Unmanned aerial vehicle assisted accident handling heterogeneous risk control data-driven robust site selection method

Abstract

This invention relates to intelligent transportation technology and aims to provide a data-driven robust site selection method for heterogeneous risk management in drone-assisted accident handling. The method includes: cleaning and preprocessing historical traffic incident data, extracting accident tags and performing quantitative classification based on the U-I-A multidimensional evaluation framework to obtain a structured dataset; decomposing uncertain demand into basic normal demand and multi-source heterogeneous fluctuations, and establishing a structured uncertainty set based on the quantitatively classified structured dataset; modeling the robust site selection planning model as a minimax robust optimization problem, and then equivalently reconstructing it into a solvable single-stage mixed integer linear programming model; using a solver to solve the problem, obtaining the globally optimal hangar site selection scheme and the number of supporting drones, and outputting the results. This site selection method enhances the reliability and robustness of the drone emergency response system; it enables scientific planning of drone site deployment, achieving rapid and stable emergency response.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of intelligent transportation systems, unmanned aerial vehicle (UAV) technology and applications, and operations research and optimization theory, and in particular to a data-driven robust location selection method for heterogeneous risk management in UAV-assisted accident handling. Background Technology

[0002] With the continuous advancement of urbanization, the complexity of large-scale transportation systems is increasing. Against this backdrop, the efficiency and cost of traffic accident emergency response have become key constraints on the overall operational effectiveness of urban transportation. In practice, the current urban emergency management paradigm suffers from insufficient allocation of dedicated incident response resources, leading to a queuing-based sequential service model for incident handling. Especially during peak traffic hours, ground response units are often constrained by traffic congestion, significantly limiting the timeliness and efficiency of incident handling.

[0003] Given the aforementioned limitations, in the ongoing wave of smart city development, drones, with their flexibility, efficiency, and multi-dimensional perception capabilities, have gradually evolved from an emerging technology into a routine operational tool for urban governance. Therefore, many cities have begun to incorporate low-altitude aviation applications into their traditional traffic management systems, making them an important component of emergency traffic response systems to enhance rapid reconnaissance, command and coordination, and communication support capabilities in accident investigation and rescue operations. Against this backdrop, research on the deployment optimization of drone-assisted traffic accident emergency response has received widespread attention.

[0004] However, most existing UAV hangar location models are built based on static demand points or simple service radii, failing to fully consider the spatiotemporal dynamic characteristics of traffic accident demands. In particular, they lack effective identification and control of the structural heterogeneity inherent in urban traffic risks, making them ill-suited to stringent emergency response time constraints. Furthermore, existing location methods generally separate site selection optimization from UAV mission scheduling optimization. This decoupling makes it difficult to accurately evaluate and verify location schemes under complex temporal and mission-level constraints, resulting in schemes that often lack sufficient robustness in real-world complex operating conditions.

[0005] In summary, there is an urgent need for a robust site selection and scheduling technology that combines heterogeneous risk management with data-driven methods to address the problems of static models, insufficient robustness, and separation of site selection and scheduling in existing technologies; thereby scientifically optimizing the deployment of UAV sites and comprehensively improving the efficiency and reliability of urban traffic accident emergency response. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a data-driven robust location selection method for heterogeneous risk management in drone-assisted accident handling.

[0007] To solve the technical problem, the solution of the present invention is:

[0008] A data-driven robust location selection method for managing heterogeneous risks in drone-assisted accident handling is provided, comprising the following steps:

[0009] (1) Data processing and quantitative classification: Preprocess the historical data of traffic incidents; extract accident labels from the processed data, and quantitatively classify the accident labels according to the UIA multidimensional evaluation framework of urgency, importance and ambiguity to obtain a structured dataset;

[0010] (2) Heterogeneity and statistical calibration: Decompose the uncertain demand into basic normal demand and multi-source heterogeneous fluctuations. Based on the structured dataset after quantitative classification, establish a structured uncertainty set based on hierarchical statistical features to characterize the heterogeneity and spatiotemporal distribution differences of emergency demand.

[0011] (3) Robust location planning model construction: The robust location planning model is modeled as a minimax robust optimization problem, with the goal of minimizing the total system shortage risk after weighting in all time periods under the premise of satisfying resource and operational stability constraints;

[0012] (4) Model transformation and solution results: The robust location planning model is equivalently reconstructed into a solvable single-stage mixed integer linear programming model using the strong duality principle; the solution is obtained by using a solver to obtain the globally optimal hangar location scheme and the number of supporting UAVs, and the results are output.

[0013] Furthermore, this invention provides an air-ground collaborative dynamic scheduling method for unmanned aerial vehicle (UAV)-assisted traffic accident handling. Based on the aforementioned heterogeneous risk management data-driven robust location selection method, it obtains the globally optimal hangar location scheme and the number of supporting UAVs by solving a single-stage mixed-integer linear programming model. This is further used to implement priority allocation and air-ground collaboration verification, thereby realizing the calculation of air-ground collaborative dynamic scheduling. The location selection model provides physical boundaries and resource constraints for the scheduling method. Specifically, it includes the following steps:

[0014] (1) Set the parameters required for the operation of the single-stage mixed-integer linear programming model, and then initialize it;

[0015] (2) Using the UIA multidimensional evaluation framework, all tasks to be processed are sorted in descending order based on their comprehensive scores, which are composed of urgency, importance, and ambiguity, to obtain an ordered task list; then, the single-stage mixed integer linear programming model performs the operations of steps (3)-(4):

[0016] (3) If the ordered task list is not empty, then take out the task with the highest priority and execute the sub-processes of best feasible UAV search, air-ground cooperative response constraint verification, task allocation and resource status update.

[0017] (4) When all tasks are completed or no resources are available for allocation, end the loop and output the final task allocation vector.

[0018] Compared with the prior art, the technical advantages of the present invention are:

[0019] 1. The robust location method of the present invention enhances the reliability and robustness of the UAV emergency response system.

[0020] This invention constructs a data-driven uncertainty set based on UIA tags, which divides sub-regions During the period The total accident demand is decomposed into the superposition of "basic normal demand" and "multi-source heterogeneous fluctuations". This effectively anchors the risk budget to the statistical quantile of the worst-case scenario in history, thereby ensuring that the model can effectively capture and manage long-tail risks with low probability of occurrence but high consequences.

[0021] 2. This invention enables the scientific planning of drone site deployment, achieving rapid and stable emergency response.

[0022] This invention incorporates a congestion-aware stability criterion into the constraints and constructs a minimax robust optimization model. This model reveals the economic mechanism of risk allocation through dual constraints, ensuring that the implicit resource costs allocated by the system are sufficient to cover the attack costs caused by type-specific events. This mechanism guarantees that the selected site selection scheme ensures that, even in the face of the worst heterogeneous risks and the most severe traffic congestion, the traffic intensity within the service area remains below the maximum permissible traffic intensity to prevent system saturation, thus fundamentally achieving robust assurance of system service quality.

[0023] 3. Addressing the inherent characteristics of the max-min robust optimization model—its two-layer structure, high complexity, and difficulty in direct solution—this invention constructs a solvable single-stage mixed-integer linear programming (MILP) model. Utilizing the strong duality principle, this invention equivalently transforms the robust constraints containing inner-layer maximization terms into a dual minimization problem. By substituting back into the original model, a single-layer equivalent formula is finally obtained, resulting in a single-stage mixed-integer linear programming (MILP) model that can be directly solved by existing solvers. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the data processing flow used as an example in this invention.

[0025] Figure 2 This is a schematic diagram of the UIA dictionary library of the present invention. Detailed Implementation

[0026] This invention proposes a widely applicable and robust facility site selection scheme, aiming to address issues in existing technologies such as the separation of site selection and task scheduling optimization, lack of strict emergency response time constraints, and absence of consideration for spatiotemporal dynamic characteristics. Supported by a real historical traffic accident dataset, this method provides a site selection approach that integrates multidimensional quantification and robust planning. Through the UIA multidimensional evaluation framework and structured uncertainty sets, it effectively identifies, quantifies, and embeds the long-tail risks of high-urgency accidents and the heterogeneity and spatiotemporal distribution differences of demand, overcoming the shortcomings of traditional models that ignore distribution differences.

[0027] The site selection method proposed in this invention aims to minimize the weighted worst-case service shortage across all time periods. It innovatively introduces an urgency (U)-importance (I)-ambiguity (A) evaluation framework, which, while addressing the high uncertainty of the spatial distribution of accidents, also considers and evaluates the structural heterogeneity of accidents. This adapts to the potential demand for emergency resources from high-urgency events (such as accidents involving casualties), solving the problem of underfitting models when dealing with high-consequence scenarios and filling a gap in existing knowledge. Simultaneously, this invention explicitly introduces buffer margin constraints to address micro-queue fluctuations and supply-demand balance constraints under heterogeneous risks, ensuring that the obtained site selection scheme guarantees the system's emergency response efficiency and robustness under the most unfavorable risk scenario. The method employs strong dual transformation and branch-and-cut algorithms, ensuring the feasibility of the model and obtaining a theoretically proven global optimal solution.

[0028] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] The method of this invention mainly includes four key steps: collecting historical data, cleaning and labeling, constructing a data-driven uncertainty set based on UIA tags, constructing an original robust optimization model, and constructing a solution algorithm. The collection, cleaning, and labeling of historical data are achieved by querying an event tag UIA quantification dictionary. This dictionary can be a predefined dictionary of this invention or a custom dictionary established and maintained according to actual business needs, data characteristics, and classification standards. Constructing a data-driven uncertainty set based on UIA tags is a structure-aware multi-dimensional demand construction, decomposing uncertain demands into "basic normal demand" and "multi-source heterogeneous fluctuations," defined as the superposition of the baseline load and heterogeneous fluctuations. The original robust optimization model constructs the problem as a mini-maximum robust optimization model, minimizing the weighted total worst-case service gap across all time periods under resource and operational stability constraints. The solution algorithm constructs the model using strong duality theory, thereby transforming the difficult-to-solve two-layer model into an equivalent, solvable single-layer mixed-integer linear programming model.

[0030] This invention addresses the scenario of drone-assisted traffic accident handling. Based on historical accident data rather than prediction, it constructs a structured uncertainty set and proposes a drone-assisted heterogeneous urban traffic accident management method. The method is then used to obtain a drone location scheme that can respond to emergencies under the most unfavorable risk scenario through a branch-and-cut algorithm.

[0031] Step 1, Data Processing and Quantitative Classification:

[0032] Historical traffic incident data is cleaned and preprocessed; accident tags are extracted from the processed data, and the event tags are quantitatively classified according to the UIA multidimensional evaluation framework of urgency, importance, and ambiguity to obtain a structured dataset.

[0033] This step involves cleaning, standardizing, and quantitatively classifying the acquired historical traffic incident data to provide structured input for subsequent heterogeneous risk modeling. The data processing flow is as follows: Figure 1 As shown, this part of the operation can be implemented in accordance with the method described in the published documents, and the present invention does not make any special requirements.

[0034] The preprocessing process includes cleaning, deduplicating, and standardizing historical traffic incident data, including the date, time, location, and description of the incident, to obtain an initial dataset; then, the data is quantitatively classified according to the UIA evaluation framework defined in this invention.

[0035] This invention defines an “Urgency-Importance-Ambiguity (UIA) Evaluation Framework”, a multi-dimensional priority evaluation framework used to quantify key event labels extracted from event description texts in the initial dataset.

[0036] The framework comprises three dimensions: urgency (U), importance (I), and ambiguity (A). Urgency measures the time sensitivity of an event and the potential risk of delayed response; a higher U value indicates the event must be handled as quickly as possible. Importance measures the potential impact of an event on the overall efficiency of the transportation system and public safety; a higher I value indicates a greater disturbance to the system. Ambiguity measures the completeness and certainty of current event information; a higher A value indicates greater uncertainty in situational awareness, necessitating immediate deployment of drones to fixed-site information.

[0037] The quantization and classification process is achieved by querying the UIA quantization dictionary library for event tags. This invention utilizes a predefined, directly callable dictionary library (such as...). Figure 2As shown in the diagram, the dictionary queries the corresponding UIA value based on the event tags extracted from the data. As a concrete example, this dictionary uses a three-dimensional feature system to quantify and classify traffic accident data. The features and hierarchical relationships of each dimension are as follows:

[0038] The urgency (U) dimension categorizes urgency levels using a numerical range of 0 to 1. 0 to 0.35 represents low urgency, corresponding to the "no casualties" characteristic; 0.35 to 0.65 represents medium urgency, corresponding to the "congestion" characteristic; and 0.65 to 1 represents high urgency, corresponding to the "casualties" characteristic. This dimension further includes scenario features such as "occurring at a traffic light" and "occurring at a crosswalk," as well as event features such as "illegal parking," "breakdown," and "high traffic volume." Event features are further subdivided into subcategories such as "other," "single non-motorized vehicle accident," "single motorized vehicle accident," "non-non-motorized vehicle accident," "motorized vehicle accident," "non-motorized vehicle accident," "non-motorized vehicle accident," "non-motorized vehicle accident," "motorized vehicle accident," and "motorized vehicle accident."

[0039] Importance (I) Dimension: Based on road attributes, roads are divided into two main categories: "Internal Roads" and "Ordinary Road Surfaces". "Internal Roads" includes subtypes such as "Internal to Units", "Internal to Residential Communities", and "Internal to Parking Lots"; "Ordinary Road Surfaces" includes subtypes such as "Urban Roads", "National Highways", "County Roads", "Rural Roads", and "Elevated Expressways".

[0040] Ambiguity (A) dimension: Based on the clarity of road location, it is divided into three levels: low, medium and high. "Intersection" corresponds to low ambiguity, "road segment" corresponds to high ambiguity, and other road types correspond to medium ambiguity.

[0041] This dictionary database achieves multi-dimensional and quantifiable classification and management of traffic accident data through the combination of features from the above three dimensions.

[0042] It is important to note that Figure 2 The dictionary structure shown is merely an example. This invention allows for the creation and maintenance of custom dictionaries based on actual business needs, data characteristics, and classification standards.

[0043] Step 2, Heterogeneity and Statistical Calibration:

[0044] Uncertain demand is decomposed into basic normal demand and multi-source heterogeneous fluctuations. Based on the structured dataset after quantitative classification, a structured uncertainty set based on hierarchical statistical features is established to characterize the heterogeneity and spatiotemporal distribution differences of emergency demand.

[0045] Robust models in traditional technologies are typically built upon the assumption of homogeneous demand fluctuations. However, urban traffic risks exhibit significant structural heterogeneity. For example, a high-urgency event involving casualties may require far more emergency resources and have stricter time constraints than the sum of multiple low-urgency events. Using simple linear aggregation methods can essentially mask such long-tail risks, ultimately leading to underfitting of the model to high-consequence scenarios.

[0046] Unlike traditional robust models, this invention constructs a robust site selection planning model based on the UIA evaluation framework, which can discover significant structural heterogeneity in urban traffic risks. Furthermore, the simple summation (linear aggregation) modeling method used in traditional techniques masks this long-tail risk, leading to underfitting of the model for high-risk scenarios. Therefore, this invention, based on the UIA framework, studies the structural heterogeneity of urban traffic risks, addressing uncertain demand... It is decomposed into "basic normal demand" and "multi-source heterogeneous fluctuations", that is, it is defined as the superposition of the baseline load and heterogeneous volatility:

[0047] (1)

[0048] In the formula, This indicates an uncertain demand for emergency resources; This represents the historical average number of base accidents. Indicates the risk category, This represents the set of risk categories obtained by clustering based on the UIA framework; This represents the resource conversion factor, reflecting the different time occupation of drone operations due to different types of accidents; This represents the historical standard deviation or volatility of the k-th type of risk; , indicating the normalized first Risk-like disturbance variables.

[0049] To strictly limit perturbations and mitigate the effects of the curse of dimensionality in parameter tuning, this invention is based on uncertain requirements. A weighted set of structural uncertainties was constructed, denoted as This set is defined by a globally conservative coefficient. By parameterizing the coefficient, the risk budget is anchored to the worst-case historical observation.

[0050] Unlike traditional dual-budget sets that treat all spatial perturbations equally, this invention constructs a weighted set of structural uncertainties. At the same time, a risk hierarchy structure derived from the UIA framework is also incorporated, as shown in the following formula:

[0051] (2)

[0052] In the formula, z represents the auxiliary uncertain variable; s represents the set of real numbers; s represents a single subregion of accident demand. This represents the set of sub-regions representing accident-related needs. Indicates belonging to a specific city area A collection of accident-related sub-regions; Indicates the risk category, This represents the set of risk categories obtained by clustering based on the UIA framework; express Budgetary consumption weights for risk categories; Indicates that for the first Normalized perturbation variables for risk class; Representing an uncertain budget, indicating the region During the period The weighted total risk budget is defined as the worst-case scenario. quantiles; This represents the global conservatism coefficient; This refers to any value that the parameter can take, such as Represents for any belong ; Indicates a separate city area. Represents a collection of urban areas; Indicates a single time period. This represents a typical set of traffic periods.

[0053] Representing an uncertain budget, indicating the region During the period The weighted total risk budget is defined as the worst-case scenario. Quantiles; specifically as shown in the following formula:

[0054] (3)

[0055] In the formula, This represents the global conservatism coefficient; This represents historical observation data; This indicates the peak of risk found across all historical observation data. The definitions of the remaining symbols are consistent with the preceding definitions (hereinafter the same).

[0056] After weighted and risk hierarchy design, The computational model achieves effective dimensionality reduction, enabling the budgeting of high-dimensional uncertainties dependent on specific indicators. This is compressed into a single, interpretable control lever. This approach avoids the inherent dimensional challenges of city-level networks while preserving local specificity, thus ensuring feasibility through a large set of hyperparameters that do not require manual calibration. Furthermore, uncertainty budgeting... Directly anchored to historical observation data This rationalizes the construction of the uncertainty set, shifting the paradigm from arbitrary parameter selection to determination based on worst-case historical budgets. Finally, this parameterization provides a continuous spectrum for sensitivity analysis, enabling policymakers to quantify the marginal cost of resilience and track the Pareto frontier between system costs and risk coverage levels, thus allowing for informed trade-offs based on fiscal constraints and risk appetite.

[0057] The structured uncertainty set serves as both the input to the robust location planning model and the core domain of constraints. Its special significance for model computation lies in the fact that, during the transformation from a robust location planning model to a MILP model, the structured uncertainty set... It provides a scientific search boundary for risk identification, and through its unique linear convex set property, it transforms the originally difficult-to-solve "two-layer maximum-minimum" optimization problem into a computationally efficient single-layer mixed-integer linear programming problem through strong duality theory. This transformation avoids the curse of dimensionality in large-scale city-level block networks, ensuring that the model can find the globally optimal hangar location and resource allocation scheme within a reasonable time.

[0058] The following section will describe the specific method for utilizing the structured uncertainty set in the congestion-aware stability constraint of step 3.

[0059] Step 3, Construction of a robust site selection planning model (and its mathematical description):

[0060] The robust site selection planning model is modeled as a minimax robust optimization problem, with the goal of minimizing the weighted total system shortage risk over all time periods while satisfying resource and operational stability constraints.

[0061] This step aims to construct a core mathematical optimization model that integrates site selection, scheduling, heterogeneous risks, and response constraints into a single framework to obtain the most robust UAV hangar site selection scheme. This model is modeled as a max-min robust optimization to achieve optimal site selection in the worst-case scenario with uncertain parameters. The optimization objective of this model is to minimize the weighted worst-case underserved state across all time periods under resource and operational stability constraints. Its core feature lies in its "congestion awareness": unlike traditional capacity-constrained models, this invention explicitly introduces a safe utilization threshold into the capacity constraints, forcing the system to retain sufficient buffer capacity to cope with micro-queue fluctuations during random arrivals.

[0062] The objective function of the robust location planning model is shown in the following equation:

[0063] (4)

[0064] The minimize function aims to find a set of input parameters that minimizes (or makes as small as possible) the output of a given function (objective function).

[0065] In the formula, Indicates the time-period importance weights driven by data; Indicates the site During the period Worst-case scenario service shortage hours; Indicates a single candidate hangar site. This represents the set of candidate hangar sites.

[0066] In the robust site selection planning model of this invention, the constraints mainly include congestion-aware stability constraints, maximum resource constraints, service non-inferiority constraints, service coverage constraints, site activation logic constraints, and decision variable domain constraints. Specifically, they are shown in the following equations:

[0067] (5) (6) (7) (8) (9) (10)

[0068] In the above formulas, This indicates the maximum permissible traffic intensity to prevent system saturation; This represents the effective operation rate coefficient of unmanned aerial vehicles (UAVs). Indicates time period Total effective duration; Indicates assignment to the site The number of drones; Indicates the steady-state capacity limit; Indicates that the drone departed from the site Response sub-region The complete closed-loop time for a single task; , indicates the assignment of variables, if time period Middle Station Responsible for covering sub-regions The value is 1; express Subregion during period The historical average demand. Represents a binary variable of 0-1 if and only if at candidate point When constructing a drone hangar, its value is 1; otherwise, its value is 0. Indicates the maximum number of hangars that are allowed to be built; This represents the number of drones assigned to station j; This indicates the total number of drones that can be purchased. The service qualification matrix is ​​represented if and only if site j can cover the sub-area with a response time that is strictly superior to that of ground response personnel. Its value is 1 if it is true, and 0 otherwise.

[0069] Constraint (5) represents the congestion-aware stability criterion, which is a key constraint in the model. It requires the configured service capacity, i.e., the threshold of safe utilization. Standard adjusted capacity, plus allowable undercapacity. It must be sufficient to meet the demand, even in the face of uncertainty. In the worst-case scenario, this worst-case scenario involves structuring the uncertainty set. This is achieved as a "worst-case risk identification mechanism." Within the heterogeneous fluctuation range defined by the uncertainty set, this mechanism automatically searches for and identifies the combination of demand distributions that pose the greatest threat to system stability. This is achieved through mandatory utilization rates. This constraint implicitly keeps the system within the linear region of the queuing delay curve, preventing exponential growth in response time. It is important to note that constraint (5) employs a global capacity planning approach, rather than micro-level scheduling. By ensuring that the "configuration redundancy capacity" exceeds the "worst-case global demand" determined by this risk identification mechanism, the system's operational capacity can be systematically assessed. This approach avoids the complexity of modeling every UAV flight path during the system design phase, while the safe utilization threshold... This implicitly takes into account the dynamic queuing effect.

[0070] Constraint (6) imposes fiscal budget restrictions on infrastructure construction and fleet procurement.

[0071] Constraint (7) ensures non-inferiority of service, stipulating that service is only available when it originates from the base. Compared to ground units, drones have a significantly better response time. And the base is in an active state. Only when it is in this condition can it be allocated to a sub-region. .

[0072] Constraint (8) ensures that each sub-region has at least one base coverage to avoid service blind spots.

[0073] Constraint (9) is a site enable logic constraint that ensures that drones are only assigned to open bases.

[0074] Finally, the domain of the decision variables was defined using constraint (10).

[0075] Step 4, Model Transformation and Solution Results: Using the strong duality principle, the robust location planning model is equivalently reconstructed into a solvable single-stage mixed integer linear programming model; the solution is obtained using a solver to obtain the globally optimal hangar location scheme and the number of supporting UAVs, and the results are output.

[0076] The model constructed in step 3 presents significant computational challenges because constraints (4) and (5) exhibit a “minimal-maximal” nested structure. This form belongs to semi-infinite programming, requiring the system to remain resilient to worst-case uncertainties. This non-convex structure makes it impossible to solve directly using standard mixed-integer programming solvers.

[0077] To address this, this invention reconstructs the adversarial problem using the strong duality theorem, thereby transforming the tractable two-layer model into an equivalent, solvable single-layer mixed-integer linear programming (MILP) model. The specific steps are as follows:

[0078] Step 4-1 involves separating the internal maximization term from the constraints described in formula (5), which is termed the adversarial subproblem (P-Sub). That is, for a given time period... Hangar-sub-area allocation scheme Nature (or uncertainty) seeks to maximize the impact of fluctuations by manipulating the perturbation vector z. .

[0079] The P-Sub subproblem is defined as follows:

[0080] (11)

[0081] Subject to constraints:

[0082] (12) (13)

[0083] In the formula, This represents a collection of urban areas.

[0084] Since the original subproblem (P-Sub) is feasible and bounded, the strong duality theorem can be used to guarantee that its optimal objective value is equal to the optimal objective value of its dual subproblem (D-Sub).

[0085] The D-Sub subproblem is defined as follows:

[0086] (14)

[0087] Subject to constraints:

[0088] (15) (16)

[0089] In the formula, Indicates site During the period The worst-case total risk load; The dual variable representing the regional risk budget; Represents the dual variable of the perturbation upper limit constraint; This represents the corresponding total risk budget Γ; This indicates the upper limit of the corresponding single-point disturbance; Indicates associated sub-regions With its city area The risk budget dual value. and These are non-negative dual variables, corresponding to the regional risk budget constraint and the physical disturbance boundary, respectively.

[0090] By substituting the minimization form of the dual subproblem (D-Sub) back into the original robust constraint condition (5), the single-layer equivalent formula is finally obtained as follows:

[0091] (17)

[0092] This formula, namely the aforementioned formula (4), is subject to the following constraints:

[0093] (18) (19) (20) (twenty one)

[0094] Here, constraint (18) is an equivalent linear constraint derived from the aforementioned constraint (5) through strong duality transformation and derivation, enabling the model to be transformed from a mathematical definition to an engineering solution. Compared to constraint (5), constraint (18) eliminates the inner maximization operator in the original model. This transforms the non-convex semi-infinite programming problem, which originally had a nested "minimal-maximal" characteristic, into a deterministic single-layer linear constraint. After the transformation, the model can be directly solved to the global optimum using an existing solver. In short, constraint (18) is the equivalent linear dual of constraint (5) mentioned above, representing the congestion-aware stability criterion; compared with constraint (5), constraint (18) transforms the abstract fluctuation prediction into a deterministic resource demand.

[0095] Duality constraint (19) reveals the economic mechanism of risk allocation. Its left-hand side represents the cost of defense: where... The shadow price for regional risk budgeting reflects the marginal systemic cost of mitigating uncertainty; while This represents the local shadow price generated when the physical perturbation boundary of a specific sub-region and incident type reaches a tight constraint. The right end represents the attack cost, i.e., when the base... Service sub-region When, it happens The impact of such accidents. Therefore, constraint (19) stipulates a fundamental optimality condition: the system must allocate sufficient defense resources so that the cost of resisting any allowed accident scenario is at least no less than the damage it may cause, thereby ensuring that there is no unmitigated risk exposure within the pre-defined set of uncertainties. In short, the original model constraints (6)-(10) contained in formula (20) integrate the inherent deterministic factors in the model and define the physical boundary and coherent logic of the system. Constraint (21) defines the non-negativity of variables, ensuring that the contribution of risk disturbances to resource demand is positive, which conforms to the principle of robust optimization defense conservatism.

[0096] Through the above processing, the robust location planning model constructed in step (3) is transformed into a mixed-integer linear programming model. Based on this, the model can utilize conventional solvers and obtain a globally optimal solution. For example, by directly solving the model using a branch-and-cut algorithm compatible with modern commercial solvers, a theoretically proven globally optimal solution can be obtained in polynomial time.

[0097] The output solution results include , , and This serves as the basis for subsequent analysis in dynamic scheduling of air-ground coordination. and This constitutes the fixed physical configuration for subsequent simulations. The system's primary responsibility area is defined, providing capacity guarantees for handling worst-case scenarios. It should be noted that in discrete event simulations, The static binding constraints will be relaxed, and the system will implement dynamic scheduling based on real-time availability to further optimize execution efficiency on this robust baseline.

[0098] Indicates the site During the period Worst-case service shortage hours represent a resilience defense boundary during the scheduling process. This indicates a shortage of working hours, and the scheduling system can proactively reallocate tasks or share resources across regions to meet the needs of the stations. During the period The service needs.

[0099] Step 5, Air-Ground Cooperative Dynamic Scheduling (UIA-DCA)

[0100] In step 4, the globally optimal hangar location scheme and the number of supporting UAVs are obtained by solving a single-stage mixed-integer linear programming model. This step further uses it to implement priority allocation and air-to-ground coordination verification, realizing the operation of dynamic scheduling for air-to-ground coordination, while the location model provides physical boundaries and resource constraints for the scheduling method.

[0101] In this step, the UAVs are scheduled in real time using the air-ground cooperative UAV dynamic scheduling method (named UIA-DCA algorithm). The output of step 4 will be used as the physical boundary to verify the true emergency response capability of the site selection coordinates. Among them, UIA refers to the task scoring mechanism based on the urgency (U), importance (I), and ambiguity (A) of the task to be processed. DCA is short for dynamic cooperation zone mechanism, which specifically refers to the dynamic scheduling algorithm for finding the best feasible UAV. Constraints (7) and (20) are a time-based dynamic screening mechanism, which stipulates that: only when the UAV leaves the site... Reaching sub-region Its response time is significantly better than that of ground units ( Only when the response time is met is the site allowed to serve the region. This means that the "cooperation zone" is dynamically formed based on real-time comparisons of response times, rather than a pre-defined static administrative region.

[0102] This scheduling method belongs to the field of real-time decision-making and resource scheduling, and can be used to solve the problem of dynamic allocation of heterogeneous resources for multiple priority tasks in urban emergency response. Through the task scoring and dynamic cooperation zone mechanism of the UIA algorithm, it can achieve rapid and robust response to emergencies. Its core mechanism adopts a two-stage framework of priority-based sequential allocation and air-ground cooperation verification: the first stage evaluates and ranks tasks online, and the second stage dynamically allocates optimal UAV resources for high-priority tasks and verifies the necessity of ground cooperation. Through this phased collaborative optimization process, it can optimize the overall resource utilization efficiency while ensuring timely response to high-risk tasks.

[0103] In addition, this step can also be used for monitoring (i.e., by comparing the "actual gap" with the "worst gap predicted by the model"). This provides decision-making reference for management departments; if the actual gap frequently approaches the model's warning value, it means that a new round of system planning is needed.

[0104] The specific steps of the air-ground coordinated dynamic scheduling process are as follows:

[0105] Step (1): Input parameters and initialization

[0106] Setting the parameters required for the Air-Ground Cooperative Dynamic Scheduling (UIA-DCA) operation is a user-operated step; then, the real-time operating environment required for the operation is initialized. During execution, the site selection results (physical coordinates of the takeoff and landing field) and resource allocation scheme (maximum number of UAVs) pre-calculated by the aforementioned single-stage mixed-integer linear programming model must be invoked. Input parameters include: the set of tasks to be processed. (Located in the area) Idle drone collection (Located in the area) Current simulation time Drone parameters (speed) Vertical takeoff and landing time Homework time Battery capacity ), Ground responder parameters (velocity) Preparation time ).

[0107] Step (2): Task ranking based on UIA scores

[0108] This step is the evaluation phase using UIA: for the region All pending tasks within The tasks are sorted in descending order based on a comprehensive score consisting of urgency (U), importance (I), and ambiguity (A) to obtain an ordered list of tasks. .

[0109] The following steps (3) and (4) belong to the real-time scheduling and allocation logic, which is executed online by the dynamic scheduling system based on the static physical configuration (location and capacity limit) output by the aforementioned MILP model.

[0110] Step (3): Sequential task allocation loop

[0111] Determine an ordered task list Is it non-empty?

[0112] If so, retrieve the task with the highest current priority. And execute the sub-processes of steps (3.1) to (3.3);

[0113] If not, proceed to step (4) and return the result.

[0114] Step (3.1): Best feasible drone search

[0115] Initialize the best drone Empty, minimum arrival time The value is infinity. Iterate through all idle drones. Perform the following sub-steps:

[0116] Step (3.1.1): Calculate the critical time

[0117] Computational drones From current location Fly to the mission point Flight time With drones From the mission point Return to its respective hangar return time .

[0118] Step (3.1.2): Energy feasibility verification

[0119] Calculate execution task Total energy demand ,Right now , , With safety buffer energy The sum of the values ​​determines the drone's performance. Remaining battery power Is it greater than or equal to? :

[0120] If so, then continue comparing their flight times. With the current minimum time ;

[0121] If not, the drone is unavailable; check the next one.

[0122] Step (3.1.3): Update the best candidate

[0123] If drone Enough energy and Then update And tentatively .

[0124] Step (3.2): Verification of air-ground coordinated response constraints

[0125] Determine if a usable drone has been found:

[0126] If so, then perform the following collaborative verification sub-step:

[0127] Calculate the estimated arrival time of the drone ; Calculate the distance from the task point Recent Ground Responder Database distance ; Calculate the estimated arrival time of ground responders ;

[0128] Compare the arrival times of the two:

[0129] like Then determine the task. It is better for ground responders to handle this; mark it as "ground handling" and remove it from the pending list. and Remove from the middle; at this point, skip the current loop and return to step (3) to process the next task;

[0130] like Then proceed to step (3.3).

[0131] If not, the current task j cannot be assigned temporarily and will remain in the list for the next task to be processed.

[0132] Step (3.3): Task allocation and resource status update

[0133] The task Officially assigned to the best drone Upgrade drones The status is "in flight";

[0134] The task From the list of pending processes and ordered lists Remove from;

[0135] drones From idle drone ensemble Removed from the middle.

[0136] Return to step (3) to process the next priority task.

[0137] Step (4): Output the allocation scheme and end.

[0138] The loop ends when all tasks are completed or no resources are available for allocation, and the final task allocation vector is output. .

[0139] In summary, the implementation of this invention can be roughly divided into the following three stages:

[0140] Phase 1: Offline layer, using a robust site selection planning model to implement the original site selection model, determining the site selection and the number of drones.

[0141] Phase Two is the computationally equivalent form of Phase One. Through dual transformation, the originally incalculable two-layer model is transformed into a single-layer formula that the computer can compute, namely, a single-stage mixed-integer linear programming (MILP). Its function is the same as that of Phase One. This model belongs to the system planning layer and is responsible for determining the hangar location and the number of UAVs. Its specific calculations are completed before UIA-DCA starts running and it does not directly participate in each round of UIA-DCA scheduling.

[0142] Phase 3: The Air-Ground Cooperative Dynamic Scheduling Method (UIA-DCA) is an online operation layer. It performs UAV scheduling operations based on the results of Phase 2, and returns the actual gap situation that may exist. UAV task sorting, assignment, and path calculation belong to the real-time scheduling layer, which inherits the results of MILP as constraints.

[0143] The usage of robust site selection planning models can be summarized as follows:

[0144] (1) After integrating historical data through the UIA mechanism, input the following into the robust site selection planning model: region Task demand frequency Each station To the area Response time Geographic coverage matrix Risk budget (Used to characterize the conservatism of task fluctuations); Total site construction limit Maximum number of drones Preset security utilization threshold .

[0145] (2) Then run the MILP model that has been transformed.

[0146] After receiving data input, the UAV emergency response system transforms the robust site selection planning model into a single-stage mixed-integer linear programming model using strong duality theory. Subsequently, a high-performance optimization solver is invoked to perform global optimization on this single-stage mixed-integer linear programming model, directly achieving coverage of the "worst-case scenario" for the site selection scheme in a single calculation.

[0147] (3) Output: Output of the entire single-stage mixed-integer linear programming model , , and The data is received by UIA-DCA and used for analysis. It refers to the location of the drone take-off and landing site selected by the model; This refers to the number of drones configured at each takeoff and landing site selected by the model; Determine the site For the region exist The system's primary responsibility for a given time period (i.e., the expected task distribution pattern under this configuration). Inform the scheduling system of potential service gaps under worst-case scenarios for reference.

[0148] In the above output, The physical coordinates of the UAV take-off and landing sites within the urban space were determined. These coordinates formed the starting and ending points for all path planning algorithms in the subsequent dynamic scheduling simulated / commanded by UIA-DCA. A resource holding limit is defined for each site. The dynamic scheduling system must allocate tasks within this constraint, and scheduling commands cannot allocate more than this limit without triggering cross-regional support protocols. Resources of scale. Sites were specified Pair of sub-regions exist The master-slave relationship within a time period. In dynamic scheduling, To provide reserve capacity for worst-case scenarios, the scheduling system will prioritize ensuring that tasks within the primary responsibility area receive an immediate response.

[0149] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A data-driven robust location selection method for heterogeneous risk management in drone-assisted accident handling, characterized in that, Includes the following steps: (1) Preprocess the historical data of traffic incidents; extract accident labels from the processed data, and quantify and classify the accident labels according to the UIA multidimensional evaluation framework of urgency, importance and ambiguity to obtain a structured dataset; (2) Decompose the uncertain demand into basic normal demand and multi-source heterogeneous fluctuations. Based on the structured dataset after quantitative classification, establish a structured uncertainty set based on hierarchical statistical features to characterize the heterogeneity and spatiotemporal distribution differences of emergency demand. (3) The robust site selection planning model is modeled as a minimax robust optimization problem, with the goal of minimizing the total system shortage risk after weighting over all time periods while satisfying resource and operational stability constraints; (4) The robust location planning model is reconstructed into a solvable single-stage mixed integer linear programming model by using the strong duality principle; the solution is obtained by using a solver to obtain the globally optimal hangar location scheme and the number of supporting UAVs, and the results are output.

2. The method according to claim 1, characterized in that, The historical traffic incident data includes at least the following information: the date, time, location, and description of the incident; the preprocessing includes cleaning, deduplication, and standardization.

3. The method according to claim 1, characterized in that, The specific operations for quantifying and classifying data include: First, defining a dictionary that can be directly accessed; the data in the dictionary contains quantified UIA accident tags, and through the combination of features of three dimensions—urgency (U), importance (I), and ambiguity (A)—multi-dimensional and quantifiable classification and management of traffic accident data is achieved; then, accident tags are extracted from the data, and the corresponding UIA values ​​are queried using the dictionary.

4. The method according to claim 3, characterized in that, The features and hierarchical relationships in the dictionary are as follows: Urgency level dimension: The urgency level is divided into numerical ranges from 0 to 1, including: low urgency corresponding to the "no casualties" feature, medium urgency corresponding to the "congestion" feature, and high urgency corresponding to the "casualties" feature; each urgency level is further divided into scene features and event features, and the event features are further subdivided into subtypes composed of event participants. Importance dimension: Based on road attributes, roads are divided into two main categories: "internal roads" and "ordinary road surfaces"; each category is further subdivided into multiple subtypes. Ambiguity dimension: Based on the clarity of road location, it is divided into three levels: low, medium and high. "Intersection" corresponds to low ambiguity, "road segment" corresponds to high ambiguity, and other road types correspond to medium ambiguity.

5. The method according to claim 1, characterized in that, The total accident demand is decomposed into the superposition of basic normal demand and multi-source heterogeneous fluctuations, and a risk hierarchy structure derived from the UIA multidimensional assessment framework is incorporated to construct a weighted uncertainty set, as shown in the following formula: ； In the formula, z represents the auxiliary uncertain variable; s represents the set of real numbers; s represents a single subregion of accident demand. This represents the set of sub-regions representing accident-related needs. Indicates belonging to a specific city area A collection of accident-related sub-regions; Indicates the risk category, This represents the set of risk categories obtained by clustering based on the UIA framework; express Budgetary consumption weighting for risk categories; Indicates that for the first Normalized perturbation variables for risk class; Representing an uncertain budget, indicating the region During the period The weighted total risk budget is defined as the worst-case scenario. quantiles; This represents the global conservatism coefficient; This refers to any value that the parameter can take, such as Represents for any belong ; Indicates a separate city area. Represents a collection of urban areas; Indicates a single time period. This represents a typical set of traffic periods.

6. The method according to claim 1, characterized in that, The objective function of the robust location planning model is shown in the following equation: ； In the formula, Indicates the time-period importance weights driven by data; Indicates the site During the period Worst-case scenario service shortage hours; Indicates a single candidate hangar site. Represents the set of candidate hangar sites; Indicates a single time period. This represents a typical set of traffic periods.

7. The method according to claim 1, characterized in that, The constraints of the robust site selection planning model include: congestion perception stability constraint, maximum resource constraint, service non-inferiority constraint, service coverage constraint, site activation logic constraint, and decision variable domain constraint.

8. The method according to claim 7, characterized in that, For the most critical congestion-aware stability constraint in the robust location planning model, its internal maximization term is separated as an adversarial subproblem. Then, the strong duality theorem is used to ensure that the optimal objective value of the adversarial subproblem is equal to the optimal objective value of its dual subproblem. By substituting the minimization form of the dual subproblem back into the original congestion-aware stability constraint, a single-layer equivalent formula that can be directly solved using a solver is finally obtained.

9. A dynamic scheduling method for air-ground cooperation in unmanned aerial vehicle (UAV)-assisted traffic accident handling, characterized in that, Based on the heterogeneous risk management data-driven robust location method described in claim 1, this method obtains the globally optimal hangar location scheme and the number of supporting UAVs by solving a single-stage mixed-integer linear programming model; further, it is used to implement priority allocation and air-to-ground coordination verification, realizing the operation of dynamic air-to-ground coordinated scheduling; the location model provides physical boundaries and resource constraints for the scheduling method; specifically, it includes the following steps: (1) Set the parameters required for the air-ground coordinated dynamic scheduling operation, and then initialize the operating environment; (2) Sort all tasks in descending order based on their urgency, importance, and ambiguity as a comprehensive score to obtain an ordered task list; then, the single-stage mixed-integer linear programming model performs steps (3)-(4): (3) If the ordered task list is not empty, then take out the task with the highest priority and execute the sub-processes of best feasible UAV search, air-ground cooperative response constraint verification, task allocation and resource status update. (4) When all tasks are completed or no resources are available for allocation, end the loop and output the final task allocation vector.

10. The method according to claim 9, characterized in that, During the search for the best feasible UAV, the following sub-processes are executed: calculating critical time, verifying energy feasibility, and updating the best candidate. When performing air-to-ground cooperative response constraint verification, determine whether an available UAV has been found. If so, perform the following cooperative verification sub-steps: calculate and compare the estimated arrival times of the UAV and the ground responder, and determine and mark the better one.

Images