Low-altitude air route resource quantitative evaluation method and system and storage medium

Abstract

The application relates to the technical field of low-altitude airspace intelligent management, in particular to a low-altitude air route resource quantitative evaluation method and system and a storage medium, which comprises the following steps: inputting standardized multi-modal data into a multi-modal deep fusion evaluation model, extracting deep features of each mode through a mode-specific encoder, realizing information fusion between modes through a cross-modal attention fusion module, and extracting space-time dependent features through a space-time feature extraction network; calculating a plurality of basic evaluation indexes of low-altitude air route resources based on the space-time dependent features, dynamically fusing the plurality of basic evaluation indexes through an evaluation output layer using a weighting mechanism, and generating a comprehensive score of low-altitude air route resources. Through the innovative mode-specific encoder and cross-modal attention fusion module, deep features of each mode are extracted, the complex interaction and dependency relationship between modes are more accurately modeled, and the evaluation is more comprehensive and accurate.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology for low-altitude airspace, specifically to a quantitative evaluation method, system, and storage medium for low-altitude airway resources. Background Technology

[0002] With the rapid development of industries such as drone logistics, urban air mobility (UAM), and general aviation, low-altitude airspace is becoming a new type of national land resource with strategic value. To ensure the safety, orderliness, and efficiency of low-altitude flight activities, the scientific planning and dynamic management of airway networks are crucial. Its core foundation lies in the accurate and real-time quantitative evaluation of airway resources.

[0003] Traditional airway resource evaluation methods mainly rely on expert experience or static rule models based on single-dimensional data (such as traffic flow and meteorological thresholds), which have obvious shortcomings: existing methods mostly focus on single dimensions such as traffic situation or meteorological conditions, and fail to integrate multi-source heterogeneous data such as airspace structure, geographical environment, and electromagnetic spectrum, resulting in evaluation results that cannot fully reflect the comprehensive status of airway resources. For example, a smooth airway may contain hidden high risks due to complex terrain or signal interference, and a single-dimensional evaluation cannot identify such complex risks. Traditional models often use fixed weights or simple linear superposition to integrate limited indicators, which cannot dynamically adapt to changes in the importance of each evaluation dimension under different scenarios (such as severe weather, peak hours, and special missions), and also cannot capture complex nonlinear relationships between data. Most methods evaluate airway resources instantaneously and in isolation, failing to fully consider the continuity and propagation effects of traffic flow and meteorological systems in time and space, resulting in evaluation results that lack foresight and are difficult to support proactive dynamic airway planning and traffic management. The evaluation process heavily relies on domain experts to set rules and thresholds, and the level of automation and intelligence is insufficient, making it difficult to cope with the characteristics of massive and rapidly changing data in low-altitude environments, and the evaluation efficiency and consistency need to be improved.

[0004] Therefore, there is an urgent need for an intelligent quantitative evaluation method for low-altitude airway resources that can deeply integrate multimodal data, dynamically model spatiotemporal correlations, and possess self-learning capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a quantitative evaluation method, system, and storage medium for low-altitude airway resources, in order to solve the technical problems in the prior art, such as single evaluation dimensions, poor model adaptability, weak spatiotemporal correlation, and excessive reliance on human experience, which leads to one-sided, static, and insufficiently intelligent evaluation results.

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

[0007] A quantitative evaluation method for low-altitude airway resources includes the following steps:

[0008] Step 1: Collect multimodal data required for low-altitude airway resource assessment, including airspace structure modal data (A), meteorological environment modal data (W), geographic information modal data (G), traffic dynamics modal data (T), and electromagnetic environment modal data (E);

[0009] Step 2: Standardize the collected modal data to form a standardized data matrix. , where m represents the modal index, m∈{A,W,G,T,E};

[0010] Step 3: Construct a multimodal deep fusion evaluation model, which includes a modality-specific encoder, a cross-modal attention fusion module, a spatiotemporal feature extraction network, and an evaluation output layer;

[0011] Step 4: Input the standardized multimodal data into the multimodal deep fusion evaluation model, extract the deep features of each modality through the modality-specific encoder, realize intermodal information fusion through the cross-modal attention fusion module, and extract spatiotemporal dependency features through the spatiotemporal feature extraction network;

[0012] Step 5: Calculate multiple basic evaluation indicators for low-altitude airway resources based on the aforementioned spatiotemporal dependence characteristics, including capacity indicators, safety indicators, efficiency indicators, economic indicators, and flexibility indicators;

[0013] Step 6: The multiple basic evaluation indicators are dynamically fused using a weighted mechanism through the evaluation output layer to generate a comprehensive score for low-altitude airway resources.

[0014] As a preferred embodiment of the present invention, in step 1, the airspace structure mode data A includes route width, altitude layer and control area boundary data;

[0015] The meteorological environmental modal data W includes wind speed. ,visibility Precipitation and turbulence intensity data;

[0016] The geographic information modal data G includes terrain height. Obstacle distribution and population density data;

[0017] The traffic dynamic modal data T includes traffic flow density. Number of conflict points and delay index data;

[0018] The electromagnetic environment mode data E includes signal strength. Interference level and spectrum occupancy data.

[0019] As a preferred embodiment of the present invention, the modality-specific encoder in step 3 includes:

[0020] The spatial structure encoder employs a combination of convolutional neural networks (CNNs) and graph convolutional networks (GCNs) to extract spatial structure features. Where N is the number of route segments, for Feature dimensions;

[0021] The meteorological environment encoder, employing a ConvLSTM network, is used to extract spatiotemporal features of the meteorological environment. Where H and W are the height and width of the spatial grid, respectively. for Feature dimensions;

[0022] The traffic dynamic encoder, employing a Transformer encoder, is used to extract dynamic traffic features. ,in, for Feature dimensions;

[0023] Both the geographic information encoder and the electromagnetic environment encoder employ convolutional neural networks to extract geographic information features. and geomagnetic environment characteristics ,in and They are respectively and The feature dimensions.

[0024] As a preferred embodiment of the present invention, in step 3, the cross-modal attention fusion module employs a cross-attention mechanism, including:

[0025] Summarize all modal features extracted by each modality-specific encoder { , , , , };

[0026] Use any one of these modal features as the reference modal feature. Calculation and residual modal characteristics Cross attention between ,in , Let i be the query matrix from modality i. Let J be the bond matrix from mode j. Let j be the value matrix from mode j. , , For learnable parameter matrix, Where is the dimension of the key vector, and M is the total number of modal features;

[0027] Based on the cross-attention calculation results, the multimodal features are fused and output to obtain the multimodal fused features. .

[0028] As a preferred embodiment of the present invention, the spatiotemporal feature extraction network in step 3 adopts a spatiotemporal convolutional network, and its calculation process is as follows:

[0029] ;

[0030] in, Graph convolution operations representing spatial dimensions, This represents the standard convolution operation in the time dimension, where Γ and Θ are learnable convolution kernel parameters. The fused features are output by the cross-modal attention fusion module. For the extracted spatiotemporal dependency features, for The feature dimension is T', where T' is the length of the time series.

[0031] As a preferred embodiment of the present invention, the method for calculating the basic evaluation index in step 5 includes:

[0032] Capacity indicators Where σ is the sigmoid activation function, and the weights are... and bias For learnable parameters, The airspace characteristics of the i-th segment of the route, The traffic dynamics characteristics of the i-th route at time t;

[0033] Safety indicators Where α and β are risk weight parameters, and conflict risk Weather risk , Let be the number of conflict points at time t on the i-th route. Let be the flow density at time t for the i-th route. Let be the turbulence intensity at time t on the i-th segment of the route. Let be the visibility at time t along the i-th flight segment;

[0034] efficiency indicators , Let be the ideal travel time for the i-th segment of the route. Let i be the actual travel time of the i-th segment of the route. Let be the delay exponent at time t for the i-th route;

[0035] Economic indicators ,in and Here are the weighting coefficients, FuelEfficiency is the fuel efficiency factor, and InfrastructureUtilization is the infrastructure utilization factor;

[0036] Elasticity index ,in, Let be the number of available alternative routes for the i-th route, and TotalSegments be the total number of routes. Let represent the congestion level of the i-th route at time t.

[0037] As a preferred embodiment of the present invention, the method for obtaining the comprehensive score of low-altitude airway resources through a weighted mechanism in step 6 includes:

[0038] ;

[0039] in, To evaluate the overall evaluation score of the i-th route output by the output layer at time t, This is the normalized value of the k-th basic evaluation index;

[0040] For dynamic attention weights;

[0041] Weight and bias These are learnable parameters;

[0042] Extract features from the spatiotemporal feature network output;

[0043] satisfy .

[0044] As a preferred embodiment of the present invention, it further includes a model training step, wherein the model training employs a multi-task loss function:

[0045] ;

[0046] Among them: regression loss , The expert-annotated score for the i-th route at time t. To evaluate the comprehensive evaluation score of the i-th route output by the output layer at time t, N is the total number of routes and T is the total number of times;

[0047] Ranking loss ,in, Let the hinge loss function be... For symbolic functions, For boundary hyperparameters, To evaluate the overall evaluation score of the j-th route output by the output layer at time t, The expert-annotated score is given for the j-th segment of the route at time t;

[0048] spatiotemporal consistency loss Where α, β, and γ are hyperparameters. The comprehensive evaluation score of the i-th route output by the output layer at time t+1 is used to evaluate the overall score.

[0049] As a preferred embodiment of the present invention, the present invention provides a quantitative evaluation system for low-altitude airway resources, applied to a quantitative evaluation method for low-altitude airway resources, the system comprising:

[0050] The data acquisition unit is used to collect multimodal data required for low-altitude airway resource assessment, including airspace structure modal data (A), meteorological environment modal data (W), geographic information modal data (G), traffic dynamics modal data (T), and electromagnetic environment modal data (E).

[0051] The data preprocessing unit is used to standardize the collected modal data to form a standardized data matrix. , where m represents the modal index, m∈{A,W,G,T,E};

[0052] A multimodal deep fusion evaluation model unit is used to construct a multimodal deep fusion evaluation model, which includes a modality-specific encoder, a cross-modal attention fusion module, a spatiotemporal feature extraction network, and an evaluation output layer.

[0053] The feature extraction and fusion unit is used to input the standardized multimodal data into the multimodal deep fusion evaluation model, extract deep features of each modality through the modality-specific encoder, realize intermodal information fusion through the cross-modal attention fusion module, and extract spatiotemporal dependent features through the spatiotemporal feature extraction network.

[0054] The basic evaluation index calculation unit is used to calculate multiple basic evaluation indicators of low-altitude airway resources based on the spatiotemporal dependence characteristics, including capacity indicators, safety indicators, efficiency indicators, economic indicators, and flexibility indicators.

[0055] The comprehensive score generation unit is used to dynamically fuse the multiple basic evaluation indicators using an attention weighting mechanism to generate a comprehensive score for low-altitude airway resources.

[0056] As a preferred embodiment of the present invention, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for quantitative evaluation of low-altitude airway resources.

[0057] Compared with the prior art, the present invention has the following advantages:

[0058] This invention integrates five modal data: airspace structure (A), meteorological environment (W), geographic information (G), traffic dynamics (T), and electromagnetic environment (E). Through an innovative modality-specific encoder and cross-modal attention fusion module, it not only extracts the deep features of each modality but also accurately models the complex interactions and dependencies between modalities (such as how severe weather amplifies navigation risks under specific terrain). This overcomes the shortcomings of traditional methods, such as single data dimensions and information silos, making the evaluation basis more comprehensive and solid.

[0059] This invention abandons the traditional approach of fixed-weight summation. Instead, it captures the spatiotemporal evolution of flight path status through a spatiotemporal feature extraction network and drives the generation of dynamic attention weights. This allows the fusion weights of the five basic indicators—capacity, safety, efficiency, economy, and resilience—to be automatically adjusted according to real-time scenarios (such as time, location, and environment). For example, in airport terminal areas, the safety weight will automatically increase; under severe weather conditions, the safety and resilience weights will dominate. This dynamic fusion mechanism greatly enhances the scenario adaptability and decision-making rationality of the evaluation system. Attached Figure Description

[0060] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0061] Figure 1 A flowchart of a method for quantitative evaluation of low-altitude airway resources based on multimodal data provided in an embodiment of the present invention;

[0062] Figure 2 This is a model structure diagram of the multimodal deep fusion evaluation model provided in the embodiments of the present invention;

[0063] Figure 3 A block diagram of a low-altitude airway resource quantitative evaluation system based on multimodal data provided in an embodiment of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] like Figure 1 As shown, this invention provides a method for quantitative evaluation of low-altitude airway resources, comprising the following steps:

[0066] Step 1: Collect multimodal data required for low-altitude airway resource assessment and standardize it.

[0067] In this embodiment, targeting the low-altitude airspace range Ω, this application embodiment first collects the multimodal data required for evaluation. The airspace is discretized into N route segments, and data at time t is collected synchronously at a fixed period (e.g., 5 minutes). The specific collection content is as follows:

[0068] Airspace structure data (A): Obtained from air traffic control databases and aeronautical charts, including static or low-frequency update features such as the design width of each airway segment, the range of available altitude layers, and the boundary coordinates of the control sector or airspace category to which it belongs.

[0069] Meteorological and environmental data (W): Access gridded forecasts and real-time observation data from meteorological departments to obtain wind speeds covering each grid point of the flight route. ),visibility( ), precipitation ( ) and turbulence intensity index ( ).

[0070] Geographic information data (G): extracted from Geographic Information Systems (GIS) and Digital Elevation Models (DEMs), including topographic elevation (G). The distribution of known obstacles (such as tall buildings, iron towers, and mountain peaks) ) and population density distribution data ( ).

[0071] Traffic dynamic data (T): Based on real-time data sources such as Unmanned Aerial Vehicle Traffic Management (UTM) systems and ADS-B, the number of aircraft on each airway segment is calculated (for calculating traffic density). ), the predicted number of short-term conflict points ( ) and average delay time (delay index) ).

[0072] Electromagnetic environment data (E): Signal strength of key communication and navigation frequency bands along the route is obtained through deployed spectrum monitoring networks or simulation data. Background noise and interference levels ) and spectrum occupancy ( ).

[0073] To prevent differences in the dimensions and numerical ranges of data from different modalities from adversely affecting subsequent model training, the original data matrix for each modality needs to be modified. Standardize the data, m∈{A,W,G,T,E}. Use meteorological modes. For example, calculate the mean of all data within the historical time window. and standard deviation Then apply the formula:

[0074] ;

[0075] After processing The dataset conforms to a distribution with a mean of 0 and a standard deviation close to 1. Other modes (A, G, T, E) undergo the same standardization process to obtain a standardized multimodal dataset. ,in For modal indexing, the total number of modalities in this embodiment is... This step is crucial, as it standardizes the data scale and lays the foundation for stable and efficient training of the neural network.

[0076] Specifically, all data is obtained through legally authorized channels or is publicly available, anonymized, and compliant data. Data involving geographic information, electromagnetic environment, and other potentially national security or sensitive areas has been declassified, aggregated, or simulated to ensure that specific information regarding state secrets, military facilities, and sensitive infrastructure is not disclosed. Data involving aircraft and population has been anonymized or aggregated to protect personal privacy and trade secrets. This evaluation model is deployed in a secure and controllable environment, and the evaluation results are only used to assist in decision-making. The final route planning and management decisions still need to be reviewed by authorized personnel based on the actual situation to ensure ultimate human control over key security decisions. The fundamental purpose of this invention is to improve the safety, efficiency, and resource utilization of low-altitude airspace, serving public interest areas such as drone logistics, emergency rescue, and urban transportation. There is no design or application that violates social morality or harms public interests.

[0077] Step 2: Construct a multimodal deep fusion evaluation model, which includes a modality-specific encoder, a cross-modal attention fusion module, a spatiotemporal feature extraction network, and an evaluation output layer.

[0078] like Figure 2As shown in this embodiment, this application requires the construction of an end-to-end multimodal deep fusion evaluation model. The core of this model lies in its hierarchical structure design, including a modality-specific encoder, a cross-modal attention fusion module, a spatiotemporal feature extraction network, and an evaluation output layer.

[0079] Specifically, the construction and initialization of modality-specific encoders involve designing dedicated encoders to extract deep features based on the data structure characteristics of different modalities.

[0080] Airspace Structure Encoder: Due to the graph structure characteristics of the route network, this embodiment employs a combination of Convolutional Neural Network (CNN) and Graph Convolutional Network (GCN). First, a two-layer CNN is used to initially extract local features of the route. Then, a two-layer GCN is connected to model the topological connectivity relationships (such as adjacency relationships) between route segments. Finally, the airspace structure features are output. ,in Number of segments for the route .

[0081] Specifically, the CNN part uses two 2D convolutional layers. The first layer has the following input channels: [spatial feature dimension] (the original number of features in the spatial structure data; for example, if the data includes three features such as flight path width and altitude, the original number of features = 3), 64 output channels, a 3x3 kernel, a stride of 1, padding of 1, and the ReLU activation function. The second layer has the following input channels: 64, 128 output channels, a 3x3 kernel, a stride of 1, padding of 1, and the ReLU activation function.

[0082] The GCN component employs a two-layer graph convolutional layer. The graph structure is built based on the adjacency relationships of flight path segments. The first layer has an input feature dimension of 128 and an output dimension of 256. The second layer has an input dimension of 256 and an output dimension of d. A =128. Each layer is followed by a ReLU activation function and layer normalization (LayerNorm). This invention chooses GCN because the route network is essentially a graph structure (nodes are route segments, edges are connections), and GCN can effectively model the topological dependencies between route segments (such as capacity constraints of upstream and downstream segments), which is more consistent with the physical characteristics of airspace data than conventional CNN.

[0083] Meteorological Environment Encoder: Meteorological data exhibits strong correlation across time and space. This embodiment employs a ConvLSTM network. A two-layer ConvLSTM (with a 3x3 kernel) simultaneously captures the spatial distribution patterns of meteorological elements and their temporal evolution. The input tensor shape is [batch size, time step, number of channels, height H, width W], and the output is the spatiotemporal features of the meteorological environment. ,in For spatial grid size, .

[0084] Traffic Dynamic Encoder: To effectively capture the complex temporal dependencies (such as periodicity and trends) of traffic flow, this embodiment employs a Transformer encoder. A 4-layer Transformer encoder is configured, with each layer containing a multi-head self-attention mechanism (8 attention heads) and a feedforward neural network (FFN, with intermediate layer dimensions of 512). The input feature dimensions are projected onto... Location coding employs sine and cosine coding, utilizing its self-attention mechanism to extract dynamic traffic features. .

[0085] Geographic Information and Electromagnetic Environment Encoders: These two types of data have relatively stable spatial structures. In this embodiment, a simple two-layer convolutional neural network (CNN, similar in structure to the CNN part of the spatial domain encoder) is used for feature extraction, and the outputs are respectively... .

[0086] Specifically, the implementation of the cross-modal attention fusion module is a key module for achieving deep interaction and complementarity of multi-source information.

[0087] First, summarize all the modal features output by the encoders mentioned above. This embodiment selects spatial structure features. As a benchmark feature Because it represents the most essential and stable physical attribute of a shipping route.

[0088] The cross-modal attention fusion module employs a cross-attention mechanism, including:

[0089] Summarize all modal features extracted by each modality-specific encoder { , , , , };

[0090] Use any one of these modal features as the reference modal feature. Calculation and residual modal characteristics Cross attention between ,in , Let i be the query matrix from modality i. Let J be the bond matrix from mode j. Let j be the value matrix from mode j. , , For learnable parameter matrix, Where is the dimension of the key vector, and M is the total number of modal features;

[0091] Based on the cross-attention calculation results, the multimodal features are fused and output to obtain the multimodal fused features. .

[0092] Subsequently, calculation With each other modal feature Cross-attention between them. In conjunction with meteorological characteristics. Taking the fusion as an example, its query matrix Key matrix Value matrix , in This is a learnable parameter matrix. The cross-attention output is calculated as follows:

[0093] ;

[0094] in This represents the dimension of the key vector. This calculation, from the perspective of "spatial structure," assigns a new importance weight representation to meteorological information. The outputs of all four cross-attention methods are averaged and then compared with the original baseline features. The features are added together and then the fused features are output through a layer normalization (LayerNorm) operation. This design cleverly uses the core mode as an anchor point to actively aggregate enhanced information from other environmental modes, forming a unified and information-rich comprehensive route state representation.

[0095] The spatiotemporal feature extraction network employs a spatiotemporal convolutional network, and its computation process is as follows:

[0096] ;

[0097] in, Graph convolution operations representing spatial dimensions, This represents the standard convolution operation in the time dimension, where Γ and Θ are learnable convolution kernel parameters. The fused features are output by the cross-modal attention fusion module. For the extracted spatiotemporal dependency features, for The feature dimension is T', where T' is the length of the time series.

[0098] Specifically, the construction of spatiotemporal feature extraction networks, Although it incorporates information from multiple sources, its spatiotemporal characteristics are still intertwined.

[0099] This embodiment employs two stacked spatiotemporal convolution (ST-Conv) modules. Within each module, a spatial graph convolution operation is first performed ( The topology of the airway network is used to capture the spatial relationships between airway segments (such as the influence of upstream and downstream segments); then a one-dimensional convolution operation in the time dimension is performed. This smooths out temporal fluctuations and extracts underlying trends. The final output contains features with strong spatiotemporal dependencies. ,in , This represents the shortened time series length after the convolution operation. Specifically, the spatial graph convolution operation ( First-order graph convolution using Chebyshev polynomial approximation is defined on the route adjacency graph; temporal convolution ( One-dimensional causal convolution is used with a kernel size of 3 to capture temporal dependencies, and each convolution operation is followed by an activation function (such as ReLU) and a normalization operation (such as BatchNorm).

[0100] Before inputting the spatiotemporal convolutional network, the fused features Z need to be... fusion The shape (N × 256) is combined with the time dimension information and reshaped into a tensor of [batch size, N, 256, T], where T is the length of the current time window.

[0101] Step 3: Calculate basic evaluation indicators and comprehensive scores based on spatiotemporal dependence characteristics.

[0102] Specifically, the basic evaluation indicators are calculated, and the evaluation output layer receives the results. China targets specific air routes ,time eigenvectors .

[0103] This layer computes five basic evaluation metrics through parallel fully connected branches:

[0104] Capacity indicators This reflects the carrying capacity of the airway. The calculation method is as follows: ,in For the sigmoid function, and These are learnable parameters.

[0105] Safety indicators Assess route safety risks. The calculation method is as follows: Among them, the risk of conflict Weather risk , For risk weight parameters.

[0106] efficiency indicators : Measures navigation efficiency. The calculation method is as follows: .

[0107] Economic indicators : Assess operational economics. Calculation method is as follows: , These are weighting coefficients. The fuel efficiency factor reflects the operational economy of an airway by calculating the average energy consumption per unit distance; the infrastructure utilization factor is characterized by statistically analyzing the coverage effectiveness of communication and navigation stations along the airway.

[0108] Elasticity index This reflects the route's ability to cope with disturbances. The calculation method is as follows: .

[0109] Specifically, the dynamic comprehensive score is generated not as a fixed weighted average of the basic indicators, but through a dynamic attention mechanism. First, the values ​​of the five basic indicators are normalized to obtain... .

[0110] Meanwhile, the evaluation output layer is passed through another fully connected layer and a softmax function, based on the current features. Dynamically generate attention weights for five metrics :

[0111] ;

[0112] in and These are learnable parameters. The final comprehensive route resource score is:

[0113] ;

[0114] This mechanism enables the model to automatically adjust the relative importance of each evaluation dimension based on real-time scenarios (such as severe weather or peak traffic), greatly enhancing the system's adaptability and decision-making rationality.

[0115] Step 4: Model training and optimization. The excellent performance of the model depends on sufficient supervised training.

[0116] Specifically, this embodiment collects historical multimodal data covering different seasons, weather conditions, and traffic patterns. Domain experts then perform posterior scoring based on historical operational safety records and efficiency analysis to form a training set. .

[0117] Specifically, when constructing the training dataset, historical multimodal data samples were collected covering spring, summer, autumn, and winter, including various weather conditions such as sunny, rainy, foggy, and thunderstorms, as well as various traffic modes such as peak, off-peak, and nighttime. At least three domain experts with more than five years of experience in low-altitude traffic management were invited to independently score each airway segment under each historical scenario using unified scoring criteria (such as safety as a veto factor and balancing efficiency and economy). The average score was then used as the labeled value for that sample. To reduce subjective bias.

[0118] This method employs a multi-task loss function for end-to-end training, with a total loss... The definition is as follows:

[0119] ;

[0120] Regression loss ( ): Ensure that the predicted score is numerically close to the expert score.

[0121] ;

[0122] The expert-annotated score for the i-th route at time t. To evaluate the comprehensive evaluation score of the i-th route output by the output layer at time t, N is the total number of routes and T is the total number of times.

[0123] Ranking loss ( ): Ensure that the model's relative ranking of different route qualities is consistent with expert perception. Randomly sampled route segments... ,like Then it is mandatory , , , , ;

[0124] ;

[0125] in, Let the hinge loss function be... For symbolic functions, For boundary hyperparameters, To evaluate the overall evaluation score of the j-th route output by the output layer at time t, The expert-annotated score is given for the j-th segment of the route at time t;

[0126] Spatiotemporal consistency loss ( ): This penalizes unreasonable and drastic jumps in scores between adjacent time points, making the score output smoother and more in line with the continuous changes in the physical world.

[0127] ;

[0128] Where α, β, and γ are hyperparameters. The comprehensive evaluation score of the i-th route output by the output layer at time t+1 is used to evaluate the overall score.

[0129] Among them, the ranking loss weight α is used to control the model's emphasis on the relative ranking consistency between routes; the consistency loss weight β is used to smooth the time series output and prevent drastic fluctuations in scores; and the boundary γ is used to define the minimum acceptable gap between ranking predictions.

[0130] Using the AdamW optimizer, with an initial learning rate of 3e-4, a batch size of 32, and hyperparameters... With γ=0.2, training was performed on a server equipped with a GPU until the validation set loss converged.

[0131] Specifically, the detailed steps and parameters for model training.

[0132] 1. Data Preparation: The proportions of the training, validation, and test sets (e.g., 70% / 15% / 15%). Data Augmentation Strategies: These strategies include applying small, random time shifts to the time-series data and adding Gaussian noise to the meteorological grid data.

[0133] 2. Parameter initialization: Weight initialization method (such as Xavier initialization).

[0134] 3. Loss Function: Specify the values ​​of hyperparameters α, β, and γ in the multi-task loss (e.g., α=0.1, β=0.05, γ=0.2), and explain the basis or range for their selection.

[0135] 4. Optimizer: The AdamW optimizer is used, with a base learning rate of lr=3e-4 and weight decay of weight_decay=1e-4.

[0136] 5. Learning rate scheduling: Cosine AnnealingLR is used, and the learning rate is restarted every 50 training epochs.

[0137] 6. Regularization: Use Dropout in the fully connected layer and set the dropout rate to 0.2.

[0138] 7. Training process: The batch size is 32, and the total number of training rounds is 200. Each round is evaluated on the validation set, and an early stopping strategy is adopted. Training stops when the validation set loss no longer decreases for 10 consecutive rounds.

[0139] Step 5: System application and effect verification.

[0140] Specifically, the trained model is deployed in the low-altitude airway resource quantitative evaluation system. The system executes steps S11 to S13 in real time, obtaining the full network... .

[0141] This embodiment uses a week's worth of operational data from a logistics air route network in a suburban city as the test set. A threshold is set, and when a certain air segment scores... When a low-score status persists for more than 15 minutes, the system automatically marks it as a bottleneck segment; if the low score continues for more than 15 minutes, an alert is triggered, and a pool of high-quality alternative routes is automatically selected. The recommended approach is to redesign the plan.

[0142] To verify the effectiveness, the evaluation results generated by this method were compared with those of an expert system based on static rules. In three potential risk scenarios caused by sudden weather events in the test set, this method issued early warnings of score decline for the affected flight segments in advance (an average of 12 minutes in advance), while the expert system failed to respond in time twice. Furthermore, the dynamic route scheduling based on the scores provided by this method reduced the average delay index across the entire network by approximately 18% during the test period. This demonstrates the significant advantages of the proposed method in terms of comprehensiveness, real-time performance, and effectiveness in decision support.

[0143] Specifically, baseline method A (traditional static weighted method) uses fixed weights (such as safety 0.4, efficiency 0.3, capacity 0.2, economy 0.1) to linearly weight and sum the indicators of each dimension after rule processing to generate a comprehensive score.

[0144] Baseline Method B (Single-Modal CNN-LSTM Method): Using only flow and meteorological data, spatial features are extracted through CNN, temporal features are extracted through LSTM, and finally the score is output through a fully connected layer.

[0145] The experimental data is based on 30 consecutive days of historical operational data from a low-altitude logistics demonstration zone in a certain city, including 5 main air routes, a total of N=150 air route segments, and a time granularity T of 5 minutes. Data from 7 days was selected as the test set.

[0146] Table 1 shows the mean absolute error (MAE) and root mean square error (RMSE) of the combined score predicted by the three methods on the test set compared to the expert posterior scores. The expert scores were obtained by averaging the scores given independently by three senior controllers based on the operation logs.

[0147]

[0148] As shown in Table 1, the MAE and RMSE of our proposed method are significantly lower than those of the two baseline methods. This indicates that, through multimodal deep fusion and dynamic weighting mechanisms, the comprehensive score generated by our method is closer to expert judgment, resulting in the highest evaluation accuracy. Baseline method B, due to its use of some dynamic data, outperforms the completely static baseline method A, but still falls short of our proposed method due to insufficient information dimensions.

[0149] Furthermore, a sudden thunderstorm lasting 2 hours was selected from the test set to examine the early warning capabilities of the three methods for flight segments severely affected by the weather (determined as "high-risk bottlenecks" by experts post-hoc assessment). The early warning threshold was set at a comprehensive score below 0.4. Performance indicators included early warning lead time (the time difference between the score first falling below the threshold and the start of expert-determined bottleneck assessment) and false negative rate (the percentage of actual bottlenecks that failed to receive an early warning), as shown in Table 2.

[0150]

[0151] Therefore, as shown in Table 2, under this method, the system issues warnings for impending bottleneck segments approximately 26 minutes in advance on average, with a false negative rate of only 5%. This is attributed to the deep encoding of meteorological modes and their interactive modeling with airspace and terrain modes (such as capturing the combined risks of "turbulence + complex terrain" through cross-attention), giving it a stronger ability to perceive and predict risk evolution. Baseline method B mainly relies on the meteorological data itself and has limited warning capabilities; baseline method A, on the other hand, is completely unable to dynamically respond to weather changes and performs the worst.

[0152] As can be seen from the above, the embodiments of this application first systematically collect and fuse five-dimensional data including airspace, meteorology, geography, transportation, and electromagnetic data, and achieve deep information complementarity through a customized encoder and cross-attention mechanism; secondly, dynamic evolution features are extracted using a spatiotemporal convolutional network, and this drives the adaptive adjustment of evaluation index weights to generate accurate comprehensive scores; finally, end-to-end optimization is performed through multi-task loss including regression, ranking, and consistency constraints. This method constructs a complete closed loop from multi-source perception to intelligent evaluation, providing a reliable technical tool for the refined and intelligent management of low-altitude airway resources, and effectively solving the problems of single evaluation dimensions, rigid models, and poor adaptability of traditional methods.

[0153] like Figure 3 As shown, this invention provides a quantitative evaluation system for low-altitude airway resources, applied to a quantitative evaluation method for low-altitude airway resources. The system includes:

[0154] The data acquisition unit is used to collect multimodal data required for low-altitude airway resource assessment, including airspace structure modal data (A), meteorological environment modal data (W), geographic information modal data (G), traffic dynamics modal data (T), and electromagnetic environment modal data (E).

[0155] The data preprocessing unit is used to standardize the collected modal data to form a standardized data matrix. Where m represents the modality index, m∈{1,2,...,M}, and M is the total number of modalities;

[0156] The multimodal deep fusion evaluation model unit is used to construct a multimodal deep fusion evaluation model. The model includes a modality-specific encoder, a cross-modal attention fusion module, a spatiotemporal feature extraction network, and an evaluation output layer.

[0157] The feature extraction and fusion unit is used to input standardized multimodal data into the multimodal deep fusion evaluation model, extract deep features of each modality through a modality-specific encoder, achieve intermodal information fusion through a cross-modal attention fusion module, and extract spatiotemporal dependent features through a spatiotemporal feature extraction network.

[0158] The basic evaluation index calculation unit is used to calculate multiple basic evaluation indicators of low-altitude airway resources based on spatiotemporal dependence characteristics, including capacity indicators, safety indicators, efficiency indicators, economic indicators, and flexibility indicators.

[0159] The comprehensive score generation unit is used to dynamically integrate multiple basic evaluation indicators using an attention-weighted mechanism to generate a comprehensive score for low-altitude airway resources.

[0160] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for quantitative evaluation of low-altitude airway resources.

[0161] This invention is the first to systematically integrate five modal data: airspace structure (A), meteorological environment (W), geographic information (G), traffic dynamics (T), and electromagnetic environment (E). Through an innovative modality-specific encoder and cross-modal attention fusion module, it not only extracts the deep features of each modality, but also accurately models the complex interactions and dependencies between modalities (such as how severe weather amplifies navigation risks under specific terrain). This overcomes the shortcomings of traditional methods, such as single data dimensions and information silos, making the evaluation basis more comprehensive and solid.

[0162] This invention abandons the traditional approach of fixed-weight summation. Instead, it captures the spatiotemporal evolution of flight path status through a spatiotemporal feature extraction network and drives the generation of dynamic attention weights. This allows the fusion weights of the five basic indicators—capacity, safety, efficiency, economy, and resilience—to be automatically adjusted according to real-time scenarios (such as time, location, and environment). For example, in airport terminal areas, the safety weight will automatically increase; under severe weather conditions, the safety and resilience weights will dominate. This dynamic fusion mechanism greatly enhances the scenario adaptability and decision-making rationality of the evaluation system.

[0163] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A quantitative evaluation method for low-altitude airway resources, characterized in that, Includes the following steps: Step 1: Collect multimodal data required for low-altitude airway resource assessment, including airspace structure modal data A, meteorological environment modal data W, geographic information modal data G, traffic dynamics modal data T, and electromagnetic environment modal data E; Step 2: Standardize the collected modal data to form a standardized data matrix. , where m represents the modal index, m∈{A,W,G,T,E}; Step 3: Construct a multimodal deep fusion evaluation model, which includes a modality-specific encoder, a cross-modal attention fusion module, a spatiotemporal feature extraction network, and an evaluation output layer; Step 4: Input the standardized multimodal data into the multimodal deep fusion evaluation model, extract the deep features of each modality through the modality-specific encoder, realize intermodal information fusion through the cross-modal attention fusion module, and extract spatiotemporal dependency features through the spatiotemporal feature extraction network; Step 5: Calculate multiple basic evaluation indicators for low-altitude airway resources based on the aforementioned spatiotemporal dependence characteristics, including capacity indicators, safety indicators, efficiency indicators, economic indicators, and flexibility indicators; Step 6: The multiple basic evaluation indicators are dynamically fused using a weighted mechanism through the evaluation output layer to generate a comprehensive score for low-altitude airway resources.

2. The method for quantitative evaluation of low-altitude airway resources according to claim 1, characterized in that: In step 1, the airspace structure modal data A includes route width, altitude layer, and control area boundary data; The meteorological environmental modal data W includes wind speed. ,visibility Precipitation and turbulence intensity data; The geographic information modal data G includes terrain height. Obstacle distribution and population density data; The traffic dynamic modal data T includes traffic flow density. Number of conflict points and delay index data; The electromagnetic environment mode data E includes signal strength. Interference level and spectrum occupancy data.

3. The method for quantitative evaluation of low-altitude airway resources according to claim 2, characterized in that: The modality-specific encoder in step 3 includes: The spatial structure encoder employs a combination of convolutional neural networks (CNNs) and graph convolutional networks (GCNs) to extract spatial structure features. Where N is the number of route segments, for Feature dimensions; The meteorological environment encoder, employing a ConvLSTM network, is used to extract spatiotemporal features of the meteorological environment. Where H and W are the height and width of the spatial grid, respectively. for Feature dimensions; The traffic dynamic encoder, employing a Transformer encoder, is used to extract dynamic traffic features. ,in, for Feature dimensions; Both the geographic information encoder and the electromagnetic environment encoder employ convolutional neural networks to extract geographic information features. and geomagnetic environment characteristics ,in and They are respectively and The feature dimensions.

4. The method for quantitative evaluation of low-altitude airway resources according to claim 3, characterized in that: In step 3, the cross-modal attention fusion module employs a cross-attention mechanism, including: Summarize all modal features extracted by each modality-specific encoder { , , , , }; Use any one of these modal features as the reference modal feature. Calculation and residual modal characteristics Cross attention between ,in , Let i be the query matrix from modality i. Let J be the bond matrix from mode j. Let j be the value matrix from mode j. , , For learnable parameter matrix, Where is the dimension of the key vector, and M is the total number of modal features; Based on the cross-attention calculation results, the multimodal features are fused and output to obtain the multimodal fused features. .

5. The method for quantitative evaluation of low-altitude airway resources according to claim 4, characterized in that: The spatiotemporal feature extraction network in step 3 uses a spatiotemporal convolutional network, and its calculation process is as follows: ； in, Graph convolution operations representing spatial dimensions, This represents the standard convolution operation in the time dimension, where Γ and Θ are learnable convolution kernel parameters. The fused features are output by the cross-modal attention fusion module. For the extracted spatiotemporal dependency features, for The feature dimension is T', where T' is the length of the time series.

6. The method for quantitative evaluation of low-altitude airway resources according to claim 5, characterized in that: The calculation method for the basic evaluation indicators in step 5 includes: Capacity indicators Where σ is the sigmoid activation function, and the weights are... and bias For learnable parameters, The airspace characteristics of the i-th segment of the route, The traffic dynamics characteristics of the i-th route at time t; Safety indicators Where α and β are risk weight parameters, and conflict risk Weather risk , Let be the number of conflict points at time t on the i-th route. Let be the flow density at time t for the i-th route. Let be the turbulence intensity at time t on the i-th segment of the route. Let be the visibility at time t along the i-th flight segment; efficiency indicators , Let be the ideal travel time for the i-th segment of the route. Let i be the actual travel time of the i-th segment of the route. Let be the delay exponent at time t for the i-th route; Economic indicators ,in and Here are the weighting coefficients, FuelEfficiency is the fuel efficiency factor, and InfrastructureUtilization is the infrastructure utilization factor; Elasticity index ,in, Let be the number of available alternative routes for the i-th route, and TotalSegments be the total number of routes. Let represent the congestion level of the i-th route at time t.

7. The method for quantitative evaluation of low-altitude airway resources according to claim 6, characterized in that: The method for obtaining the comprehensive score of low-altitude airway resources through a weighted mechanism in step 6 includes: ； in, To evaluate the overall evaluation score of the i-th route output by the output layer at time t, This is the normalized value of the k-th basic evaluation index; For dynamic attention weights; Weight and bias These are learnable parameters; Extract features from the spatiotemporal feature network output; satisfy .

8. The method for quantitative evaluation of low-altitude airway resources according to claim 7, characterized in that: It also includes a model training step, in which a multi-task loss function is used: ； Among them: regression loss , The expert-annotated score for the i-th route at time t. To evaluate the comprehensive evaluation score of the i-th route output by the output layer at time t, N is the total number of routes and T is the total number of times; Ranking loss ,in, Let the hinge loss function be... For symbolic functions, For boundary hyperparameters, To evaluate the overall evaluation score of the j-th route output by the output layer at time t, The expert-annotated score is given for the j-th segment of the route at time t; spatiotemporal consistency loss Where α, β, and γ are hyperparameters. The comprehensive evaluation score of the i-th route output by the output layer at time t+1 is used to evaluate the overall score.

9. A quantitative evaluation system for low-altitude airway resources, characterized in that, The system, which is applied to the quantitative evaluation method for low-altitude airway resources according to any one of claims 1-8, comprises: The data acquisition unit is used to collect multimodal data required for low-altitude airway resource assessment, including airspace structure modal data (A), meteorological environment modal data (W), geographic information modal data (G), traffic dynamics modal data (T), and electromagnetic environment modal data (E). The data preprocessing unit is used to standardize the collected modal data to form a standardized data matrix. , where m represents the modal index, m∈{A,W,G,T,E}; A multimodal deep fusion evaluation model unit is used to construct a multimodal deep fusion evaluation model, which includes a modality-specific encoder, a cross-modal attention fusion module, a spatiotemporal feature extraction network, and an evaluation output layer. The feature extraction and fusion unit is used to input the standardized multimodal data into the multimodal deep fusion evaluation model, extract deep features of each modality through the modality-specific encoder, realize intermodal information fusion through the cross-modal attention fusion module, and extract spatiotemporal dependent features through the spatiotemporal feature extraction network. The basic evaluation index calculation unit is used to calculate multiple basic evaluation indicators of low-altitude airway resources based on the spatiotemporal dependence characteristics, including capacity indicators, safety indicators, efficiency indicators, economic indicators, and flexibility indicators. The comprehensive score generation unit is used to dynamically fuse the multiple basic evaluation indicators using an attention weighting mechanism to generate a comprehensive score for low-altitude airway resources.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images