Low-altitude economic unmanned aerial vehicle data processing method and system based on large model

Abstract

The invention discloses a low-altitude economic unmanned aerial vehicle data processing method and system based on a large model, and relates to the technical field of low-altitude economic data processing, and the method comprises the steps: extracting low-altitude multi-target decision features of a low-altitude semantic decision map, carrying out the fuzzy reasoning of the low-altitude multi-target decision features, and generating an anti-interference control instruction set; based on the anti-interference control instruction set, fault logs and task execution data during operation of the unmanned aerial vehicle are collected, and an evolution strategy library is generated through association rule mining; and performing causal analysis according to the evolution strategy library, generating a cluster coordination rule upgrade package, performing dynamic verification on the cluster coordination rule upgrade package, and outputting a self-healing strategy. According to the method, the reliability and the safety of task execution of the unmanned aerial vehicle are improved by constructing the semantic association large model and generating the anti-interference control instruction set.

Description

Technical Field

[0001] This invention relates to the field of low-altitude economic data processing technology, and in particular to a method and system for processing low-altitude economic unmanned aerial vehicle (UAV) data based on a large model. Background Technology

[0002] In recent years, with the booming development of the low-altitude economy, unmanned aerial vehicles (UAVs), as an important carrier in the low-altitude field, have been widely used in many fields such as logistics transportation, agricultural and forestry plant protection, surveying and mapping, and emergency rescue. Alongside this, UAV data processing methods have also made significant progress. In terms of data acquisition, various high-precision sensors have emerged, capable of acquiring multi-source data including visible light images, lidar point clouds, IMU attitude data, GPS position data, and turbulence frequency domain signals, providing rich information for comprehensively perceiving the UAV's flight environment and mission status. Regarding data processing methods, traditional machine learning and simple rule-based algorithms are widely used in the analysis and decision-making of UAV data, which has improved the autonomy and mission execution capabilities of UAVs to a certain extent.

[0003] However, existing technologies still have some shortcomings. On the one hand, traditional data processing methods struggle to effectively integrate multi-source heterogeneous data and fail to fully explore the inherent relationships and semantic information between different types of data. This results in an incomplete and superficial understanding of the UAV's flight environment and mission, affecting the accuracy and reliability of decision-making. On the other hand, existing methods lack adaptive and self-optimizing capabilities when facing complex and ever-changing low-altitude environments and dynamic missions. They struggle to adjust decision-making strategies in real time according to actual conditions and cannot effectively cope with unexpected situations and uncertainties. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a low-altitude economic UAV data processing method based on a large model to solve the problems of difficulty in effectively fusing multi-source heterogeneous data and poor adaptive decision-making ability of UAVs.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a data processing method for low-altitude economic unmanned aerial vehicles based on a large model, comprising,

[0008] Collect low-altitude dynamic sensing data and preprocess it; perform structured transformation on the preprocessed low-altitude dynamic sensing data to generate environmental fingerprints.

[0009] The environmental fingerprint is input into the semantic association big model, the dynamic topology perception layer performs task-space fusion evolution, the multimodal intent decoupling layer performs hierarchical optimization, and a low-altitude semantic decision map is generated.

[0010] Extract low-altitude multi-target decision features from the low-altitude semantic decision map, and perform fuzzy reasoning on the low-altitude multi-target decision features to generate an anti-disturbance control instruction set;

[0011] Based on the anti-disturbance control instruction set, fault logs and task execution data of UAVs during operation are collected, and an evolution strategy library is generated through association rule mining.

[0012] Based on the evolution strategy library, causal analysis is performed to generate a cluster collaboration rule upgrade package. The cluster collaboration rule upgrade package is dynamically verified, and a self-healing strategy is output.

[0013] As a preferred embodiment of the low-altitude economic UAV data processing method based on a large model described in this invention, the low-altitude dynamic perception data includes visible light image data, lidar point cloud data, IMU attitude data, GPS location data, turbulence frequency domain signal data, and mission command text data.

[0014] The preprocessing includes data cleaning, format conversion, deduplication, normalization, and outlier handling.

[0015] As a preferred embodiment of the low-altitude economic UAV data processing method based on a large model described in this invention, the specific steps for generating the environmental fingerprint are as follows:

[0016] Dynamic scene adaptive features are extracted from visible light image data and lidar point cloud data. The dynamic scene adaptive features are then aligned with coordinates to generate a motion-invariant feature matrix.

[0017] Time synchronization and coordinate transformation are performed on IMU attitude data and GPS location data. By obtaining the spatial relationship features of IMU attitude data and GPS location data, a cooperative topological adjacency matrix is ​​constructed.

[0018] Frequency domain analysis is performed on turbulent frequency domain signal data to obtain turbulence spectrum feature vectors;

[0019] Semantically compress the task instruction text data to obtain the intent encoding vector;

[0020] By integrating motion-invariant feature matrices, cooperative topological adjacency matrices, intent encoding vectors, and turbulence spectrum feature vectors, environmental fingerprints are obtained.

[0021] As a preferred embodiment of the low-altitude economic UAV data processing method based on a large model described in this invention, the semantic association large model is constructed using the following steps:

[0022] Invoke and initialize the fully connected network and the gated linear network to build a dynamic topology perception layer and a multimodal intent decoupling layer;

[0023] By using the cross-entropy loss function to perform backpropagation and gradient update on the dynamic topology perception layer and the multimodal intent decoupling layer, a large semantic association model is constructed.

[0024] As a preferred embodiment of the low-altitude economic UAV data processing method based on a large model described in this invention, the specific steps for generating the low-altitude semantic decision map are as follows:

[0025] The environmental fingerprint is input into the semantic association model, and the dynamic topology perception layer performs task-space fusion evolution through dynamic graph convolution to generate a dynamic collaborative constraint graph.

[0026] The multimodal intent decoupling layer is optimized through a gating attention mechanism to form multi-dimensional balanced decision features;

[0027] The dynamic collaborative constraint graph and multi-dimensional balanced decision features are integrated through the feature splicing channel to generate a low-altitude semantic decision map.

[0028] As a preferred embodiment of the low-altitude economic UAV data processing method based on a large model described in this invention, the specific steps for extracting low-altitude multi-target decision features from the low-altitude semantic decision map are as follows:

[0029] Multi-dimensional feature decoupling is performed on the low-altitude semantic decision map to extract safety constraint features and efficiency optimization features;

[0030] Perform task semantic association on the low-altitude semantic decision map and extract task semantic features;

[0031] By integrating safety constraint features, efficiency optimization features, and task semantic features, low-altitude multi-objective decision features are generated.

[0032] As a preferred embodiment of the low-altitude economic UAV data processing method based on a large model described in this invention, the specific steps for generating the anti-disturbance control instruction set are as follows:

[0033] The low-altitude multi-target decision-making features are fuzzified to generate a multi-dimensional fuzzy input vector.

[0034] A membership degree weighting function is applied to the multidimensional fuzzy input vector to calculate the disturbance rejection correction amount. The disturbance rejection correction amount is then encapsulated in a structured manner to generate a disturbance rejection control instruction set.

[0035] As a preferred embodiment of the low-altitude economic UAV data processing method based on a large model described in this invention, the specific steps for generating the evolution strategy library are as follows:

[0036] The UAV executes an anti-disturbance control command set, synchronously collects fault logs and mission execution data during UAV operation, and generates a related dataset.

[0037] Extract the fault propagation patterns and task efficiency features from the associated datasets to generate a fault-efficiency correlation feature table;

[0038] We perform association rule mining on the fault-efficiency association feature table to generate an initial policy library, and optimize the policy weights based on the association dataset to generate an evolved policy library.

[0039] As a preferred embodiment of the low-altitude economic UAV data processing method based on a large model described in this invention, the output self-healing strategy comprises the following specific steps.

[0040] Extract historical evolution strategies from the evolution strategy library, analyze the causal relationship between historical evolution strategies and cluster collaboration effects using causal inference algorithms, and output a causal analysis report;

[0041] Perform causal inference and logical reconstruction on the causal analysis report to generate a cluster collaboration rule upgrade package;

[0042] Verify the cluster collaboration effect of the cluster collaboration rule upgrade package, dynamically calibrate it according to preset indicator thresholds, and output a self-healing strategy.

[0043] Secondly, the present invention provides a low-altitude economic unmanned aerial vehicle (UAV) data processing system based on a large model, including a data acquisition module, a semantic association module, a fuzzy reasoning module, a strategy mining module, and a causal analysis module.

[0044] The data acquisition module is used to collect low-altitude dynamic sensing data and preprocess it, and then perform structured transformation on the preprocessed low-altitude dynamic sensing data to generate environmental fingerprints.

[0045] The semantic association module is used to input environmental fingerprints into the semantic association big model, the dynamic topology perception layer performs task-space fusion evolution, the multimodal intent decoupling layer performs hierarchical optimization, and generates a low-altitude semantic decision map.

[0046] The fuzzy inference module is used to extract the low-altitude multi-target decision features from the low-altitude semantic decision map, and to perform fuzzy inference on the low-altitude multi-target decision features to generate an anti-disturbance control instruction set.

[0047] The strategy mining module is used to collect fault logs and task execution data of UAVs based on the anti-disturbance control instruction set, and generate an evolution strategy library through association rule mining.

[0048] The causal analysis module is used to perform causal analysis based on the evolution strategy library, generate cluster collaboration rule upgrade packages, dynamically verify the cluster collaboration rule upgrade packages, and output self-healing strategies.

[0049] The beneficial effects of this invention are as follows: by using a semantic association model to perform task-space fusion evolution and hierarchical optimization, the intrinsic correlation between data is explored, improving the accuracy and adaptability of UAVs in making decisions in complex scenarios; simultaneously, by generating an anti-disturbance control instruction set, dynamic optimization and adaptive evolution are performed, improving the reliability, safety and success rate of UAVs in mission execution. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of a data processing method for low-altitude economic drones based on a large model.

[0052] Figure 2 This is a schematic diagram of a low-altitude economic unmanned aerial vehicle (UAV) data processing system based on a large model.

[0053] Figure 3 A flowchart for generating the disturbance rejection control instruction set.

[0054] Figure 4 The flowchart for generating the strategy library and self-healing strategy. Detailed Implementation

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0058] Reference Figures 1-4 This is one embodiment of the present invention, which provides a data processing method for low-altitude economic unmanned aerial vehicles based on a large model, including the following steps:

[0059] S1. Collect low-altitude dynamic sensing data and preprocess it. Perform structured transformation on the preprocessed low-altitude dynamic sensing data to generate environmental fingerprints.

[0060] S1.1 Low-altitude dynamic sensing data includes visible light image data, lidar point cloud data, IMU attitude data, GPS location data, turbulence frequency domain signal data, and mission command text data;

[0061] It should be noted that visible light image data is acquired by RGB (color mode) and panoramic cameras mounted on the UAV fuselage; lidar point cloud data is generated by laser pulses emitted by lidar on the top and bottom of the UAV; IMU (Inertial Measurement Unit) attitude data is acquired by an IMU integrating accelerometers and gyroscopes and fused to output attitude angles; GPS position data is acquired by receiving satellite signals through a GNSS (Global Navigation Satellite) receiver on the top; turbulence frequency domain signal data is acquired by sensors such as ultrasonic anemometers at different locations on the UAV fuselage to collect airflow fluctuations; and mission command text data is acquired through a ground control station.

[0062] S1.2 Preprocessing includes data cleaning, format conversion, deduplication, normalization, and outlier handling;

[0063] It should be noted that the collected low-altitude dynamic sensing data, including visible light image data, lidar point cloud data, IMU attitude data, GPS location data, turbulence frequency domain signal data, and mission command text data, undergoes data cleaning to remove incomplete, erroneous, or irrelevant data records. Format conversion is performed on the collected low-altitude dynamic sensing data, including visible light image data, lidar point cloud data, IMU attitude data, GPS location data, turbulence frequency domain signal data, and mission command text data, to ensure that all data sources adhere to a unified format standard for subsequent processing. Deduplication is performed on the collected low-altitude dynamic sensing data, including visible light image data, lidar point cloud data, IMU attitude data, GPS location data, turbulence frequency domain signal data, and mission command text data, to eliminate duplicate data entries and ensure the uniqueness of the dataset. A normalization step is used to map data values ​​from different sources to the same scale, avoiding undue influence on the results due to significant differences in magnitude of certain features. Outlier handling identifies and corrects data points that deviate from the normal range, ensuring the accuracy and reliability of the dataset.

[0064] S1.3 Extract dynamic scene adaptability features from visible light image data and lidar point cloud data, align the coordinates of the dynamic scene adaptability features, and generate a motion invariant feature matrix;

[0065] It should be noted that visual features such as edges and textures from visible light image data and geometric features such as key points (e.g., edges, corners, high curvature areas) and normal vectors from lidar point cloud data are extracted respectively. Dynamic scene adaptability features that are not sensitive to changes in illumination and viewpoint shifts (such as undeformed target contours in visible light image data) are selected. Coordinate alignment is completed through timestamp registration to eliminate spatiotemporal misalignment caused by UAV movement. The registered dynamic scene adaptability features are mapped into matrix form according to spatial position to generate motion-invariant feature matrix.

[0066] S1.4. Perform time synchronization and coordinate transformation on IMU attitude data and GPS position data, and construct a cooperative topological adjacency matrix by obtaining the spatial relationship characteristics of IMU attitude data and GPS position data.

[0067] It should be noted that time synchronization and coordinate transformation are performed on the IMU attitude data and GPS position data to convert the local coordinate system of the IMU (with the UAV's centroid as the origin) to the global coordinate system of the GPS. Then, spatial relationship features are extracted, including the position difference, attitude angle difference and relative velocity between the IMU attitude data and the GPS position data. Based on the spatial relationship features, a cooperative topological adjacency matrix is ​​constructed, where the rows and columns of the cooperative topological adjacency matrix correspond to the UAV and the environment nodes, respectively, and the element values ​​are the cosine similarity of the spatial relationship features between the environment nodes.

[0068] S1.5 Perform frequency domain analysis on the turbulence frequency domain signal data to obtain the turbulence spectrum feature vector;

[0069] It should be noted that high-frequency noise (e.g., above 10Hz) and low-frequency trend terms (e.g., below 0.1Hz) in the turbulence frequency domain signal data are removed by bandpass filtering, while retaining the effective frequency bands that reflect turbulence fluctuations (e.g., vortex generation and dissipation). The processed turbulence frequency domain signal data is converted into a frequency domain energy distribution using short-time Fourier transform, and turbulence spectrum feature vectors such as frequency domain energy entropy, dominant frequency peak, and frequency band energy ratio are further extracted from the frequency domain energy distribution.

[0070] S1.6. Perform semantic compression on the task instruction text data to obtain the intent encoding vector;

[0071] It should be noted that the task instruction text data is segmented using regular expression matching, and function words and punctuation marks in the stop word list are filtered out to obtain a cleaned priority word sequence. Through context co-occurrence statistics based on a sliding window, the priority word sequence is transformed into a high-dimensional dense vector set containing task priorities. The high-dimensional dense vector set is then weighted and aggregated to compress the vector dimensions, outputting a low-dimensional compressed vector. The low-dimensional compressed vector is then matched with a predefined intent feature template using cosine similarity to generate an intent encoding vector. Each dimension of the intent encoding vector corresponds to an intent category, and the value of each dimension corresponds to the cosine similarity score between the current task instruction text data and the corresponding intent category.

[0072] It should also be noted that the stop word list is defined based on the co-occurrence statistics of high-frequency meaningless words (such as function words and punctuation marks) in historical corpora. Specifically, by performing word frequency statistical analysis on historical task instruction text data, function words and punctuation marks that appear frequently but do not make a substantial contribution to the expression of semantic information such as task priority are selected, and high-frequency meaningless words are integrated to form the stop word list.

[0073] The intent feature template is defined based on key terms and clustering analysis results of high-frequency meaningful terms (such as words that appear more than 500 times and rank in the top 10% of TF-IDF weight) in historical task instruction texts. Specifically, it extracts high-frequency key terms that are representative of task intent from historical task instruction text data (such as "reconnaissance", "obstacle avoidance", "material transportation", etc.), analyzes the co-occurrence patterns of key terms in historical task instruction text data (such as "reconnaissance" is often associated with "target area A") and semantic relationships (such as "obstacle avoidance" and "turbulence anomaly" have a causal relationship), clusters key terms with similar functions and scene associations into associated feature groups, and standardizes the associated feature groups to form the intent feature template.

[0074] S2. Input the environmental fingerprint into the semantic association big model, perform task-space fusion evolution in the dynamic topology perception layer, perform hierarchical optimization in the multimodal intent decoupling layer, and generate a low-altitude semantic decision map.

[0075] S2.1 Call and initialize the fully connected network and the gated linear network to build a dynamic topology perception layer and a multimodal intent decoupling layer;

[0076] It should be noted that the fully connected network sets the input dimension to the joint dimension of the environmental fingerprint (e.g., 680-dimensional), and contains two hidden fully connected sub-layers. Each fully connected sub-layer is followed by a ReLU activation function to enhance the nonlinear expression, mapping the environmental fingerprint into a high-dimensional dense fusion feature. A gated linear network is initialized synchronously, consisting of three gated networks connected in series. Each gated network contains a linear transformation layer and a Sigmoid activation function. The fusion feature output by the fully connected network is dynamically adjusted through gate weights (initialized with small random numbers), preserving key information (such as the vortex features of the turbulence spectrum feature vector) and suppressing redundant information (such as environmental background noise), outputting the adjusted feature. The dynamic topology sensing layer takes the adjusted feature output by the gated linear network as input and, based on a self-attention mechanism, inputs the adjusted feature to three independent linear layers, performing Q (query) space mapping, K (key) space mapping, and V (value) space mapping respectively, generating corresponding Q, K, and V matrices to the input. The feature linear transformation performs Q-space mapping, K-space mapping, and V-space mapping respectively. The Q matrix is ​​used to capture the "topology query requirements" of the control features (e.g., whether the current sensor anomaly is related to historical anomalies of neighboring cameras); the K matrix is ​​used to encode the "topology key information" of other data (e.g., the spatial distribution characteristics of historical anomalies of neighboring cameras); and the V matrix is ​​used to store the "topology value content" of other data (e.g., the ambient temperature and humidity when historical anomalies occurred). Next, the correlation score between each data point is obtained by the dot product of the Q and K matrices, and then normalized using the Softmax function to generate the attention weight matrix. Then, the V matrix is ​​weighted and aggregated using causal attention weights: the attention weight matrix and the V matrix are summed element-wise to obtain the context-aware feature of each data point. The context-aware feature is then mapped to the target dimension through the ReLU nonlinear activation function and a fully connected layer to generate a dynamic topology-aware feature matrix. The multimodal intent decoupling layer takes the dynamic topology-aware feature matrix of the dynamic topology-aware layer as input and separates independent intents of different modalities through a multi-head attention mechanism: it divides the subspace according to the modality type of the environmental fingerprint, and each subspace extracts local features of a specific modality (such as the vortex intensity features of the turbulence spectrum feature vector) through an independent linear transformation layer; the output is a structured decoupling intent feature that reflects the independent intents of the multimodalities, thus completing the construction of the dynamic topology-aware layer and the multimodal intent decoupling layer.

[0077] S2.2. Utilize the cross-entropy loss function to perform backpropagation and gradient update on the dynamic topology perception layer and the multimodal intent decoupling layer to construct a large semantic association model;

[0078] It should be noted that the dynamic topology-aware feature matrix output by the dynamic topology-aware layer and the decoupled intent features output by the multimodal intent decoupling layer are concatenated to generate a comprehensive feature vector. Intent classification prediction is then performed on the comprehensive feature vector. The comprehensive feature vector is then linearly transformed and normalized to obtain the probability distribution map of each intent category. The cross-entropy loss function is used to obtain the error value between the probability distribution map and the intent category, constructing a prediction error matrix. Backpropagation is performed on the prediction error matrix to generate the gradient tensors of the dynamic topology-aware layer, the multimodal intent decoupling layer, and the classification head. The Adam optimizer is used to update the gradient tensors, obtaining the updated network weights. Based on the updated network weights, layer normalization is used to perform hyperparameter tuning on the feature dimensions of the dynamic topology-aware layer and the multimodal intent decoupling layer. Residual connections are then used to fuse the dynamic topology-aware layer and the multimodal intent decoupling layer across layers, enhancing the transitivity of semantic association. Finally, the constructed semantic association model is output.

[0079] Next, the semantic association model is trained. Specifically, the environmental fingerprints are first divided into a sample set, a training set, and a validation set according to their data type (e.g., a 7:2:1 ratio). The sample set is then augmented with spatiotemporal features using linear interpolation and standardized using batch normalization to eliminate dimensional differences and stabilize the training process, forming enhanced standard samples. The Adam optimizer is used to update the parameters of these enhanced standard samples, while early stopping is applied to monitor the training process: using the validation set loss as the monitoring metric, early stopping is triggered when the validation set loss fails to decrease for five consecutive rounds to avoid overfitting, and intermediate parameters are recorded simultaneously for each round. On the validation set, the cross-entropy loss function is used to quantify the loss of the intermediate parameters, obtaining the classification error on the validation set. Training terminates when the classification error on the validation set does not fall below the convergence threshold for eight consecutive rounds, and the trained semantic association model is output simultaneously.

[0080] It should be noted that the convergence threshold is defined based on the rate of change of the prediction error on the validation set, and its value ranges from 0.0001 to 0.01.

[0081] S2.3. Input the environmental fingerprint into the semantic association big model, and the dynamic topology perception layer performs task-space fusion evolution through dynamic graph convolution to generate a dynamic collaborative constraint graph.

[0082] It should be noted that, taking the environmental fingerprint as input, the dynamic topology perception layer generates a dynamic topology perception feature matrix based on a self-attention mechanism, constructs an initial task-space association graph, where associated nodes in the initial task-space association graph are defined as key environmental elements (such as sensor location points), and the features of associated nodes are fused with environmental spatial coordinates and task-related attributes output by the semantic association model (such as "obstacles to be avoided"); edges are defined as the spatial distance between environmental elements, and the edge weights are initialized as the product of the inverse of the spatial distance and the task relevance. The dynamic topology perception layer uses dynamic graph convolution to weight and aggregate the edge weights of each associated node feature in the initial task-space association graph with the relevant features of neighboring nodes to obtain local context features. Based on the local context features, dynamic attention weights between associated nodes are obtained through a graph attention mechanism, and the edge weights are adjusted according to the dynamic attention weights to generate a dynamic collaborative constraint graph.

[0083] It should also be noted that the task relevance is based on the spatial location and attribute definition of task instruction text data and environmental elements. In specific operation, the functional attributes of environmental elements in the environmental fingerprint (such as the spatial coordinates of sensor locations) are obtained and semantically matched with the task instruction text data (e.g., the matching degree between the keyword "reconnaissance" and the functional attribute of "sensor"). The co-occurrence frequency of environmental elements and intent categories in historical task instruction text data is counted (e.g., the proportion of the number of times a sensor in a certain area appears in the "reconnaissance task" to the total number of times the task is executed). The co-occurrence frequency is normalized to obtain the historical co-occurrence frequency. The semantic matching result and the historical co-occurrence frequency are weighted and summed with equal weights to generate the task relevance.

[0084] S2.4 The multimodal intent decoupling layer is optimized through a gating attention mechanism to form multi-dimensional balanced decision features;

[0085] It should be noted that the environmental fingerprint fusion based on the dynamic collaborative constraint graph is optimized hierarchically through a three-layer gated attention mechanism: The first-layer gated attention mechanism inputs the associated node features and collaborative topological adjacency matrix of the dynamic collaborative constraint graph into the linear transformation layer to obtain the spatial-task association score. After Sigmoid gating adjustment, the spatial-task constraint features (such as the "adjacency weight between target region A and sensor S1") are output. The second-layer gated attention mechanism projects the motion invariant feature matrix and the turbulence spectrum feature vector to the shared feature space to obtain the cross-modal attention score. The contribution ratio of each modality to the decision is dynamically adjusted through the gating weight (such as enhancing the contribution of the image target contour to the "target detection" task), and the modality enhancement features are output. The third-layer gated attention mechanism weights and aggregates the spatial-task constraint features and modality enhancement features output from the first two layers with the intent encoding vector. The decision weights of each dimension are generated through Softmax normalization and mapped to the target dimension, outputting multi-dimensional balanced decision features.

[0086] S2.5 Integrate the dynamic collaborative constraint graph and multi-dimensional balanced decision features through the feature splicing channel to generate a low-altitude semantic decision guide map;

[0087] It should be noted that in the feature concatenation channel, spatiotemporal alignment and feature fusion are performed on the dynamic collaborative constraint graph and the multi-dimensional balanced decision features: a graph-level global average pooling layer is used to aggregate features of the dynamic collaborative constraint graph to generate graph-level feature vectors. At the same time, the multi-dimensional balanced decision features are projected to the dimension matching the graph-level feature vectors through a linear transformation layer to eliminate feature dimension differences. Then, feature fusion is performed through a dual-channel attention gating layer: the graph-level feature vectors are compressed by channels, while the multi-dimensional balanced decision features are smoothed by a Gaussian sliding window. The graph-level feature vectors and multi-dimensional balanced decision features are then weighted and fused through the feature concatenation channel to generate a hybrid feature tensor. The hybrid feature tensor is then subjected to nonlinear feature transformation and dimension enhancement using the ReLU activation function to generate high-dimensional nonlinear features. The spatial resolution of the high-dimensional nonlinear features is restored through transposed convolution, outputting a spatial structure feature map. The spatial structure feature map is then semantically segmented to generate a low-altitude semantic decision map.

[0088] S3. Extract the low-altitude multi-target decision features from the low-altitude semantic decision map, and perform fuzzy reasoning on the low-altitude multi-target decision features to generate an anti-disturbance control instruction set.

[0089] S3.1 Decouple the low-altitude semantic decision map into multiple dimensions and extract safety constraint features and efficiency optimization features;

[0090] It should be noted that when extracting safety constraint features, the dynamic collaborative constraint graph and the collaborative topology adjacency matrix are used as inputs. A linear transformation layer projects the edge weights of the dynamic collaborative constraint graph and the positional differences in the collaborative topology adjacency matrix to the safety dimension space to obtain a safety attention score. After Sigmoid gating adjustment, the safety constraint information of high-risk nodes (such as "edge weight of obstacle region B is 0.1") is retained, while the redundant information of low-risk nodes (such as "edge weight of open area is 0.9") is suppressed, and the safety constraint features are output. When extracting efficiency optimization features, the associated node features and edge weights of the dynamic collaborative constraint graph are used as inputs. A linear transformation layer projects the task priority labels (such as "high priority" marked as 1) and edge weights in the associated node features to the efficiency dimension space to obtain an efficiency attention score. After Softmax gating normalization, efficiency attention weights are generated. Based on the efficiency attention weights, the associated node features and edge weights of the dynamic collaborative constraint graph are weighted and aggregated to output efficiency optimization features.

[0091] S3.2 Perform task semantic association on the low-altitude semantic decision map and extract task semantic features;

[0092] It should be explained that the associated node features and intent encoding vectors of the low-altitude semantic decision map are projected onto the shared semantic space through a linear transformation layer, and the cosine similarity between the associated node features and the intent encoding vectors is used as the task-node semantic association score. At the same time, the decision weights in the multi-dimensional balanced decision features and the edge weights of the dynamic collaborative constraint graph are weighted and fused to generate the task-space collaborative semantic score. The associated node features and the task-space collaborative semantic score are dynamically aggregated through a graph attention mechanism: the task-node semantic association score is used as the attention coefficient to aggregate the task-related attributes and intent encoding vectors of the neighboring nodes to the current associated node. At the same time, the task-space collaborative semantic score is used as the gating weight to suppress the semantic interference of non-task-related nodes and obtain the task semantic features.

[0093] S3.3. Perform fuzzification processing on the low-altitude multi-target decision features to generate a multi-dimensional fuzzy input vector;

[0094] It should be noted that the low-altitude multi-target decision features are fuzzified using a three-level decision threshold. Specifically, the task-environment correlation is fuzzified based on the task semantic features (task priority, efficiency optimization semantic correlation, and safety constraint features such as the cosine similarity of the position difference between IMU attitude data and GPS position data). For example, the range of the first-level decision threshold is set to [0.8, 1.0], the range of the second-level decision threshold is [0.5, 0.8], and the range of the third-level decision threshold is [0, 0.5]. When the task-environment correlation is within the first-level decision threshold range, it is fuzzified to the "high correlation" level; when the task-environment correlation is within the second-level decision threshold range, it is fuzzified to the "medium correlation" level; when the task-environment correlation is within the third-level decision threshold range, it is fuzzified to the "low correlation" level. The fuzzified low-altitude multi-objective decision features are integrated and normalized to generate a multi-dimensional fuzzy input vector, which includes the fuzzified and normalized values ​​of the safety constraint features, the fuzzified and normalized values ​​of the efficiency optimization features, and the fuzzified and normalized values ​​of the task semantic features.

[0095] It should also be noted that the three-level decision threshold is defined based on the task execution effect of low-altitude multi-target decision features. Specifically, historical observations of task priority, semantic relevance, and cosine similarity of position difference in historical data are divided into several feature value intervals (e.g., [0, 0.5), [0.5, 0.8), [0.8, 1.0]). Each feature value interval represents the value range of a feature in low-altitude multi-target decision features. The task execution effect of all low-altitude multi-target decision features within each feature value interval is statistically analyzed, including task completion rate, response delay, and failure rate. The relationship between each feature value interval and task execution effect is analyzed (e.g., whether a high feature value interval corresponds to a high completion rate). Based on the analysis results, the levels are divided as follows: feature value intervals with task completion rates higher than the average level are defined as "high relevance" level; feature value intervals with task completion rates at the average level are defined as "medium relevance" level; and feature value intervals with task completion rates lower than the average level are defined as "low relevance" level. The level division results are integrated to form the three-level decision threshold.

[0096] S3.4 Perform function mapping on the multidimensional fuzzy input vector, calculate the anti-disturbance correction amount, and encapsulate the anti-disturbance correction amount in a structured manner to generate an anti-disturbance control instruction set;

[0097] The expression for calculating the disturbance rejection correction is:

[0098] ;

[0099] in, This indicates the interference immunity correction amount. Represents safety constraint characteristics The dynamic weighting coefficients, Represents safety constraint characteristics The normalized value after fuzzification Representing efficiency optimization features The dynamic weighting coefficients, Representing efficiency optimization features The normalized value after fuzzification Representing the semantic features of the task The dynamic weighting coefficients, Representing the semantic features of the task The normalized value after fuzzification;

[0100] It should be noted that the dynamic weighting coefficients of the safety constraint features are defined based on the normalized values ​​of the turbulence spectrum eigenvectors, with an exemplary range of 0.3 to 0.8.

[0101] The dynamic weight coefficients of the efficiency optimization features are defined based on the normalized values ​​of the dynamic scene adaptability features, with an exemplary value range of 0.1 to 0.4.

[0102] The dynamic weight coefficients of the task semantic features are defined based on the normalized value of the intent encoding vector, with an exemplary range of 0.1 to 0.3.

[0103] S4. Collect fault logs and task execution data of UAVs during operation based on the anti-disturbance control instruction set, and generate an evolution strategy library through association rule mining;

[0104] S4.1 Execute the anti-disturbance control instruction set, synchronously collect the fault logs and mission execution data during UAV operation, and generate a related dataset;

[0105] It should be noted that the anti-disturbance control command set is input into the UAV for execution, and the UAV's operating status is monitored synchronously. Various UAV operating status data are collected in real time during the execution of anti-disturbance control, including flight attitude, power parameters, navigation and positioning data, raw sensor data, and task execution data including efficiency indicators such as task target achievement rate, path planning deviation, and anti-disturbance control response delay. At the same time, fault logs generated by the equipment self-test are recorded, covering fault information such as hardware abnormal alarms, software abnormal prompts, and environmental abnormal responses. The anti-disturbance control command set, UAV operating status data, and fault logs are correlated and matched by timestamp alignment to generate a correlated dataset containing three-dimensional information of "control command - execution status - fault event".

[0106] S4.2 Extract the fault propagation patterns and task efficiency features of the associated datasets and generate a fault-efficiency association feature table;

[0107] It should be noted that, based on time series analysis, correlation mining is performed on UAV operational status data and fault logs to identify the propagation path and temporal pattern of faults spreading from a single sensor and actuator to related devices (such as attitude control failure, power reduction), and integrate them to form fault propagation patterns (such as "IMU anomaly → attitude control failure → path deviation"). At the same time, the task efficiency change rate before and after the fault occurs is obtained through the sliding window statistical method, and the efficiency indicators of task execution data and the task efficiency change rate are integrated to generate task efficiency features. The fault propagation patterns and corresponding task efficiency features are structurally correlated according to fault type and task stage to generate a fault-efficiency correlation feature table.

[0108] S4.3. Perform association rule mining on the fault-efficiency association feature table to generate an initial policy library, and optimize the policy weights based on the association dataset to generate an evolved policy library.

[0109] It should be noted that the Apriori algorithm is used to mine high-frequency co-occurrence association rules (such as "when the IMU is abnormal and the turbulence intensity is >0.7, there is a 90% probability that the efficiency will decrease by more than 15%) based on the fault propagation mode, task efficiency characteristics, and environmental conditions and real-time status of the UAV in the associated dataset (such as remaining battery power and flight altitude). An initial policy library containing preconditions, association results and confidence scores is generated. Based on the historical execution records in the associated dataset, the effectiveness weight of each policy in the initial policy library in different scenarios is obtained, and the policy weights are dynamically adjusted in combination with online feedback data (such as real-time task efficiency correction). An evolved policy library is then generated.

[0110] S5. Perform causal analysis based on the evolution strategy library, generate cluster collaboration rule upgrade packages, dynamically verify the cluster collaboration rule upgrade packages, and output self-healing strategies.

[0111] S5.1 Extract historical evolution strategies from the evolution strategy library, analyze the causal relationship between historical evolution strategies and cluster collaboration effects using causal inference algorithms, and output a causal analysis report;

[0112] It should be noted that historical evolution strategies in the evolution strategy library are used as processing variables, and core quantitative indicators of UAV swarm collaboration (such as task completion rate, collaboration response time, and resource allocation balance) are selected as outcome variables. At the same time, environmental interference factors and confounding variables such as the basic state of the swarm (such as the number of UAVs and initial distribution) are controlled. By comparing the collaboration performance of swarms that have applied evolution strategies with those that have not applied evolution strategies in the same scenario, the causal effect value of historical evolution strategies on swarm collaboration is obtained. Furthermore, by identifying which intermediate variables (such as reduced communication latency and optimized task allocation) indirectly affect the swarm collaboration effect through which evolution strategies, the action path and causal relationship of evolution strategies are analyzed in depth. A causal analysis report is then output.

[0113] S5.2 Perform causal inference and logical reconstruction on the causal analysis report to generate a cluster collaboration rule upgrade package;

[0114] It should be noted that deep causal inference is performed on the causal relationships revealed in the causal analysis report. The causal effects and action paths are transformed into logical rules in the form of "IF-THEN" through a logical rule extraction algorithm (such as "when the environmental turbulence intensity is >0.7 and the policy triggering condition is met, the attitude compensation action can shorten the collaborative response time by 20%). Redundant and contradictory logical rules are optimized and integrated by combining priority sorting; and a cluster collaborative rule upgrade package is generated.

[0115] S5.3. The cluster collaboration effect of the cluster collaboration rule upgrade package is verified by simulation, and dynamic calibration is performed according to the preset index threshold to output the self-healing strategy.

[0116] It should be noted that by constructing a virtual simulation environment highly matched to the actual operating environment, the cluster collaboration rule upgrade package is transformed into specific behavioral guidelines for individual drones (such as environmental condition-triggered actions and task priority response logic). Multiple typical scenarios covering various challenging conditions, including complex terrain, dense obstacles, and emergency needs, are set up, and the initial state and collaboration objectives of the drone cluster are deployed according to historical operational logic. During the virtual simulation, the flight status of each drone and the overall cluster collaboration effect are tracked in real time. Simulated values ​​of collaboration indicators such as task completion rate, collaboration response time, and conflict avoidance success rate are extracted. The simulated values ​​of these collaboration indicators are compared one by one with preset indicator thresholds to determine the effectiveness of the cluster collaboration rule upgrade package. If the simulated value of any collaboration indicator in the overall cluster collaboration effect does not reach the preset indicator threshold, then... This study analyzes the behavioral trajectory and interaction process of UAVs when making decisions based on the swarm collaboration rule upgrade package, identifies defects in the UAV's action rule logic and related rule parameter settings (such as unreasonable triggering conditions and action response deviations). It dynamically adjusts rule parameters to address these defects (such as correcting sensitivity and optimizing action execution logic), and re-runs simulations to verify the adjustment effects. Rule parameters whose simulated values ​​of collaboration indicators reach preset thresholds are retained, and the optimal parameter combination is gradually approximated through multiple iterations. A self-healing strategy is output, which includes optimized rule parameters, responses to different abnormal situations, and logic for self-adjustment of rule parameters. By dynamically optimizing the relevant rule parameters and response logic of the swarm collaboration rule upgrade package, the self-healing strategy ensures that the UAV autonomously adapts to complex environments and continuously improves its collaborative efficiency.

[0117] It should also be noted that when constructing a virtual simulation environment that closely matches the actual operating environment, based on historical operating data (such as terrain, meteorological conditions, and electromagnetic characteristics), the basic parameters of the environment and the rules for dynamic changes in the environment (such as meteorological evolution and fluctuations in lighting conditions) are defined; according to the actual operating range and task type of the UAV swarm, the spatial grid of the simulation area is divided and key elements (such as no-fly zones, task target points, and resource distribution points) are marked, and environmental interaction logic is configured, including the influence of airflow on flight and collision judgment rules for obstacles; through parameter calibration, the physical characteristics and dynamic events in the environment are made consistent with the actual operating environment, forming a reproducible and realistic simulation base scenario;

[0118] The preset indicator thresholds are defined based on the optimal and standard performance of the drone swarm in a normal task scenario in historical operation data. For example, the preset indicator threshold range can be set as: task completion rate (0.9, 1), collaborative response time (0, 2 seconds), and conflict avoidance success rate (0.95, 1).

[0119] This embodiment also provides a low-altitude economic unmanned aerial vehicle (UAV) data processing system based on a large model, including: a data acquisition module, a semantic association module, a fuzzy reasoning module, a strategy mining module, and a causal analysis module;

[0120] The data acquisition module is used to collect low-altitude dynamic sensing data and preprocess it, and then perform structured transformation on the preprocessed low-altitude dynamic sensing data to generate environmental fingerprints.

[0121] The semantic association module is used to input environmental fingerprints into the semantic association big model, the dynamic topology perception layer performs task-space fusion evolution, the multimodal intent decoupling layer performs hierarchical optimization, and generates a low-altitude semantic decision map.

[0122] The fuzzy inference module is used to extract the low-altitude multi-target decision features from the low-altitude semantic decision map, and to perform fuzzy inference on the low-altitude multi-target decision features to generate an anti-disturbance control instruction set.

[0123] The strategy mining module is used to collect fault logs and task execution data of UAVs based on the anti-disturbance control instruction set, and generate an evolution strategy library through association rule mining.

[0124] The causal analysis module is used to perform causal analysis based on the evolution strategy library, generate cluster collaboration rule upgrade packages, dynamically verify the cluster collaboration rule upgrade packages, and output self-healing strategies.

[0125] This embodiment also provides a computer device applicable to the data processing method for low-altitude economic unmanned aerial vehicles based on a large model, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the data processing method for low-altitude economic unmanned aerial vehicles based on a large model as proposed in the above embodiment.

[0126] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0127] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the data processing method for low-altitude economic UAVs based on a large model as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0128] In summary, this invention utilizes a semantic association model for task-space fusion evolution and hierarchical optimization to uncover the intrinsic relationships between data, thereby improving the accuracy and adaptability of UAVs in making decisions in complex scenarios. Simultaneously, by generating an anti-disturbance control instruction set for dynamic optimization and adaptive evolution, it enhances the reliability, safety, and success rate of UAVs in mission execution.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A data processing method for low-altitude economic unmanned aerial vehicles based on a large model, characterized in that: include, Collect low-altitude dynamic sensing data and preprocess it; perform structured transformation on the preprocessed low-altitude dynamic sensing data to generate environmental fingerprints. The environmental fingerprint is input into the semantic association big model, the dynamic topology perception layer performs task-space fusion evolution, the multimodal intent decoupling layer performs hierarchical optimization, and a low-altitude semantic decision map is generated. Extract low-altitude multi-target decision features from the low-altitude semantic decision map, and perform fuzzy reasoning on the low-altitude multi-target decision features to generate an anti-disturbance control instruction set; Based on the anti-disturbance control instruction set, fault logs and task execution data of UAVs during operation are collected, and an evolution strategy library is generated through association rule mining. Based on the evolution strategy library, causal analysis is performed to generate a cluster collaboration rule upgrade package. The cluster collaboration rule upgrade package is dynamically verified, and a self-healing strategy is output.

2. The data processing method for low-altitude economic unmanned aerial vehicles based on a large model as described in claim 1, characterized in that: The low-altitude dynamic sensing data includes visible light image data, lidar point cloud data, IMU attitude data, GPS location data, turbulence frequency domain signal data, and mission command text data. The preprocessing includes data cleaning, format conversion, deduplication, normalization, and outlier handling.

3. The data processing method for low-altitude economic unmanned aerial vehicles based on a large model as described in claim 2, characterized in that: The specific steps for generating the environmental fingerprint are as follows: Dynamic scene adaptive features are extracted from visible light image data and lidar point cloud data. The dynamic scene adaptive features are then aligned with coordinates to generate a motion-invariant feature matrix. Time synchronization and coordinate transformation are performed on IMU attitude data and GPS location data. By obtaining the spatial relationship features of IMU attitude data and GPS location data, a cooperative topological adjacency matrix is ​​constructed. Frequency domain analysis is performed on turbulent frequency domain signal data to obtain turbulence spectrum feature vectors; Semantically compress the task instruction text data to obtain the intent encoding vector; By integrating motion-invariant feature matrices, cooperative topological adjacency matrices, intent encoding vectors, and turbulence spectrum feature vectors, environmental fingerprints are obtained.

4. The low-altitude economic UAV data processing method based on a large model as described in claim 3, characterized in that: The specific construction steps of the semantic association model are as follows. Invoke and initialize the fully connected network and the gated linear network to build a dynamic topology perception layer and a multimodal intent decoupling layer; By using the cross-entropy loss function to perform backpropagation and gradient update on the dynamic topology perception layer and the multimodal intent decoupling layer, a large semantic association model is constructed.

5. The data processing method for low-altitude economic unmanned aerial vehicles based on a large model as described in claim 4, characterized in that: The specific steps for generating the low-altitude semantic decision map are as follows: The environmental fingerprint is input into the semantic association model, and the dynamic topology perception layer performs task-space fusion evolution through dynamic graph convolution to generate a dynamic collaborative constraint graph. The multimodal intent decoupling layer is optimized through a gating attention mechanism to form multi-dimensional balanced decision features; The dynamic collaborative constraint graph and multi-dimensional balanced decision features are integrated through the feature splicing channel to generate a low-altitude semantic decision map.

6. The low-altitude economic UAV data processing method based on a large model as described in claim 5, characterized in that: The specific steps for extracting low-altitude multi-target decision features from the low-altitude semantic decision map are as follows: Multi-dimensional feature decoupling is performed on the low-altitude semantic decision map to extract safety constraint features and efficiency optimization features; Perform task semantic association on the low-altitude semantic decision map and extract task semantic features; By integrating safety constraint features, efficiency optimization features, and task semantic features, low-altitude multi-objective decision features are generated.

7. The low-altitude economic UAV data processing method based on a large model as described in claim 6, characterized in that: The specific steps for generating the disturbance rejection control instruction set are as follows: The low-altitude multi-target decision-making features are fuzzified to generate a multi-dimensional fuzzy input vector. A membership degree weighting function is applied to the multidimensional fuzzy input vector to calculate the disturbance rejection correction amount. The disturbance rejection correction amount is then encapsulated in a structured manner to generate a disturbance rejection control instruction set.

8. The low-altitude economic UAV data processing method based on a large model as described in claim 7, characterized in that: The specific steps for generating the evolution strategy library are as follows: The UAV executes an anti-disturbance control command set, synchronously collects fault logs and mission execution data during UAV operation, and generates a related dataset. Extract the fault propagation patterns and task efficiency features from the associated datasets to generate a fault-efficiency correlation feature table; We perform association rule mining on the fault-efficiency association feature table to generate an initial policy library, and optimize the policy weights based on the association dataset to generate an evolved policy library.

9. The low-altitude economic UAV data processing method based on a large model as described in claim 8, characterized in that: The specific steps of the self-healing output strategy are as follows: Extract historical evolution strategies from the evolution strategy library, analyze the causal relationship between historical evolution strategies and cluster collaboration effects using causal inference algorithms, and output a causal analysis report; Perform causal inference and logical reconstruction on the causal analysis report to generate a cluster collaboration rule upgrade package; Verify the cluster collaboration effect of the cluster collaboration rule upgrade package, dynamically calibrate it according to preset indicator thresholds, and output a self-healing strategy.

10. A data processing system for low-altitude economic unmanned aerial vehicles (UAVs) based on a large model, comprising the data processing method for low-altitude economic UAVs based on a large model as described in any one of claims 1 to 9, characterized in that: It includes a data acquisition module, a semantic association module, a fuzzy reasoning module, a strategy mining module, and a causal analysis module; The data acquisition module is used to collect low-altitude dynamic sensing data and preprocess it, and then perform structured transformation on the preprocessed low-altitude dynamic sensing data to generate environmental fingerprints. The semantic association module is used to input environmental fingerprints into the semantic association big model, the dynamic topology perception layer performs task-space fusion evolution, the multimodal intent decoupling layer performs hierarchical optimization, and generates a low-altitude semantic decision map. The fuzzy inference module is used to extract the low-altitude multi-target decision features from the low-altitude semantic decision map, and to perform fuzzy inference on the low-altitude multi-target decision features to generate an anti-disturbance control instruction set. The strategy mining module is used to collect fault logs and task execution data of UAVs based on the anti-disturbance control instruction set, and generate an evolution strategy library through association rule mining. The causal analysis module is used to perform causal analysis based on the evolution strategy library, generate cluster collaboration rule upgrade packages, dynamically verify the cluster collaboration rule upgrade packages, and output self-healing strategies.

Images