scale feature maps through multiple attention windows and add them to the cross-modal contrastive learning module, obtain semantic consistency features by calculating the mutual information in different views and add them to the dynamic routing capsule network, update the routing weights by calculating the coupling coefficients between feature vectors and output a high-dimensional feature vector containing target posture information, add the high-dimensional feature vector to a feature enhancement module including a channel recalibration unit and a spatial recalibration unit, update the adaptive weights by calculating global statistical information to obtain an enhanced feature vector, align the enhanced feature vectors according to the adversarial feature alignment method to minimize the maximum mean difference between different data distributions, and obtain an aligned feature representation;

[0150] The second unit is used to add the alignment feature representation to a pre-built graph attention diffusion model and represent it as a spatiotemporal graph structure, denoise the spatiotemporal graph structure through a diffusion process to obtain an accurate state representation, input the state representation into an adaptive causal convolutional network, extract multi-scale time series features through a hole convolution layer with a variable receptive field, solve it in combination with the continuous-time dynamic system in the neural ordinary differential equation network to obtain a state evolution trajectory, determine the posterior distribution of the state evolution trajectory through a variational inference network and map the posterior distribution to a standard normal distribution in combination with an autoregressive flow model to obtain a probability density, add the probability density to a conditional neural process through a neural process network, model data distribution in combination with a context encoder and a target decoder, predict time type and time in combination with a dynamic point process model, and obtain a state prediction sequence containing uncertainty;

[0151] The third unit is used to add the state prediction sequence to the first discriminator of the hierarchical generative adversarial imitation learning network, determine the implicit distribution of the expert strategy through adversarial training learning and decompose the expert strategy into sub-tasks at multiple abstract levels, construct a state transition diagram of the return task in combination with the option chain, calculate the state evolution gradient through the dual proximal strategy optimization algorithm and add it to the second discriminator based on neural architecture search, determine the optimal feature extraction structure, update the optimal feature extraction structure and the corresponding network parameters through multiple parallel evaluation units in the distributed trust domain evolution strategy algorithm to obtain the optimal parameter configuration, map the optimal parameter configuration to a multi-dimensional feature vector corresponding to the return state, perform state prediction through a multi-dimensional evaluation network, determine the uncertainty of the state prediction through a sequential Monte Carlo method, and obtain a final judgment result.

[0152] According to a third aspect of the embodiments of the present invention,

[0153] An electronic device is provided, comprising:

[0154] processor;

[0155] a memory for storing processor-executable instructions;

[0156] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0157] A fourth aspect of the embodiments of the present invention is:

[0158] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0159] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The drone return detection method based on AI recognition is characterized by: include: The multi-source image data in the drone return process is obtained and added to the improved window-shifting hierarchical transformer model. The multi-scale feature maps are extracted through multiple attention windows and added to the cross-modal contrastive learning module. The semantic consistency features are obtained by calculating the mutual information in different views and added to the dynamic routing capsule network. The routing weights are updated by calculating the coupling coefficients between the feature vectors and a high-dimensional feature vector containing the target posture information is output. The high-dimensional feature vector is added to the feature enhancement module including the channel recalibration unit and the spatial recalibration unit. The adaptive weights are updated by calculating the global statistical information to obtain the enhanced feature vector. The enhanced feature vector is aligned according to the adversarial feature alignment method to minimize the maximum mean difference between different data distributions to obtain the aligned feature representation. The aligned feature representation is added to a pre-built graph attention diffusion model and represented as a spatiotemporal graph structure, the spatiotemporal graph structure is denoised through a diffusion process to obtain an accurate state representation, the state representation is input into an adaptive causal convolutional network, multi-scale temporal features are extracted through a dilated convolutional layer with a variable receptive field, the state evolution trajectory is obtained by combining with the continuous-time dynamic system in the neural ordinary differential equation network, the posterior distribution of the state evolution trajectory is determined through a variational inference network and the posterior distribution is mapped to a standard normal distribution in combination with an autoregressive flow model to obtain a probability density, the probability density is added to a conditional neural process through a neural process network, the data distribution is modeled in combination with a context encoder and a target decoder, the time type and time are predicted in combination with a dynamic point process model, and a state prediction sequence containing uncertainty is obtained; The state prediction sequence is added to the first discriminator of the hierarchical generative adversarial imitation learning network, the implicit distribution of the expert strategy is determined through adversarial training and learning, and the expert strategy is decomposed into sub-tasks at multiple abstract levels. The state transition diagram of the return task is constructed in combination with the option chain, the state evolution gradient is calculated by the dual proximal strategy optimization algorithm and added to the second discriminator based on neural architecture search, the optimal feature extraction structure is determined, the optimal feature extraction structure and the corresponding network parameters are updated by multiple parallel evaluation units in the distributed trust domain evolution strategy algorithm to obtain the optimal parameter configuration, the optimal parameter configuration is mapped to a multidimensional feature vector corresponding to the return state, the state prediction is performed through the multidimensional evaluation network, the uncertainty of the state prediction is determined by the sequential Monte Carlo method, and the final judgment result is obtained.

2. The method according to claim 1, characterized in that The multi-source image data in the drone return process is obtained and added to the improved window-shifting hierarchical transformer model. The multi-scale feature map is extracted through multiple attention windows and added to the cross-modal contrastive learning module. The semantic consistency feature is obtained by calculating the mutual information in different views and added to the dynamic routing capsule network. The routing weight is updated by calculating the coupling coefficient between the feature vectors and a high-dimensional feature vector containing the target posture information is output. The high-dimensional feature vector is added to the feature enhancement module including the channel recalibration unit and the spatial recalibration unit. The adaptive weight is updated by calculating the global statistical information to obtain the enhanced feature vector. The enhanced feature vector is aligned according to the adversarial feature alignment method to minimize the maximum mean difference between different data distributions. The alignment feature representation includes: Acquire infrared images, visible light images, and depth images during the drone return process as multi-source image data; The multi-source image data is input into a shifting window hierarchical transformer model, wherein the shifting window hierarchical transformer model comprises three layers of attention windows, wherein the first layer of attention window extracts target contour features, the second layer of attention window extracts target region features, and the third layer of attention window extracts environmental background features, and the features extracted by the three layers of attention windows are fused to obtain a multi-scale feature map; The multi-scale feature map is input into a cross-modal contrastive learning module, which calculates the cosine similarity between feature vectors of feature maps of different modalities at the same position, constructs a mutual information maximization objective function based on the cosine similarity, and optimizes the mutual information maximization objective function to obtain semantic consistency features; Inputting the semantic consistency feature into a dynamic routing capsule network, wherein the dynamic routing capsule network comprises a plurality of capsule layers, each of the capsule layers comprises a plurality of capsules representing different posture attributes, adjacent capsule layers are connected via a transformation matrix, the parameters of the transformation matrix are obtained by calculating the coupling coefficient between capsules and iteratively updating the routing weights, and the outputs of the lower-layer capsules are weightedly fused based on the routing weights to obtain a high-dimensional feature vector; The high-dimensional feature vector is input into a self-calibration feature enhancement module, wherein the self-calibration feature enhancement module comprises a channel recalibration unit and a spatial recalibration unit, wherein the channel recalibration unit generates a channel weight vector through global average pooling and a fully connected layer, and the spatial recalibration unit generates a spatial weight matrix through global maximum pooling, and the channel weight vector and the spatial weight matrix are respectively weightedly fused with the high-dimensional feature vector to obtain an enhanced feature vector; The enhanced feature vector is input into an adversarial feature alignment module, which includes a feature extractor and a domain discriminator. According to the adversarial feature alignment method, the enhanced feature vector is mapped to a common feature space through the feature extractor, the data domain to which the feature belongs is discriminated through the domain discriminator, and the maximum mean difference between the source domain feature and the target domain feature in the reproducing kernel Hilbert space is calculated based on the Gaussian kernel function, and the alignment feature representation is obtained by minimizing the maximum mean difference.

3. The method according to claim 2, characterized in that The mutual information maximization objective function based on the cosine similarity is constructed as shown in the following formula: ； Among them, J represents the objective function value, α(i, j) represents the attention weight of position (i, j), sim(x i ,y j ) represents the eigenvector x of the two modes at position (i, j) i and j The cosine similarity between them, λ represents the balance factor, d(x i ,y j ) represents the positive sample pair (x i ,y j ), d(x i ,y k ) represents the negative sample pair (x i ,y k ), M represents the boundary value in the contrast loss, which is used to control the distance difference between positive and negative samples.

4. The method according to claim 1, characterized in that The aligned feature representation is added to the pre-built graph attention diffusion model and represented as a spatiotemporal graph structure. The spatiotemporal graph structure is denoised through a diffusion process to obtain an accurate state representation. The state representation is input into an adaptive causal convolutional network, and multi-scale temporal features are extracted through a dilated convolutional layer with a variable receptive field. The continuous-time dynamic system in the neural ordinary differential equation network is combined for solution, and the state evolution trajectory is obtained, including: The aligned feature representation is input into a pre-built graph attention diffusion model, a spatiotemporal graph structure is constructed by setting state feature nodes of multiple time steps and edges of temporal dependencies between states, attention weights between nodes are calculated according to the graph attention layer, and node features are aggregated to obtain context node representations containing long-term and short-term dependencies; A diffusion process is performed on the spatiotemporal graph structure through multiple diffusion steps, and in each diffusion step, Gaussian noise is added to the node features and the features are updated through a multi-layer graph neural network. The feature update process of the graph neural network is calculated based on the node features of the previous diffusion step, the Gaussian noise intensity, and the time interval of adjacent time steps. The inverse diffusion process is performed according to the iteratively updated node features to remove the noise, so as to obtain an accurate state representation; The precise state representation is input into an adaptive causal convolutional network, the adaptive causal convolutional network is constructed by setting a plurality of dilated convolutional layers with different dilated rates, time series features of multiple scales are extracted according to the dilated convolutional layers with different dilated rates, and the time series features of the multiple scales are feature spliced ​​to obtain multi-scale time series features; The multi-scale time series features are input into the neural ordinary differential equation network, and the right function of the ordinary differential equation is parameterized by setting a fully connected network with multiple hidden layers. The high-order Runge-Kutta method is used to perform numerical integration at multiple time points to solve the ordinary differential equation and obtain the state evolution trajectory.

5. The method according to claim 1, characterized in that The posterior distribution of the state evolution trajectory is determined by the variational inference network and the posterior distribution is mapped to the standard normal distribution in combination with the autoregressive flow model to obtain the probability density. The probability density is added to the conditional neural process through the neural process network, the context encoder and the target decoder are combined to model the data distribution, and the dynamic point process model is combined to predict the time type and time, and the state prediction sequence containing uncertainty is obtained, including: The state evolution trajectory is input into a variational inference network constructed by setting a fully connected network with multiple hidden layers, and the variational distribution parameters are calculated according to minimizing the reconstruction error and the Kuhler-Leibler divergence, and the variational distribution parameters are input into an autoregressive flow model composed of multiple coupling layers, and the posterior distribution is mapped to a standard normal distribution to obtain a probability density function of the state evolution; Inputting the probability density function of the state evolution into a conditional neural process, constructing a context encoder of the conditional neural process by setting a multi-layer long short-term memory network, encoding the state sequence into a context vector, and constructing a target decoder of the conditional neural process by setting a multi-layer long short-term memory network according to the splicing result of the context vector and a random vector sampled from a standard normal distribution, to generate a target state sequence; The target state sequence is input into the dynamic point process model, and a fully connected network parameterized intensity function is set up with multiple hidden layers. The time and type of the next state transfer are predicted according to the current time and state. The parameters of the intensity function are optimized by maximizing the log-likelihood function of the time series. The intensity function is sampled using the Monte Carlo method to obtain the time and type of state transfer. The time and type of state transfer are combined with the target state sequence to generate a state prediction sequence containing uncertainty.

6. The method according to claim 1, characterized in that The state prediction sequence is added to the first discriminator of the hierarchical generative adversarial imitation learning network, the implicit distribution of the expert strategy is determined through adversarial training learning and the expert strategy is decomposed into subtasks at multiple abstract levels, the state transition diagram of the return task is constructed in combination with the option chain, the state evolution gradient is calculated by the dual proximal strategy optimization algorithm and added to the second discriminator based on neural architecture search, the optimal feature extraction structure is determined, the optimal feature extraction structure and the corresponding network parameters are updated by multiple parallel evaluation units in the distributed trust region evolution strategy algorithm to obtain the optimal parameter configuration, the optimal parameter configuration is mapped to a multi-dimensional feature vector corresponding to the return state, the state prediction is performed through the multi-dimensional evaluation network, the uncertainty of the state prediction is determined by the sequential Monte Carlo method, and the final judgment result includes: The image sequence, position information, and posture data of the drone's historical return process are collected as expert strategy data, and the state prediction sequence containing uncertainty is input into the first discriminator of the hierarchical generative adversarial imitation learning network. The input state sequence is encoded by the generator of the hierarchical generative adversarial imitation learning network and a state sequence that imitates the expert strategy is generated. The adversarial loss between the state sequence and the real state sequence of the expert strategy is calculated by the first discriminator, and the high-level features of the state sequence are extracted by the second discriminator. The parameters of the generator and the discriminator are alternately updated by back propagation and gradient descent to obtain the implicit distribution of the expert strategy. Based on the option chain method, the expert strategy is decomposed into subtasks at multiple abstract levels, and a state transition diagram is constructed. By setting nodes representing subtasks in the state transition diagram, edges between nodes represent migration probabilities and migration costs between subtasks, and using a dynamic programming algorithm to perform path search and planning on the state transition diagram, the optimal option sequence for completing the original task is obtained; Input the implicit distribution of the expert strategy into the dual proximal strategy optimization algorithm, set the objective function of the drone return state prediction to minimize the deviation between the predicted state sequence and the true state sequence as the original problem, transform the original problem into a dual problem by introducing a dual variable, alternately optimize the prediction deviation minimization problem and the corresponding dual problem to calculate the state evolution gradient, input the state evolution gradient into the second discriminator based on neural architecture search, construct a candidate network architecture by setting the convolution kernel size search space and the fully connected layer neuron number search space, evaluate different network architectures using the state evolution gradient as the optimization target, and select the network architecture with the smallest state evolution gradient as the optimal feature extraction structure; The optimal feature extraction structure is input into the distributed trust domain evolution strategy algorithm, and a parallel evaluation unit is constructed by dividing the evolutionary population into multiple sub-populations, and fitness evaluation and parameter update operations are performed on each sub-population in parallel to obtain a local optimal solution. The individual migration strategy is determined by calculating the trust between sub-populations, and the individuals with the highest fitness in each sub-population are migrated to other sub-populations to update the global optimal solution, thereby obtaining the optimal network parameter configuration; The optimal network parameter configuration is mapped to a multidimensional feature vector corresponding to the return state through a fully connected network, a multidimensional evaluation network is constructed by setting up multiple parallel evaluation subnetworks, the evaluation subnetwork is used to extract and evaluate state features from different dimensions, the state features are repeatedly sampled multiple times using the sequential Monte Carlo method, the sampling results are input into the multidimensional evaluation network to obtain multiple prediction results, the statistical distribution characteristics of the prediction results are calculated to obtain the expected value and uncertainty measure of the state prediction, and the final return state determination result is generated.

7. The method according to claim 6, characterized in that Fitness evaluation and parameter update operations are performed in parallel on each sub-population to obtain the local optimal solution. The individual migration strategy is determined by calculating the trust between sub-populations, including: Construct a multi-objective fitness evaluation function to calculate the fitness score of individuals in each sub-population. The fitness evaluation function includes a prediction error term, a parameter complexity penalty term, and a network structure penalty term. The parameter distribution similarity and historical iteration effect between sub-populations are calculated to establish a trust matrix between sub-populations. Arrange the trust matrix in descending order according to the trust value, take the two sub-populations with the largest trust value as the first group, and take the two sub-populations with the largest trust value in the remaining sub-populations as the second group, and repeat the grouping until all sub-populations are grouped. In each group, set the sub-population with a large trust value as the source population, and the sub-population with a small trust value as the target population; Calculate the trust difference between the source population and the target population, use the product of the trust difference and the preset benchmark migration number as the actual number of migrated individuals, form an individual migration strategy based on the source population, the target population and the number of migrated individuals, and select the individual with the highest fitness score from the source population as the migration individual based on the individual migration strategy, and replace the individual with the lowest fitness score in the target population with the migration individual; Perform local parameter adjustment on the migrated individuals, where the adjustment amplitude is equal to the difference in trust between the source population and the target population multiplied by the preset adjustment coefficient. Recalculate the trust between sub-populations based on the adjusted individuals, update the trust matrix, and execute the individual migration strategy repeatedly until the preset termination condition is reached.

8. A drone return detection system based on AI recognition, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain multi-source image data in the process of returning the drone and add it to the improved window-shifting hierarchical transformer model, extract multi-scale feature maps through multiple attention windows and add them to the cross-modal contrastive learning module, obtain semantic consistency features by calculating the mutual information in different views and add them to the dynamic routing capsule network, update the routing weights by calculating the coupling coefficients between feature vectors and output a high-dimensional feature vector containing target posture information, add the high-dimensional feature vector to a feature enhancement module including a channel recalibration unit and a spatial recalibration unit, update the adaptive weights by calculating global statistical information to obtain an enhanced feature vector, align the enhanced feature vectors according to the adversarial feature alignment method to minimize the maximum mean difference between different data distributions, and obtain an aligned feature representation; The second unit is used to add the alignment feature representation to a pre-built graph attention diffusion model and represent it as a spatiotemporal graph structure, denoise the spatiotemporal graph structure through a diffusion process to obtain an accurate state representation, input the state representation into an adaptive causal convolutional network, extract multi-scale time series features through a hole convolution layer with a variable receptive field, solve it in combination with the continuous-time dynamic system in the neural ordinary differential equation network to obtain a state evolution trajectory, determine the posterior distribution of the state evolution trajectory through a variational inference network and map the posterior distribution to a standard normal distribution in combination with an autoregressive flow model to obtain a probability density, add the probability density to a conditional neural process through a neural process network, model data distribution in combination with a context encoder and a target decoder, predict time type and time in combination with a dynamic point process model, and obtain a state prediction sequence containing uncertainty; The third unit is used to add the state prediction sequence to the first discriminator of the hierarchical generative adversarial imitation learning network, determine the implicit distribution of the expert strategy through adversarial training learning and decompose the expert strategy into sub-tasks at multiple abstract levels, construct a state transition diagram of the return task in combination with the option chain, calculate the state evolution gradient through the dual proximal strategy optimization algorithm and add it to the second discriminator based on neural architecture search, determine the optimal feature extraction structure, update the optimal feature extraction structure and the corresponding network parameters through multiple parallel evaluation units in the distributed trust domain evolution strategy algorithm to obtain the optimal parameter configuration, map the optimal parameter configuration to a multi-dimensional feature vector corresponding to the return state, perform state prediction through a multi-dimensional evaluation network, determine the uncertainty of the state prediction through a sequential Monte Carlo method, and obtain a final judgment result.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images