Wounded post-delivery decision-making method, device and equipment and storage medium of wounded post-delivery decision-making method

Abstract

The invention provides a wounded post-delivery decision-making method, device and equipment and a storage medium thereof, and relates to the technical field of medical informationization, and the method comprises the steps: obtaining environment state information at a current moment; performing feature vectorization processing to obtain a medical institution embedded vector and a transportation tool embedded vector; calculating through an attention mechanism to obtain an initial attention preference matrix; correcting the initial attention preference matrix; and performing action decoding to generate a joint scheduling action of the transportation tool. According to the method, a variable-length sequence is processed by using a Transform structure, and global adaptive modeling of dynamic medical resources is realized. By introducing a distance attenuation item to correct an attention matrix and explicitly introducing geographic space constraints, invalid long-distance pairing is effectively inhibited. The mechanism gives consideration to timeliness and resource utilization rate, and the decision-making efficiency and robustness of the system in a complex environment are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of medical information technology, and in particular to a method, device, equipment and storage medium for decision-making regarding the evacuation of wounded patients. Background Technology

[0002] In major emergency rescue operations, the evacuation of the wounded faces extremely high complexity and timeliness challenges. The large number of wounded, their wide geographical distribution, and the dynamic changes in their conditions, coupled with the continuous fluctuations in external factors such as road conditions, weather, and unknown threats, create a highly uncertain decision-making environment. Against this backdrop, how to achieve rapid and precise allocation of medical resources and transportation capacity under limited resources and complex and ever-changing constraints has become a core issue in improving the response capability and operational efficiency of the emergency medical rescue system.

[0003] Traditional medical evacuation dispatch primarily relies on a "static-centralized" model based on experience-based plans and manual command, making decisions centrally through standardized processes. To improve dispatch efficiency, existing technologies attempt to introduce operations research methods, abstracting the medical evacuation system as a weighted directed graph and using linear programming or mixed-integer programming to find the optimal solution. Furthermore, for large-scale dispatching problems, heuristic algorithms such as genetic algorithms, ant colony optimization, and particle swarm optimization are widely used, searching for approximate solutions by simulating natural evolution mechanisms. In recent years, with the development of artificial intelligence technology, reinforcement learning, as a data-driven paradigm, has begun to be applied to medical evacuation scenarios. Utilizing algorithms such as deep Q-networks (DQN) or proximal policy optimization (PPO), it learns dispatching strategies through interaction with the environment to cope with nonlinear and high-dimensional decision-making tasks.

[0004] However, the aforementioned existing technologies have significant limitations when facing highly dynamic environments. Manual command, under conditions of incomplete information and time constraints, is limited by the cognitive abilities of commanders and information transmission delays, making it difficult to achieve globally optimal coordination and easily leading to Pareto suboptimal resource allocation. Traditional operations research methods rely on strict linear or convexity assumptions, making it difficult to characterize the ubiquitous non-stationary characteristics, and their computational complexity increases exponentially with scale, failing to meet the requirements for second-level response. While heuristic algorithms reduce computational pressure to some extent, they generally suffer from strong randomness in results, high parameter sensitivity, and insufficient generalization ability. Although reinforcement learning improves dynamic response capabilities, conventional reinforcement learning models typically rely on fixed-structure neural networks, facing the challenge of unadaptive matching of input structures when dealing with scenarios where the number of wounded, available vehicles, and medical institutions change dynamically over time.

[0005] In summary, current medical evacuation decision-making technologies primarily face the problem of insufficient adaptability to complex and dynamic environments. On the one hand, the spatial topology and road accessibility of the field environment are highly uncertain, and simple feature vector inputs are insufficient to fully represent complex geospatial constraints and logical relationships. On the other hand, existing decision-making models struggle to effectively cope with the dynamic expansion or contraction of the state space and action space as the task progresses. How to overcome the dual challenges of variable input structure and complex spatial relationships under multidimensional nonlinear constraints, and achieve adaptive intelligent decision-making under multi-objective trade-offs, is a problem that current medical evacuation auxiliary decision-making technologies urgently need to solve. Summary of the Invention

[0006] In a first aspect, the present invention provides a method for decision-making on the evacuation of wounded personnel, which obtains environmental status information at the current moment; the environmental status information includes at least the status information of medical institutions and the status information of transportation vehicles. The status information of the medical institution and the status information of the transportation vehicle are respectively processed into feature vectors to obtain the embedding vectors of the medical institution and the transportation vehicle. A policy network based on the Transformer structure is constructed. The embedding vectors of the medical institutions and the transportation vehicles are used as input sequences. An initial attention preference matrix is ​​calculated through an attention mechanism. The attention preference matrix is ​​used to represent the transfer priority relationship between the transportation vehicles and the medical institutions. The initial attention preference matrix is ​​corrected using a distance attenuation term to obtain a corrected attention preference matrix; wherein the distance attenuation term is calculated based on the distance between the vehicle and the target medical institution. Action decoding is performed based on the modified attention preference matrix to generate joint scheduling actions for transportation vehicles.

[0007] In an optional implementation, the step of performing feature vectorization processing on the status information of the medical institution and the status information of the transportation vehicle includes: For each medical institution, features such as institution level, geographical location, acceptable injury vector, number of beds, material reserves, inpatient load, and turnover time are extracted and used to obtain a vector representation of the medical institution through a mapping function; For each mode of transport, features such as transport capacity, current location, available injury distribution, and available time window are extracted and concatenated or fused with the embedding vector of its affiliated medical institution. The vector representation of the mode of transport is then obtained through a mapping function.

[0008] In an optional implementation, obtaining the vector representation of the medical institution via the mapping function, and obtaining the vector representation of the transportation vehicle via the mapping function, includes: Construct a learnable linear projection layer or multilayer perceptron as the mapping function; The mapping function is used to project the features of medical institutions and transportation vehicles with different feature dimensions into a unified feature space with the same dimension, so as to meet the input dimension requirements of the Transformer policy network.

[0009] In an optional implementation, the initial attention preference matrix is ​​corrected using a distance attenuation term, as expressed by the following expression: ; in, This represents the revised attention score; Represents environmental constraints; Represents the distance penalty coefficient; Represents the distance decay function; This represents the distance between the current location of the vehicle or its affiliated organization and the target medical facility.

[0010] In an optional implementation, the distance decay function may be of the form of linear decay, exponential decay, or embedding decay; and / or, The environmental constraints are used to quantify the impact of road conditions and weather conditions on the transportation process.

[0011] In an optional implementation, when the distance decay function is an embedded decay form, the distance is divided into several discrete intervals, and a corresponding trainable scalar weight is established for each discrete interval as the function output; or, When the distance decay function is in the form of exponential decay, the function value of the distance decay function increases exponentially with the increase of the distance traveled.

[0012] In an optional implementation, the step of decoding the action based on the modified attention preference matrix includes: The modified attention preference matrix is ​​treated as a soft policy distribution; Based on the soft strategy distribution, and under the conditions of satisfying transportation capacity, injury matching, agency reception constraints and time window constraints, feasible transportation tasks are selected sequentially from high preference values ​​to low preference values ​​to construct joint scheduling actions for transportation vehicles.

[0013] In an optional implementation, the construction of the policy network based on the Transformer structure further includes: A value identifier, token, is introduced into the input sequence; A multilayer perceptron structure is connected outside the output vector of the value identifier token as a value network to fit the global state value.

[0014] In an optional implementation, the casualty evacuation decision-making method further includes the step of training the policy network; Training the policy network includes: The PPO algorithm is optimized using a near-end strategy, and the policy network parameters are updated based on the reward signal from environmental feedback. The reward signal is calculated based on a combination of factors including transfer time, resource utilization, institutional load balancing, success rate of wounded personnel treatment, and delay penalty.

[0015] In an optional implementation, the step of constructing a policy network based on the Transformer structure, before calculating the initial attention preference matrix through the attention mechanism, further includes: The embedding vectors of the medical institutions and the transportation vehicles are combined to form an input sequence; Type encoding and position encoding are superimposed on each element in the input sequence to distinguish the category attributes and sequence position information of medical institutions and transportation vehicles.

[0016] In an optional implementation, the step of calculating the initial attention preference matrix through the attention mechanism includes: The input sequence is modeled using a multi-head self-attention mechanism to calculate an attention score that reflects the dependency between the transportation vehicle and the medical institution. The attention scores are normalized to output the initial attention preference matrix.

[0017] In an optional implementation, after generating the joint scheduling action of the transportation vehicles, the method further includes: Perform joint scheduling actions for the transportation vehicles; and update the load information in the status information of the medical institution, the availability information in the status information of the transportation vehicles, and the road condition information in the environmental status information based on the results of the joint scheduling actions. The updated status information of the medical institution, the status information of the transportation vehicle, and the road condition information are used as the environmental status information for the next moment. The process of performing feature vectorization on the status information of the medical institution and the status information of the transportation vehicle is then returned to perform the next moment's evacuation decision.

[0018] Secondly, the present invention provides a casualty evacuation decision-making device, comprising: The acquisition module is used to acquire the environmental status information at the current moment; the environmental status information includes at least the status information of the medical institution and the status information of the transportation vehicle; The vector module is used to perform feature vectorization processing on the state information of the medical institution and the state information of the transportation vehicle to obtain the embedding vector of the medical institution and the embedding vector of the transportation vehicle. The computation module is used to construct a policy network based on the Transformer structure. It takes the embedding vectors of the medical institutions and the embedding vectors of the transportation vehicles as input sequences and calculates an initial attention preference matrix through an attention mechanism. The attention preference matrix is ​​used to characterize the transfer priority relationship between the transportation vehicles and the medical institutions. The correction module is used to correct the initial attention preference matrix using a distance attenuation term to obtain a corrected attention preference matrix; wherein the distance attenuation term is calculated based on the distance between the vehicle and the target medical institution. The decoding module is used to decode actions based on the corrected attention preference matrix and generate joint scheduling actions for transportation vehicles.

[0019] Thirdly, the present invention provides a computer device including a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the casualty evacuation decision-making method described in any of the foregoing embodiments.

[0020] Fourthly, the present invention provides a computer storage medium storing a computer program, which, when executed on a processor, implements the casualty evacuation decision-making method according to any one of the foregoing embodiments.

[0021] The embodiments of this application have the following beneficial effects: By constructing a policy network based on the Transformer structure and combining it with an attention mechanism to represent the preferences and relevance between nodes, this method effectively addresses the challenge of uncertain decision input dimensions caused by the dynamic changes in the number of wounded, medical institutions, and transportation vehicles in an environment. Unlike traditional models that rely on fixed input structures, the Transformer structure processes the state information of medical institutions and transportation vehicles as variable-length sequence inputs and uses an attention mechanism to perform global context modeling of the dependencies between all units. This approach not only enables the model to adaptively perceive the current resource distribution situation but also captures high-dimensional matching features between transportation vehicles and medical institutions in complex heterogeneous information, thus laying the foundation for achieving multi-node collaborative scheduling.

[0022] The key to this method's improved physical interpretability and effectiveness lies in modifying the initial attention preference matrix using a distance decay term. By transforming the actual distance between the transportation vehicle and the target medical institution into an explicit decay constraint, which is directly applied to the decision layer of the neural network, this method successfully integrates geospatial constraints into the deep learning inference process. This mechanism allows the model to naturally suppress invalid pairings involving long distances or high costs while focusing on abstract features such as injury matching, making the decision results more consistent with the transportation patterns of the physical world. This not only avoids the long-distance detour scheduling problems that may arise from purely data-driven approaches but also significantly improves the efficiency of searching for the optimal solution.

[0023] Finally, joint scheduling actions are generated based on the preference matrix after spatial constraint correction, achieving an effective transformation from soft policy distribution to hard execution instructions. This method integrates data-driven intelligent perception capabilities with rule-based physical constraints, enabling the output of scheduling schemes that balance timeliness and resource utilization while ensuring treatment needs are met. This significantly enhances the system's rapid response capability and decision robustness in highly dynamic and uncertain environments. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and therefore should not be considered as a limitation on the scope of protection of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the hardware operating environment involved in an embodiment of the casualty evacuation decision-making method of the present invention; Figure 2 This is a flowchart illustrating Embodiment 1 of the wounded evacuation decision-making method of the present invention; Figure 3 This is a detailed flowchart of step S200 in Embodiment 2 of the wounded evacuation decision-making method of the present invention; Figure 4 This is a detailed flowchart of step S500 in embodiment 4 of the wounded evacuation decision-making method of the present invention; Figure 5 This is a detailed flowchart of step S300 in embodiment 5 of the wounded evacuation decision-making method of the present invention; Figure 6 This is a schematic diagram of the module connections of the casualty evacuation decision-making device of the present invention. Detailed Implementation

[0026] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0027] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0028] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.

[0029] Furthermore, the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0030] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0031] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0032] like Figure 1 The diagram shown is a structural schematic of the hardware operating environment of the terminal involved in an embodiment of the present invention.

[0033] The casualty evacuation decision-making system (device) of this invention can be a PC, or a mobile terminal device such as a smartphone, tablet, or portable computer. This visual navigation optimization system may include: a processor 1001, such as a CPU; a network interface 1004; a user interface 1003; a memory 1005; and a communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, an input unit such as a keyboard, or a remote control; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or a stable memory, such as a disk storage device. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001. Optionally, the casualty evacuation decision-making system may also include RF (Radio Frequency) circuitry, audio circuitry, a Wi-Fi module, etc. In addition, the visual navigation optimization system can also be equipped with other sensors such as gyroscopes, barometers, hygrometers, thermometers, and infrared sensors, which will not be described in detail here.

[0034] Those skilled in the art will understand that Figure 1 The visual navigation optimization system shown is not intended to limit it and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a data interface control program, a network connection program, and a visual navigation optimization program.

[0035] In summary, the method provided by this invention utilizes the Transformer structure to adapt to dynamic changes in resources and introduces spatial constraints through distance decay correction, effectively avoiding ineffective long-distance scheduling and significantly improving decision-making efficiency and robustness in complex environments.

[0036] Example 1 Reference Figure 2 This embodiment provides a method for making decisions on the evacuation of wounded personnel, including: Step S100: Obtain the current environmental status information; the environmental status information includes at least the status information of the medical institution and the status information of the transportation vehicle.

[0037] This step refers to the system collecting environmental data in real time as a basis for decision-making. The processing includes data acquisition and integration to ensure that the acquired information reflects the current situation.

[0038] Its input content contains at least the states of two types of core objects: (1) Status information of medical institutions: This can refer to the current reception capacity and load of medical institutions (including but not limited to field hospitals and first aid stations).

[0039] (2) Status information of transport vehicles: This usually refers to the location and availability of vehicles (such as ambulances and helicopters) performing evacuation missions.

[0040] This step yields a raw, unstructured, or numerical dataset of environmental states.

[0041] Specifically, this information can be obtained through sensor networks, information systems, or manual reporting. For example, it can be obtained regarding the current number of injured people, available beds, and resource occupancy rates at various medical institutions; and the current coordinates and carrying capacity of various transportation vehicles.

[0042] Step S200: The status information of the medical institution and the status information of the transportation vehicle are respectively processed into feature vectors to obtain the embedding vector of the medical institution and the embedding vector of the transportation vehicle.

[0043] This step involves transforming heterogeneous raw data into a numerical form that can be understood by computer models (especially neural networks). Specifically, feature extraction and embedding techniques can be used to obtain two sets of high-dimensional vectors: medical institution embedding vectors (representing institutional features) and transportation vehicle embedding vectors (representing vehicle features).

[0044] This step eliminates the dimensional differences between different types of data (such as location coordinates, category labels, and numerical values), mapping them to a unified feature space to facilitate subsequent computation by the neural network.

[0045] For medical institutions, features such as their level, location, and material reserves are extracted and transformed into a vector e through a mapping function (such as a fully connected layer or a lookup table). mj For transportation vehicles, features such as location and capacity are extracted and transformed into a vector e using a mapping function. ui .

[0046] Step S300: Construct a policy network based on the Transformer structure, using the embedding vectors of the medical institutions and the transportation vehicles as input sequences, and calculate an initial attention preference matrix through an attention mechanism; wherein, the attention preference matrix is ​​used to characterize the transfer priority relationship between the transportation vehicles and the medical institutions.

[0047] In this embodiment, the method (feature extraction module) is constructed using a Transformer architecture.

[0048] It should be noted that the "Transformer architecture" here refers broadly to various deep learning models or their variants based on the Transformer architecture. As a specific, non-limiting, preferred embodiment, the BERT (Bidirectional Encoder Representations from Transformers) model can be used as a concrete implementation of this Transformer architecture. The BERT model is essentially the encoder part of the Transformer, and compared to the standard Transformer, it performs better in handling contextual semantic understanding tasks.

[0049] The above steps constitute the core reasoning process for decision-making. The Transformer architecture is used to process the aforementioned vector. Since the number of medical institutions and transportation vehicles is dynamically changing, the Transformer structure is naturally well-suited for handling such variable-length sequence inputs. The processing utilizes an attention mechanism to calculate the interdependencies between elements in the input sequence, thereby obtaining an initial attention preference matrix.

[0050] This matrix can be a two-dimensional numerical table, where each value (weight) represents the "matching degree" or "transfer priority" between a particular mode of transport and a particular medical institution. This step can capture global dependencies. For example, it not only considers whether a vehicle is close to a hospital, but also combines global supply and demand relationships (such as a hospital that is close but full, in which case the attention weight will be reduced), thereby achieving global coordination.

[0051] Step S400: The initial attention preference matrix is ​​corrected using a distance attenuation term to obtain a corrected attention preference matrix; wherein the distance attenuation term is calculated based on the distance between the vehicle and the target medical institution.

[0052] The steps described above explicitly introduce the geographical constraints of the physical world into the neural network decision-making process. The initial matrix calculated by the Transformer is mainly based on feature matching and may ignore the cost of excessively long actual distances. This step uses mathematical operations to "penalize" the initial score, thereby obtaining a revised attention preference matrix. In this matrix, the preference value for task pairings that are too far apart will be significantly reduced.

[0053] This step avoids generating scheduling schemes that do not conform to physical laws (such as choosing the longer route instead of the shorter one), thus improving the practical feasibility and execution efficiency of the decision.

[0054] Step S500: Decode the action based on the corrected attention preference matrix to generate the joint scheduling action of the transportation vehicle.

[0055] This step transforms the probability distribution (soft policy) output by the neural network into specific execution instructions (hard actions), thereby generating a joint scheduling action for transportation vehicles. Specifically, it specifies where each transportation vehicle should go and which wounded individuals it should transport. This step solves the combinatorial optimization problem in scenarios involving multiple vehicles and multiple hospitals, outputting a directly executable scheduling scheme.

[0056] Specifically, the modified attention matrix can be viewed as a probability distribution or a priority list. Based on this matrix, and under the premise of satisfying practical constraints (such as vehicle passenger capacity limits and hospital admission limits), matching is performed sequentially from high preference values ​​to low preference values. For example, if transportation vehicle A has the highest modified preference value for medical institution B, and both have resources / available spaces, then a scheduling instruction of "A to B" is generated.

[0057] Example 2 Reference Figure 3 Based on the foregoing embodiments, this embodiment provides a method for decision-making regarding the evacuation of wounded personnel. Step S200 involves performing feature vectorization processing on the status information of the medical institution and the status information of the transport vehicle, including: Step S210: For each medical institution, extract the features of institution level, geographical location, acceptable injury vector, number of beds, material reserves, inpatient load and turnover time, and obtain the vector representation of the medical institution through a mapping function.

[0058] This step aims to construct a digital twin description of the medical institution. Specifically, it begins by acquiring the raw data. For discrete / categorical features, this includes "institution level" and "acceptable injury vector" (e.g., whether the institution can handle burns, traumatic brain injuries, etc., typically encoded as 0 / 1); for continuous / numerical features, this includes "geographic location" (latitude and longitude coordinates), "number of beds" (integers), "supply reserves" (quantity or percentage), "inpatient load" (current number of injured / total capacity), and "turnover time" (average treatment time). Then, a mapping function is applied, which is a mathematical transformation operation. Since the raw data has different dimensions and physical units (e.g., distance is in kilometers, while the number of beds is in units), it cannot be directly input into the Transformer. The mapping function φ... m (·) acts as an "encoder".

[0059] Specifically, for categorical features (such as institution level), one-hot encoding or embedding layers can be used to convert them into low-dimensional dense vectors. For numerical features (such as the number of beds), normalization can be performed first (e.g., mapping to 0-1) to eliminate the influence of dimensions.

[0060] Specifically, the mapping function concatenates all the processed features into a long vector, and then projects it onto a fixed dimension (i.e., the Transformer's model dimension d) through a multilayer perceptron (MLP) or a fully connected layer. model Thus, a numerical vector e with fixed dimensions is obtained. mj , representing the overall status of the j-th medical institution.

[0061] This step abstracts the complex state of medical resources into a computer-computable mathematical form, preserving not only static attributes (location, level) but also integrating dynamic attributes (load, turnover), enabling the decision-making model to perceive whether the hospital is "busy" and "can treat patients".

[0062] Step S220: For each means of transport, extract features such as transport capacity, current location, transportable injury distribution, and available time window, and concatenate or fuse these features with the embedding vector of its affiliated medical institution. The vector representation of the means of transport is then obtained through a mapping function.

[0063] This step is used to construct the feature representation of the transportation vehicle, which includes a key "fusion" operation. This involves extracting its own features, such as "transportation capacity," which can be the number of passengers (e.g., 2 seriously injured or 4 slightly injured); "current location," which can be real-time coordinates; "transportable injury distribution," which can be the emergency medical equipment the vehicle possesses and the types of injured it can transport; and "available time window," which can be when the vehicle is available and the remaining fuel allows for a certain amount of driving time. Then, splicing or fusion (a key processing step) can be performed, specifically by operating on the "embedded vector of the affiliated medical institution." This means that the transportation vehicle is not an independent entity; it carries the genes of its affiliated unit (e.g., the hospital). The processing action can be to extract the transportation vehicle's own feature vector v... self Vector e of its affiliated medical institution mj Perform splicing (e.g., [v]) self , e mj () or weighted summation.

[0064] Then, similar to the process used in medical institutions, the combined features obtained from the above fusion are mapped using the mapping function φ. m Transformation is performed on (·) (such as fully connected layers).

[0065] Specifically, assuming the vector of the transportation vehicle after feature processing is The vector of the organization to which it belongs is The splicing operation is as follows: The mapping operation is as follows: , where W and b are learnable parameters, and the final output dimension is adjusted to be consistent with the dimension of the medical institution vector.

[0066] The final vector representation e of the transportation vehicle can be obtained through the above processing. ui .

[0067] The vector of a transportation vehicle implicitly contains information about its "origin." This allows the model to know not only where the vehicle is when calculating attention, but also implicitly which base the vehicle might need to return to after transporting the wounded, or whether the base has the capability to support long-distance evacuation. Furthermore, it addresses the problem of insufficient information for a single transportation vehicle, enhances the expressive power of features, and helps the Transformer better calculate the matching relationship between transportation vehicles and target hospitals.

[0068] In some embodiments, step S210, obtaining a vector representation of the medical institution via a mapping function, and step S220, obtaining a vector representation of the transportation vehicle via a mapping function, include: (1) Construct a learnable linear projection layer or multilayer perceptron as the mapping function.

[0069] This step defines the specific structure of the mathematical model used to process eigenvectors.

[0070] The term "learnable" means that the mapping function includes weights and biases. These parameters are not fixed but are continuously optimized during subsequent reinforcement learning training as the policy network is updated to extract the most useful feature information for decision-making.

[0071] The aforementioned "linear projection layer" refers to a single-layer fully connected network that performs only linear transformations.

[0072] The aforementioned "Multilayer Perceptron (MLP)" refers to a neural network structure that includes at least one hidden layer and a nonlinear activation function, and has the ability to extract nonlinear features.

[0073] Specifically, during the system initialization phase, two independent neural network modules can be established for "medical institution characteristics" and "transportation vehicle characteristics," respectively. Define the function φ. m (·), the network structure used to process data from medical institutions; and, defining the function φ. u (·), a network structure used for processing transportation data.

[0074] Specifically, this involves the following two aspects: A. Implementation of linear projection layer: Construct a matrix operation y = Wx + b, where x represents the input features and W represents the dimension d. out ×d in The weight matrix is ​​denoted by b, where b represents the bias vector.

[0075] B. Multilayer Perceptron (MLP) Implementation: Construct a hierarchical structure containing non-linear activation functions (such as ReLU, Tanh). For example: y = W2(σ(W1x + b1)) + b2, where σ represents the activation function.

[0076] The above methods provide flexibility in choosing model complexity. For simple feature transformations, computationally inexpensive linear projections can be used; for complex features requiring the extraction of higher-order implicit associations, MLPs can be used for deep abstraction.

[0077] (2) The mapping function is used to project the features of medical institutions and transportation vehicles with different feature dimensions into a unified feature space with the same dimension, so as to meet the input dimension requirements of the Transformer policy network.

[0078] The above steps are key preprocessing steps to meet the input requirements of the subsequent Transformer network.

[0079] The aforementioned "different feature dimensions" mean that medical institutions extract 7 types of features (such as grade, number of beds, etc.), while transportation vehicles extract 4 types of features plus the embedding of their affiliated institution (such as location, capacity, etc.). This implies that their original feature vector lengths (Input Dimension) are usually not equal. For example, the vector length of a medical institution might be 10, while the vector length of a transportation vehicle might be 18.

[0080] The aforementioned "unified feature space" means that the attention mechanism of the Transformer model requires all element vectors in the input sequence to have the same dimension (usually denoted as d). model (e.g., 128 or 512 dimensions), otherwise matrix dot product operations cannot be performed.

[0081] The specific processing procedure can be as follows: input the original feature vector (generated in the aforementioned implementation) into the mapping function constructed above for calculation. The input dimension is d. in1 Characteristics of medical institutions V hosp After the mapping function φ m The output dimension is d target vector e mj The input dimension is d. in2 Transportation characteristics V tool After the mapping function φ u The output dimension is also d. target vector e ui This results in two vector sets: the medical institution embedding vector set e. mj and the vehicle embedding vector set e uiEach vector in these two sets, although representing different physical objects, has the same mathematical length (dimension).

[0082] This step eliminates the dimensional differences between heterogeneous data, allowing "medical institutions" and "transportation vehicles" to be treated as tokens of equal status and mixed into the same input sequence; and enabling subsequent steps to use the Transformer's self-attention mechanism to calculate the correlation strength between transportation vehicles and medical institutions (cross-relationship) and between medical institutions (competitive relationship).

[0083] For example, the Transformer requires an input dimension of 64. The original feature vector length of the medical institution is 15. The mapping function φ... m It's essentially a fully connected layer that transforms 15-dimensional data into 64-dimensional data. The original feature vector length of the transportation vehicle is 20. The mapping function φ u It's essentially a fully connected layer that transforms 20-dimensional data into 64-dimensional data. After processing, both become 64-dimensional vectors, which can be concatenated into a shape of (N... agents +M hospitals A 64-fold matrix is ​​input into the network.

[0084] Example 3 Based on the foregoing embodiments, this embodiment provides a method for decision-making regarding the evacuation of wounded personnel. In step S400, the initial attention preference matrix is ​​corrected using a distance attenuation term, as shown in the following expression (Formula 1): ; in, This represents the revised attention score; The above This represents the physical distance between the current location or affiliated institution of transportation vehicle i and the target medical institution j. This is a data acquisition and calculation step. The system needs to acquire map data, locate the coordinates of points i and j, and combine this with road network information to calculate the actual travel distance between the two, rather than a simple straight-line distance. The shortest path length between the two points can be calculated using path planning algorithms in Geographic Information System (GIS) (such as Dijkstra's algorithm or A* algorithm).

[0085] The above Represents the distance decay function; This is a mathematical function used to calculate the physical distance d. i,j This is mapped to a numerical "attenuation" or "cost value." Its purpose is to quantify the physical law that "the greater the distance, the greater the cost." Specifically, this involves calculating the distance d... i,jThe function is input as its independent variable, and its value is calculated. For example, it could be: Linear decay: Distance directly serves as the cost.

[0086] Exponential decay: As the distance increases, the cost increases dramatically.

[0087] Embedding attenuation: Divide the distance into segments, such as 0~5km, 5~10km, etc., with each segment corresponding to a learnable weight value.

[0088] The above This represents the distance penalty coefficient; it's a scalar weighting parameter (hyperparameter or learnable parameter) used to control the proportion of distance factors in the overall decision weights. It acts as a regulating valve. If... The large size indicates that the system highly values ​​distance costs and tends to deliver to the nearest location; if The small value indicates that the system places greater emphasis on the matching degree of medical resources (from the initial A). i,j (Decision), allowing for a certain degree of roundabout route selection. The specific processing involves adjusting the output value of the distance decay function. Multiply by a coefficient .

[0089] The above This represents the environmental constraint term, which quantifies the impact of external environmental factors (such as road conditions and weather) on a specific transportation route, in addition to distance. It represents a bias value for a specific route (i, j). If the road is damaged or the weather is bad, this term may be negative (lowering the overall score); if the road is clear and safe, this term may be positive or zero. This allows the decision to be based not only on "distance" but also on "road conditions." Specifically, an environmental matrix can be constructed, and the corresponding route can be obtained directly by looking up a table or calculating based on real-time information (such as road congestion coefficients and weather levels). value.

[0090] The above correction calculation This is the final matrix operation. Its input is the initial attention score A calculated by the Transformer in the aforementioned implementation. i,j (This primarily reflects the matching strategy between injury severity and hospital capabilities). Specifically, for each element in the matrix, subtract the distance penalty term and add the environmental constraint term to obtain the corrected attention preference matrix (A on the left side of the formula). i,j ).

[0091] The algorithm described above eliminates the need for additional encoding sequences, explicitly considering the "edge" information (distance and environment) between points through mathematical corrections. Furthermore, it allows the attention weights for long-distance pairings to naturally decay with increasing distance, forcing the model-generated scheduling scheme to be physically more reasonable and efficient. Specifically, element-wise operations on matrices can be employed. For example, in Python / PyTorch, attention tensors and distance tensors can be directly added or subtracted.

[0092] In some implementations, the distance decay function may take the form of any one of linear decay, exponential decay, and embedding decay.

[0093] This method can limit the specific type of mathematical model used to calculate the distance cost. Among them, the distance decay function f... dist The distance d between the transportation vehicle and the target medical facility. i,j The attenuation term is converted into a numerical value. Based on preset configurations, the system can select a mathematical function to process the input distance data. (1) Linear attenuation: Establish a linear mapping relationship between distance and attenuation value.

[0094] (2) Exponential decay: Establish an exponential mapping relationship between distance and decay value.

[0095] (3) Embedding decay: Establish the mapping relationship between distance interval and learnable weights.

[0096] In some implementations, the environmental constraints are used to quantify the impact of road conditions and weather conditions on the transportation process.

[0097] This method clarifies the physical meaning and input source of the environmental constraint term. It serves as a correction bias, specifically designed to characterize the impact of external environmental factors, other than distance, on the feasibility or cost of a particular transportation route. This allows the decision-making model to not only be based on geographical distance but also to perceive dynamic environmental changes. For example, even if two points are very close, if the road is impassable or the weather does not permit flight, this constraint term will significantly reduce its attention score, thereby avoiding the generation of unexecutable scheduling instructions.

[0098] In some implementations, when the distance decay function is an embedded decay form, the distance is divided into several discrete intervals, and a corresponding trainable scalar weight is established for each discrete interval as the function output; or, when the distance decay function is an exponential decay form, the function value of the distance decay function increases exponentially with the increase of the distance.

[0099] The distance decay function is an embedded decay form, which is a non-linear processing method based on deep learning. It does not assume that there is a fixed mathematical formula (such as linear or exponential) between distance and cost, but instead allows the model to learn the "cost weights" for different distance segments through data.

[0100] Specifically, continuous distance values ​​di,j can be mapped to corresponding interval indices. For example, defining intervals [0-5km, 5-10km, 10-20km...], if the distance is 7km, it falls into the second interval. Based on the interval index, the corresponding scalar weight value is looked up in a trainable parameter table, thus outputting a data-driven learned decay value.

[0101] This method offers exceptional flexibility. The model can learn non-monotonic or complex distance preference patterns, enabling it to more precisely fit empirical patterns in actual transportation processes.

[0102] For example, an embedding layer in a neural network framework can be used. The input is the distance interval ID, and the output is a scalar with dimension 1. The parameters of this embedding layer are trained together with the policy network.

[0103] The distance decay function is in exponential decay form, defining the numerical change law under exponential decay mode. Although named a "decay function," it refers to the fact that the "cost" or "penalty term" of the function output increases exponentially with increasing distance. The processing involves taking the input distance d... i,j An exponential operation is performed to obtain a penalty value that increases dramatically with increasing distance.

[0104] Because the formula is When f dist When the penalty increases exponentially, the final attention score A will decrease exponentially (decay).

[0105] This processing method exerts a strong inhibitory effect on long-distance transportation. This helps to strictly limit long-distance repatriation in decision-making, forcing the system to prioritize medical institutions at very close range, and is suitable for environments with extremely high timeliness requirements or extremely high risks associated with long-distance transportation.

[0106] Example 4 refer to Figure 4 Based on the foregoing embodiments, this embodiment provides a method for decision-making regarding the evacuation of wounded personnel. Step S500, which involves decoding actions based on the modified attention preference matrix, includes: Step S510: Treat the modified attention preference matrix as a soft policy distribution.

[0107] This step defines how the input data is understood. The corrected attention preference matrix P i,jEach value in the equation is no longer just an ordinary number, but is interpreted as a "probability" or "inclination." It represents the strength or priority with which a particular mode of transport i chooses a particular medical institution j.

[0108] Specifically, the system can avoid hard truncation (i.e., either zero or 1) of the values ​​in the matrix, instead preserving their continuous numerical characteristics. This typically involves normalization operations (such as using the Softmax function) to ensure that the sum of the preference values ​​for all possible target institutions for a given mode of transport is 1, thus forming a probability distribution. This allows for the creation of a probability distribution map or scoring table describing transport priorities, reflecting the relative suitability of pairing each mode of transport with each medical institution in the current state.

[0109] This method preserves the "grayscale" information of the decision. Compared to directly outputting a single "best" result, soft-policy distribution allows the system to know which choices are "best" and which are "second best." This provides leeway for resolving resource conflicts later (when the best option is occupied, it can smoothly degenerate to the second best option).

[0110] Step S520: Based on the soft strategy distribution, and under the conditions of satisfying transportation capacity, injury matching, agency reception constraints and time window constraints, feasible transportation tasks are selected sequentially from high preference values ​​to low preference values ​​to construct joint scheduling actions for transportation vehicles.

[0111] This step is a search process based on a greedy strategy combined with rule constraints. Its purpose is to solve the resource competition problem among multiple objectives and ensure that the generated solution is physically feasible.

[0112] The constraints involved include the four mandatory conditions that must be met as explicitly listed in the aforementioned implementation method: Transport capacity: The remaining passenger space in the transport vehicle must be sufficient to accommodate the wounded to be transported.

[0113] Injury matching: The target medical institution must have the capability to treat the specific injury of the patient (such as traumatic brain injury, burns, etc.).

[0114] Institutional acceptance constraints: The target medical institution must have sufficient available beds or supplies to receive new wounded.

[0115] Time window constraint: The estimated arrival time must meet the emergency medical time window requirements of the injured or the mission cycle limit of the vehicle.

[0116] Specifically, the process begins with sorting: all elements in the attention preference matrix (representing all possible "vehicle-hospital" pairings) are ranked from highest to lowest numerical value. Then, the top-ranked pairings are checked sequentially. For each pairing, it is determined whether it satisfies the four constraints mentioned above. If a pairing satisfies all conditions, it is selected as a valid task, and corresponding resources are deducted (e.g., reducing vehicle vacancies, reducing hospital beds). If resources are insufficient or conditions are not met, the pairing is skipped, and the next option in the list is considered, thus generating a joint scheduling action for the transportation vehicles (a t This is a specific set of instructions that clearly specifies which wounded should be carried on each means of transport and which medical facility they should be sent to.

[0117] Through strict constraint checks, ineffective scheduling such as "overloading," "sending to the wrong hospital," or "hospital overcrowding" is eliminated. By utilizing a "high-to-low" greedy mechanism, the resource competition problem is naturally solved: tasks with high priority (high preference value) are given priority to obtain scarce resources (such as the only ICU bed), thereby maximizing the efficiency of the overall system.

[0118] For example, a greedy search algorithm with a mask can be used. At each selection step, the resource state vector is dynamically updated, and for nodes that have exhausted their resources, their corresponding masks are invalidated to prevent repeated selections.

[0119] Example 5 refer to Figure 5 Based on the foregoing embodiments, this embodiment provides a method for decision-making regarding the evacuation of wounded personnel. Step S300, which involves constructing a policy network based on a Transformer structure, further includes: Step S310: Introduce a value identifier token into the input sequence.

[0120] This step refers to adding a special token to the input data of the Transformer, in addition to the regular "medical institution embedding vector" and "transportation vehicle embedding vector," to represent global information. This is similar to the [CLS] token in natural language processing models such as BERT.

[0121] Specifically, before feeding the feature vector sequence into the Transformer model, a learnable special vector (i.e., [VALUE] token) can be inserted at the beginning or a specific position of the sequence to form a new input sequence, which is composed of "value identifier token", "medical institution vector set" and "transportation vehicle vector set".

[0122] This special token serves as a hub for global information. Through the Transformer's self-attention mechanism, this token can interact with all medical institutions and transportation vehicles in the sequence, thereby "absorbing" and aggregating global state information, rather than being limited to the features of a single local node.

[0123] You can initialize a random vector with the same dimension as the embedding vector of the medical institution / transportation vehicle, and use it as a parameter of the model to train the network.

[0124] Step S320: A multilayer perceptron structure is connected outside the output vector of the value identifier token as a value network to fit the global state value.

[0125] This step defines how to use the output information of the aforementioned special token to construct a value assessment function. It constructs a branch network (value network) independent of the action policy output.

[0126] In terms of processing, after the input sequence undergoes multi-layer self-attention computation by a Transformer, the final output vector corresponding to the "value identifier token" position is extracted. This output vector is then input into a multilayer perceptron (MLP) structure for computation. The MLP then maps the high-dimensional vector to a scalar value, thus outputting a numerical value representing the "global state value." This value is a prediction and quantification of the long-term gains of the current situation.

[0127] This step simultaneously evaluates the long-term cumulative reward of the global state, determining whether the current situation is "favorable" or "unfavorable," providing a "benchmark signal" for reinforcement learning strategies (such as the PPO algorithm). By comparing the actual reward with the predicted value, the advantage function is calculated, thereby reducing training variance and accelerating the convergence of the policy network.

[0128] Specifically, an MLP consisting of fully connected layers and activation functions can be constructed, with its input dimension being the same as the hidden layer dimension of the Transformer (e.g., 512) and its output dimension being 1 (representing a value scalar).

[0129] In some embodiments, the casualty evacuation decision-making method further includes: step S600, training the policy network; Step S600, training the policy network, includes: Step S610: The PPO algorithm is optimized using a near-end strategy, and the strategy network parameters are updated based on the reward signal from environmental feedback; wherein, the reward signal is calculated based on a comprehensive calculation of transfer time, resource utilization rate, institutional load balancing, success rate of wounded personnel treatment, and delay penalty.

[0130] The phrase "updating policy network parameters using the proximal policy optimization (PPO) algorithm" specifies a particular type of deep reinforcement learning algorithm. PPO (Proximal Policy Optimization) is an algorithm based on policy gradients. Its expression (Equation 2) can be: .

[0131] in, This represents the long-term cumulative reward (or objective function value) that needs to be maximized. E represents the trainable parameters of the policy network. t This represents the expectation at time step t; This represents the reward decay coefficient, with a value range of [value range missing]. ∈(0,1)$, used to weigh long-term and short-term returns; t represents the time exponent or time step; This represents the decision feedback information (i.e., reward) during the state transition process. This value comprehensively considers transfer time, resource utilization, institutional load balancing, success rate of wounded personnel treatment, and penalties for delays.

[0132] "Update parameters" refers to adjusting the values ​​of the connection weights in the neural network so that, when faced with the same state input, the network has a higher probability of outputting a "good action" and a lower probability of outputting a "bad action". The process involves the policy network interacting with the simulation environment to collect a batch of data samples (trajectories) composed of "state, action, and reward"; calculating the loss function using PPO's unique "clipping" mechanism; calculating the ratio of the new policy to the old policy and limiting the range of this ratio to prevent policy collapse due to excessively large update steps; and updating the parameters of the Transformer network and the value network (MLP) using gradient descent based on the calculated loss, thereby obtaining an optimized policy network model that can output high-value scheduling actions according to the environmental state.

[0133] Compared to other algorithms (such as DQN or the original policy gradient), PPO ensures smooth convergence during training by limiting the magnitude of policy updates, making it less prone to significant performance fluctuations. It also achieves a balance between sample efficiency and data utilization, ensuring relatively simple algorithm implementation while providing good data utilization efficiency, making it suitable for handling scheduling problems with complex action spaces.

[0134] Specifically, the PPO algorithm body can be built using existing deep learning frameworks (such as PyTorch or TensorFlow), hyperparameters (such as pruning threshold, learning rate, etc.) can be set, and a "sampling-update" loop can be executed periodically.

[0135] The aforementioned "reward signal based on environmental feedback" is the core driving force of reinforcement learning. This reward signal is a numerical value (scalar) that acts as a "teacher," telling the model whether the series of scheduling actions were "right" or "wrong," and how "good" they were. The process involves the environment calculating a score based on the execution result after the model outputs a joint scheduling action and it is executed (or simulated in a simulation). If the score is positive (reward), the model reinforces the tendency to perform that action; if the score is negative (penalty), the model suppresses the tendency to perform that action, thus providing the algorithm with optimization direction guidance (i.e., gradient direction).

[0136] The aforementioned "Comprehensive Calculation of Reward Signals" specifies that reward signals must be comprehensively calculated based on the following five specific dimensions (usually a weighted sum). This reflects that medical evacuation is a typical "multi-objective optimization problem".

[0137] (1) Based on transfer time: refers to the total time it takes for an injured person to be transported from the point of injury or current institution to the target medical institution. It is usually used as a negative reward (penalty). The shorter the time, the smaller the penalty (or the larger the reward). Its purpose is to encourage the model to find the shortest and fastest path, reflecting "timeliness".

[0138] (2) Based on resource utilization rate: This refers to the occupancy rate of transportation vehicles or the use of beds / supplies in medical institutions. It is usually used as a positive reward. For example, if an ambulance that can seat 4 people departs with only 1 person on board, the resource utilization rate is low and the reward value is low; if it departs fully loaded, the reward value is high. The purpose is to avoid wasting transportation capacity and encourage the model to carpool or carry out intensive transportation.

[0139] (3) Based on institutional load balancing: This refers to the degree of difference in the number of injured people received by different medical institutions. Processing: Usually, the variance or standard deviation of the load of each institution is calculated as a negative reward. If all hospitals have similar loads, the variance is small and the penalty is small; if one hospital is overloaded while another is idle, the variance is large and the penalty is large. The purpose is to prevent the medical resources of a specific node from being squeezed (congested) due to blindly sending patients to the nearest hospital, and to achieve network-wide traffic balance.

[0140] (4) Based on the success rate of wounded patient treatment: This is the ultimate medical indicator. The probability of successful treatment is estimated based on a model relating the severity of the injury (e.g., serious or minor) to the time of arrival and the hospital's level. The higher the probability, the greater the positive reward. The purpose is to reflect the principle of "life first." For example, sending a seriously injured patient to a hospital with advanced surgical capabilities (even if it's a bit far) may result in a higher success rate reward than sending them to the nearest clinic.

[0141] (5) Delay-based penalties: Additional heavy penalties for cases exceeding a specific "time window" (such as the golden hour for treatment). A threshold T can be set. max If the transit time t > T max Then a huge negative value is applied. The purpose is to force the model to avoid the risk of timeout and ensure that critically injured patients receive treatment within a critical time.

[0142] For example, the five indicators mentioned above are combined into a total reward R using a weighted formula (Formula 3). t : R t = w1·(Treatment success rate) + w2·(Utilization rate) - w3·(Transfer time) - w4·(Load variance) - w5·(Delay penalty); Here, w1~w5 are adjustable weight coefficients used to balance the importance of each objective. The PPO algorithm will directly maximize this R. t The expected value.

[0143] In some implementations, in step S300, the step of constructing a policy network based on the Transformer structure, before calculating the initial attention preference matrix through the attention mechanism, the following steps are also included: Step S330: Combine the medical institution embedding vector and the transportation vehicle embedding vector to form an input sequence.

[0144] This step refers to merging the two sets of feature vectors obtained independently in the previous steps into a unified whole.

[0145] Although medical facilities and transportation vehicles are physically distinct entities, in the input layer of a neural network, they need to be treated as different nodes in the same sequence in order to perform unified operations.

[0146] The specific processing can involve performing a union operation on sets or a concatenation operation on sequences. Assume there is a vector set e of N transportation vehicles. ui and the vector set e of M medical institutions mj By connecting them end to end or arranging them in a specific order, a mixed sequence of length N+M is formed, thus obtaining a joint input sequence containing all the current key interactive entities.

[0147] This step breaks down the barriers between different types of objects, enabling the subsequent attention mechanism to not only calculate the relationship between "vehicles" or "institutes", but also to directly calculate the core pairing relationship between "vehicles and institutes". It provides the model with a complete situational snapshot, allowing decisions to be based on global information rather than local information.

[0148] For example, in tensor operations, concatenation operations are used. For instance, concatenating a transportation tensor of shape (N, d) and a medical institution tensor of shape (M, d) along the 0th dimension yields an input tensor of shape (N+M, d).

[0149] Step S340: Type encoding and position encoding are superimposed on each element in the input sequence to distinguish the category attributes and sequence position information of medical institutions and transportation vehicles.

[0150] Regarding the "overlaying type encoding on each element in the input sequence," it's important to note that since the vectors for medical institutions and transportation vehicles are projected onto the same dimension, the Transformer model might struggle to distinguish whether a vector represents a "vehicle" or a "hospital" based solely on numerical values. Type encoding assigns a clear "category identity" to each vector.

[0151] A learnable type vector v can be defined for "vehicle". type_car Define another type vector v for "medical institutions". type_hosp The corresponding type vectors are element-wise added to their respective embedding vectors, thus implicitly containing entity attribute information in the input vector.

[0152] This step helps the model distinguish between different role functions. For example, the model will learn that a node belonging to the "Transportation Vehicle" type needs to find a node of the "Medical Institution" type as its target, rather than finding another "Transportation Vehicle".

[0153] Regarding the "overlaying positional encoding on each element in the input sequence," it's important to note that the Transformer's self-attention mechanism operates in parallel during computation and inherently cannot recognize the sequential order of elements within the sequence. Positional encoding is used to provide the model with the index or relative position information of each element within the sequence.

[0154] Based on the index position of the element in the input sequence (such as the 1st, 2nd, etc.), the corresponding position vector can be generated and superimposed (added) onto the embedding vector to form the final input vector ready to be sent to the Transformer layer.

[0155] The expression (Formula 4) can be represented as: Input=Embedding+TypeEncoding+PosEncoding.

[0156] Here, Input represents the final input sequence fed into the Transformer model. It is a comprehensive vector sequence formed by superimposing type and location information onto the basic feature vectors, used for subsequent multi-head self-attention calculations. Embedding represents the feature embedding vectors (i.e., the medical institution embedding vector and the transportation vehicle embedding vector). This is based on the medical institution's state information (e... mj ) and vehicle status information (e ui The original vector representation obtained after feature extraction and mapping. `TypeEncoding` represents the type encoding. Since the input sequence mixes two different object categories, "medical institution" and "transportation vehicle," this variable is used to distinguish whether the current element belongs to the "medical institution" or "transportation vehicle" category. `PosEncoding` represents the position encoding. Because the Transformer's self-attention mechanism itself does not have sequence order awareness, this variable is used to label the specific position information of each element in the input sequence to preserve the structural features of the sequence.

[0157] Even if two vehicles have extremely similar states (closer vector values), the location encoding based on this method can ensure that they are distinct and independent entities in the sequence. Preserving sequence structure: This allows the model to process the structural information of sequence data. Although the relative significance of location information may be weaker in ensemble optimization problems than in NLP tasks, it is still necessary for distinguishing entities with different IDs.

[0158] Specifically, position codes can be generated using fixed sine / cosine functions, or learnable position embeddings can be used.

[0159] In some implementations, step S300, calculating an initial attention preference matrix using an attention mechanism, includes: Step S350: Context modeling is performed on the input sequence using a multi-head self-attention mechanism to calculate an attention score that reflects the dependency between the transportation vehicle and the medical institution.

[0160] This step provides the core perception process of the policy network.

[0161] The aforementioned "context modeling" refers to the model's approach to understanding a specific node (such as an ambulance) not by viewing it in isolation, but by considering the states of all other nodes in the entire environmental sequence (such as all hospitals and other ambulances).

[0162] The term "multi-head" refers to the model analyzing this relationship simultaneously from multiple different feature subspaces (perspectives).

[0163] The specific processing can be as follows: First, the input sequence (a mixed vector containing position and type encodings) is mapped to a query vector (Query, Q), a key vector (Key, K), and a value vector (Value, V) through different linear projection layers; then, matrix multiplication (dot product) is performed using Q and K. This step calculates the matching degree between the "query" and the "features"; then, the above process is repeated multiple times (i.e., multiple "heads"), each head focusing on different feature dimensions (e.g., some heads focus on matching geographical locations, while others focus on matching injury severity with hospital level), and then the calculation results from multiple heads are concatenated and fused to obtain the "attention score".

[0164] This step involves an unnormalized numerical matrix, where the magnitude of the values ​​directly reflects the strength of the original association or dependence between transportation vehicle i and medical institution j at the feature level.

[0165] Its advantage lies in its ability to capture long-distance dependencies. No matter how far apart the vehicle and the hospital are in the sequence, the model can calculate the association between them. Through the "multi-head" mechanism, the model can simultaneously take into account a variety of complex matching logics (such as being fast and able to treat serious injuries), avoiding the one-sidedness of a single perspective.

[0166] Specifically, the standard scaled dot product attention formula (Formula 5) can be used: ; Where Q represents the query vector, which is one of the vectors obtained by mapping the input sequence through a linear projection layer; K represents the key vector, which is also one of the vectors obtained by mapping the input sequence through a linear projection layer. This represents the degree of match between the "query" and the "feature". k This represents the scaling factor, which is used to prevent gradient vanishing. This represents the calculated attention score. It is an unnormalized numerical matrix reflecting the strength of the association between transportation vehicles and medical institutions at the feature level.

[0167] Step S360: Normalize the attention score and output the initial attention preference matrix.

[0168] This step transforms the raw scores (Logits) calculated above into a numerical distribution with probabilistic or relative weighting significance.

[0169] Specifically, mathematical transformation functions can be applied to the attention scoring matrix. This is typically done on the "target medical institution" dimension, ensuring that for each mode of transport, the processed attention scores for all available medical institutions fall within a specific range (usually 0 to 1) and are comparable, thus outputting an "initial attention preference matrix".

[0170] The element P in the matrix i,j This represents the degree to which transportation vehicle i recommends medical institution j, considering only feature matching (before considering distance penalty).

[0171] Its advantages are that it prevents the model training from being unstable due to the original score being too large or too small; it converts absolute values ​​into relative preferences, which makes it easier to add "distance decay term" for correction and "action decoding" later.

[0172] Specifically, the Softmax function (Formula 6) can be used for processing: ; This ensures that the sum of each row of the output matrix (representing the preference distribution of a mode of transport) is 1.

[0173] In some embodiments, after generating the joint scheduling action of the transportation vehicles in step S600, the method further includes: Step S610: Perform the joint scheduling action of the transportation vehicle; and update the load information in the status information of the medical institution, the availability information in the status information of the transportation vehicle, and the road condition information in the environmental status information based on the result of the joint scheduling action.

[0174] The aforementioned step of "executing the joint scheduling actions of the transport vehicles" represents the transition from the "decision planning" stage to the "actual operation" stage. The "joint scheduling actions" generated in the previous steps were merely calculated plans, while this step involves putting those plans into practice.

[0175] The specific system can issue the generated scheduling instructions to the corresponding execution units. In practical applications, this manifests as sending specific instructions (such as "Vehicle A proceeds to coordinate X to load the wounded Y and transport him to hospital Z") to the communication terminals of each transport vehicle. In a simulation training environment (such as a reinforcement learning training process), this manifests as the simulation engine receiving the action instruction and driving virtual entities to move and perform operations digitally, thereby causing the transport vehicle to start moving, the wounded to be loaded or unloaded, and the physical state to begin to change.

[0176] This step transforms the calculation results into actual effectiveness, which is a necessary step for the decision-making system to generate real value.

[0177] For example, command data packets can be sent through the communication interface of the command and control system, or the env.step(action) function can be called in the simulation program.

[0178] The step described above, "updating information on medical facility load, transportation availability, and road conditions based on the execution results," is used to capture the chain reaction caused by the action and to update the environmental state synchronously. This embodies the "state transition" characteristic of a Markov Decision Process (MDP).

[0179] Specifically, the system can modify the status values ​​in the database or memory based on the objective facts after the action is executed: (1) Update the medical facility load: If a hospital receives new wounded, its "current number of wounded" should be increased by 1 and its "number of available beds" should be decreased by 1.

[0180] (2) Update vehicle availability: If a vehicle receives a task, mark its status as “occupied” or “in transit” and lock its availability before it reaches its destination; if a vehicle completes a task, reset it to “idle”.

[0181] (3) Update road conditions: Update the traffic parameters of the road network based on the latest intelligence or simulation results (such as a road section becoming impassable due to being bombed, or becoming congested due to increased traffic volume).

[0182] Through the above processing, a set of latest data that truly reflects the current situation (after the action is executed) can be obtained.

[0183] Its advantages are that it prevents the decision-making system from making incorrect follow-up instructions based on outdated information (such as thinking that a vehicle is still available when it has actually already departed); and it avoids conflicts when multiple vehicles simultaneously send patients to hospitals that are already full by deducting bed and vehicle resources in real time.

[0184] Step S620: The updated status information of the medical institution, the status information of the transportation vehicle, and the road condition information are used as the environmental status information for the next moment. Then, the process of performing feature vectorization on the status information of the medical institution and the status information of the transportation vehicle is returned to perform the next moment's evacuation decision.

[0185] This step defines the cyclic logic for rolling time-domain decision-making. It indicates that the method is not a one-time static programming, but a continuous decision-making process that continues over time.

[0186] Specifically, the time can be shifted from t to t+1. The updated state is defined as the new "current state" S. t+1 The program control flow jumps back to the "feature vectorization processing" step, utilizing the new state S. t+1A new round of computation is started (embedding -> Transformer -> attention correction -> decoding), thus forming a continuously running decision-making loop that constantly outputs scheduling actions for future moments.

[0187] Its advantage lies in the fact that the environmental situation is constantly changing (such as a sudden large number of casualties or a sudden road closure). This cyclical mechanism enables the system to dynamically adjust subsequent scheduling strategies based on the latest situation. Combined with the characteristics of reinforcement learning, this multi-step decision-making can pursue the maximization of cumulative benefits throughout the entire task cycle, rather than just immediate benefits.

[0188] For example, in algorithm design, a while not done: loop structure can be used to wrap the entire decision-making process until all the wounded have been treated or the mission time has ended.

[0189] This application embodiment also provides a casualty evacuation decision-making device, including: The acquisition module 10 is used to acquire the environmental status information at the current moment; the environmental status information includes at least the status information of the medical institution and the status information of the transportation vehicle; Vector module 20 is used to perform feature vectorization processing on the state information of the medical institution and the state information of the transportation vehicle to obtain the medical institution embedding vector and the transportation vehicle embedding vector, respectively. The calculation module 30 is used to construct a policy network based on the Transformer structure. It takes the embedding vectors of the medical institutions and the embedding vectors of the transportation vehicles as input sequences and calculates an initial attention preference matrix through an attention mechanism. The attention preference matrix is ​​used to characterize the transfer priority relationship between the transportation vehicles and the medical institutions. The correction module 40 is used to correct the initial attention preference matrix using a distance attenuation term to obtain a corrected attention preference matrix; wherein the distance attenuation term is calculated based on the distance between the vehicle and the target medical institution. The decoding module 50 is used to perform action decoding based on the corrected attention preference matrix to generate joint scheduling actions for transportation vehicles.

[0190] It is understood that the device in this embodiment corresponds to the casualty evacuation decision method in the above embodiment, and the options in the above embodiment are also applicable to this embodiment, so they will not be described again here.

[0191] The present invention provides a computer device, the computer device including a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the casualty evacuation decision-making method as described in any of the foregoing embodiments.

[0192] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0193] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving execution instructions.

[0194] This application also provides a computer storage medium storing a computer program, which, when executed on a processor, implements the casualty evacuation decision-making method according to any one of the foregoing embodiments.

[0195] The computer storage medium can be a readable storage medium, a non-volatile storage medium, or a volatile storage medium. For example, the computer storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0196] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0197] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0198] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0199] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for making decisions on the evacuation of wounded soldiers, characterized in that, Obtain the current environmental status information; the environmental status information includes at least the status information of the medical institution and the status information of the transportation vehicle; The status information of the medical institution and the status information of the transportation vehicle are respectively processed into feature vectors to obtain the embedding vectors of the medical institution and the transportation vehicle. A policy network based on the Transformer structure is constructed. The embedding vectors of the medical institutions and the transportation vehicles are used as input sequences. An initial attention preference matrix is ​​calculated through an attention mechanism. The attention preference matrix is ​​used to represent the transfer priority relationship between the transportation vehicles and the medical institutions. The initial attention preference matrix is ​​corrected using a distance attenuation term to obtain a corrected attention preference matrix; wherein the distance attenuation term is calculated based on the distance between the vehicle and the target medical institution. Action decoding is performed based on the revised attention preference matrix to generate joint scheduling actions for transportation vehicles.

2. The casualty evacuation decision-making method as described in claim 1, characterized in that, The step of performing feature vectorization processing on the status information of the medical institution and the status information of the transportation vehicle includes: For each medical institution, features such as institution level, geographical location, acceptable injury vector, number of beds, material reserves, inpatient load, and turnover time are extracted and used to obtain a vector representation of the medical institution through a mapping function; For each mode of transport, features such as transport capacity, current location, available injury distribution, and available time window are extracted and concatenated or fused with the embedding vector of its affiliated medical institution. The vector representation of the mode of transport is then obtained through a mapping function.

3. The casualty evacuation decision-making method as described in claim 2, characterized in that, The vector representation of the medical institution obtained through the mapping function, and the vector representation of the transportation vehicle obtained through the mapping function, include: Construct a learnable linear projection layer or multilayer perceptron as the mapping function; The mapping function is used to project the features of medical institutions and transportation vehicles with different feature dimensions into a unified feature space with the same dimension, so as to meet the input dimension requirements of the Transformer policy network.

4. The casualty evacuation decision-making method as described in claim 1, characterized in that, The initial attention preference matrix is ​​corrected using a distance decay term, as shown by the following expression: ； in, This represents the revised attention score; Represents environmental constraints; Represents the distance penalty coefficient; Represents the distance decay function; This represents the distance between the current location of the vehicle or its affiliated organization and the target medical facility.

5. The casualty evacuation decision-making method as described in claim 4, characterized in that, The distance decay function may be of any one of linear decay, exponential decay, and embedding decay; and / or, The environmental constraints are used to quantify the impact of road conditions and weather conditions on the transportation process.

6. The casualty evacuation decision-making method as described in claim 5, characterized in that, When the distance decay function is in the form of embedded decay, the distance is divided into several discrete intervals, and a corresponding trainable scalar weight is established for each discrete interval as the function output; or, when the distance decay function is in the form of exponential decay, the function value of the distance decay function increases exponentially with the increase of the distance.

7. The casualty evacuation decision-making method as described in claim 1, characterized in that, The step of decoding actions based on the modified attention preference matrix includes: The modified attention preference matrix is ​​treated as a soft policy distribution; Based on the aforementioned soft policy distribution, and under the conditions of satisfying transportation capacity, injury matching, agency reception constraints, and time window constraints, feasible transportation tasks are selected sequentially from high preference values ​​to low preference values ​​to construct joint scheduling actions for transportation vehicles; and / or, The construction of the policy network based on the Transformer structure also includes: A value identifier, token, is introduced into the input sequence; A multilayer perceptron structure is connected outside the output vector of the value identifier token as a value network to fit the global state value; and / or, The casualty evacuation decision-making method also includes the step of training the policy network; Training the policy network includes: The PPO algorithm is optimized using a near-end strategy, and the policy network parameters are updated based on the reward signal from environmental feedback. The reward signal is calculated based on a combination of factors including transfer time, resource utilization, institutional load balancing, success rate of casualty treatment, and delay penalty. In the step of constructing the policy network based on the Transformer structure, before calculating the initial attention preference matrix through the attention mechanism, the following steps are also included: The embedding vectors of the medical institutions and the transportation vehicles are combined to form an input sequence; Type encoding and location encoding are superimposed on each element in the input sequence to distinguish the category attributes and sequence location information of medical institutions and transportation vehicles; and / or, The initial attention preference matrix calculated through the attention mechanism includes: The input sequence is modeled using a multi-head self-attention mechanism to calculate an attention score that reflects the dependency between the transportation vehicle and the medical institution. The attention scores are normalized to output the initial attention preference matrix; and / or, Following the joint scheduling action of generating transportation vehicles, the following is also included: Perform joint scheduling actions for the transportation vehicles; and update the load information in the status information of the medical institution, the availability information in the status information of the transportation vehicles, and the road condition information in the environmental status information based on the results of the joint scheduling actions. The updated status information of the medical institution, the status information of the transportation vehicle, and the road condition information are used as the environmental status information for the next moment. The process of performing feature vectorization on the status information of the medical institution and the status information of the transportation vehicle is then returned to perform the next moment's evacuation decision.

8. A decision-making device for evacuating wounded soldiers, characterized in that, include: The acquisition module is used to acquire the current environmental status information. The environmental status information includes at least the status information of medical institutions and the status information of transportation vehicles; The vector module is used to perform feature vectorization processing on the state information of the medical institution and the state information of the transportation vehicle to obtain the embedding vector of the medical institution and the embedding vector of the transportation vehicle. The computation module is used to construct a policy network based on the Transformer structure. It takes the embedding vectors of the medical institutions and the embedding vectors of the transportation vehicles as input sequences and calculates an initial attention preference matrix through an attention mechanism. The attention preference matrix is ​​used to characterize the transfer priority relationship between the transportation vehicles and the medical institutions. The correction module is used to correct the initial attention preference matrix using a distance attenuation term to obtain a corrected attention preference matrix; wherein the distance attenuation term is calculated based on the distance between the vehicle and the target medical institution. The decoding module is used to decode actions based on the corrected attention preference matrix and generate joint scheduling actions for transportation vehicles.

9. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the casualty evacuation decision-making method according to any one of claims 1-7.

10. A computer storage medium, characterized in that, It stores a computer program that, when executed on a processor, implements the casualty evacuation decision-making method according to any one of claims 1-7.

Images