Emergency rescue method for heaven and earth cooperative mountain torrent disasters

Abstract

The invention relates to the technical field of mountain torrent disaster emergency rescue, and particularly discloses a sky-ground collaborative mountain torrent disaster emergency rescue method, which comprises the following steps: deploying and training the following emergency rescue key technologies based on a federal learning framework; the method comprises the following steps: firstly, proposing an unmanned aerial vehicle disaster area cluster wide-area optimal deployment algorithm based on a dual-target multi-arm tiger machine, and balancing the coverage range and communication efficiency of an unmanned aerial vehicle cluster; secondly, developing a space-ground collaborative two-stage semi-definite relaxation disaster area positioning method to ensure that the disaster area is accurately positioned; finally, an emergency rescue method based on unmanned aerial vehicle remote sensing vision and dynamic path optimization is provided, and the emergency rescue and evacuation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of emergency rescue for mountain torrent disasters, and in particular to a method for emergency rescue for mountain torrent disasters in coordination with heaven and earth. Background Art

[0002] Flash floods and secondary disasters have a wide range of impacts and spread rapidly, posing severe challenges to the coverage and response speed of drone clusters. In flash flood disasters, GPS signals are easily interfered with or even fail, making it difficult to locate the disaster area. Flash floods are accompanied by a large amount of clouds and heavy rainfall, which seriously affect the imaging quality of drone remote sensing, especially in small target identification tasks such as searching for disaster victims, where the degradation of imaging quality becomes particularly obvious.

[0003] Currently, data from different modalities (images, radars, sensor data, etc.) are integrated through advanced data fusion algorithms. For example, convolutional neural networks (CNNs) are used to combine remote sensing images with ground sensor data to achieve accurate analysis of the situation in the disaster area. However, due to the different resolutions, formats, and acquisition frequencies of different data sources, inconsistencies are prone to occur during the fusion process, affecting the refined analysis effect in complex scenarios. With the collaborative work of edge computing and cloud computing, existing technologies can achieve fast data transmission and processing between ground equipment and aerial equipment. This architecture allows some computing tasks to be performed on drones or ground terminals, reducing data transmission delays. Although cloud and edge collaborative computing improve processing efficiency, data transmission and processing will still be delayed in large-scale disaster scenarios or when the network is unstable, making it difficult to meet the real-time response requirements in emergency rescue. Summary of the invention

[0004] In response to the technical problems mentioned in the background technology, the present invention proposes a method for emergency rescue of mountain torrent disasters in coordination with heaven and earth.

[0005] The present invention adopts the following technical solution:

[0006] A method for emergency rescue of mountain torrent disasters in coordination with heaven and earth, comprising the following steps:

[0007] Based on the federated learning framework, the following key emergency rescue technologies are deployed and trained; first, in response to the problem that the flash flood disaster area is vast and difficult to respond in time, a wide-area optimal deployment algorithm for drone disaster area clusters based on a dual-target multi-armed tiger machine is proposed to balance the coverage range and communication efficiency of the drone cluster; secondly, in response to the dilemma of GPS signal failure under flash flood disasters, a two-stage semi-definite relaxation disaster area positioning method for space-ground collaboration is developed to ensure that the disaster area can be accurately located; finally, in order to overcome the problems of interference with remote sensing images, low accuracy in small target recognition, and difficulty in emergency rescue and retreat under flash flood disasters, an emergency rescue method based on drone remote sensing vision and dynamic path optimization is studied to improve the efficiency of emergency rescue and retreat.

[0008] As a further technical solution of the present invention, the federated learning framework realizes the collaborative learning of clusters by combining the feature extractor update mechanism guided by the affinity matrix and the feature aggregation mechanism driven by the maximum mean difference (MMD);

[0009] First, the federated learning framework constructs an affinity matrix to quantify the data similarity between different drones, and dynamically adjusts the matrix weights according to the performance of each drone's local model, so that drones with similar feature distributions obtain higher weights; based on the weights, the federated learning framework performs enhanced aggregation on the drone's local feature extractors, thereby improving the overall feature extraction capability of the cluster; then, the federated learning framework introduces MMD to measure the differences in feature distributions in different rounds, and performs weighted averaging of the features based on the differences to generate a comprehensive feature representation.

[0010] As a further technical solution of the present invention, a wide-area optimal deployment algorithm for drone disaster area clusters based on a dual-target multi-armed bandit is provided:

[0011] The first optimization goal is to maximize the coverage to achieve full coverage of the disaster area. This process is shown in Formula 1:

[0012] Formula 1:

[0013] where θ = {θ 1 ,θ 2 ,…,θ U} is the beamwidth group assigned to the drone, N i (θ i ) is the beam width θ of drone i i The number of deployed environmental information sensors covered by OL It refers to the number of times multiple drones cover the sensor simultaneously;

[0014] The second optimization goal is to minimize the maximum overall communication time to achieve the fastest deployment under full coverage of the disaster area. This process is shown in Formula 2:

[0015] Formula 2:

[0016] in is the overall communication time, is the average communication time, T i,G is the global update time, T i,L It is the local update time;

[0017] The present invention intends to use the Pareto UCB1 algorithm to solve the above dual-objective optimization problem. The Pareto UCB1 algorithm directly maximizes the reward vector in the reward space of the two objectives. In the initial stage, the Pareto UCB1 algorithm pulls each arm once to complete the initialization. In each round of iteration, the selection process of the Pareto optimal arm set A' is shown in Formula 3:

[0018] Formula 3:

[0019] in is the empirical mean reward of arm α, is the empirical mean reward of arm k, n α is the number of times arm α is pulled, n k is the number of times arm k is pulled, n is the total number of arms, Q is the target number, |A * | is the number of Pareto optimal arm sets. After finding the optimal arm set A', the algorithm randomly selects an arm from it to pull and updates its reward value.

[0020] Finally, the Pareto UCB1 algorithm evaluates its performance by calculating the Pareto regret.

[0021] As a further technical solution of the present invention, a two-stage semi-definite relaxation disaster area positioning method coordinated by the sky and the earth is as follows:

[0022] Using drones as relay nodes, the drones collaborate with the flash flood situation awareness sensors deployed on the ground to achieve reliable positioning. The drones move at a constant speed. Therefore, the modeling process of the instantaneous position of the mth UAV at time step k is shown in Formula 4:

[0023] Formula 4:

[0024] in is the starting position of the mth UAV, j = (k-1)M+m, m = 1, 2, ..., M, k = 1, ..., K;

[0025] Due to the change in the position of the anchor node and the high nonlinearity between the UAV ranging and the location of the disaster area, the complex nonlinear positioning problem can be transformed into a linear problem through the semidefinite relaxation technology, thereby ensuring that the positioning algorithm converges to the global optimal solution. The specific process is shown in Formula 5-6:

[0026] Formula 5:

[0027] Formula 6:

[0028] where s iis the position of the i-th ground sensor, s i T is the position of the ith sensor on the ground at time T, s i o is the initial position of the i-th ground sensor, is the receiver position, ε p is the additional noise, u o is the disaster area, d p is the measurement range of the propagation path,

[0029] In the proposed reliable positioning algorithm, P1 = [(x 1 ,y 1 ,z 1 ),(x 2 ,y 2 ,z 2 ),...,(x n ,y n ,z n )] represents the position set of other n common nodes calculated by the positioning algorithm when the anchor node position is completely accurate, P2 = [(x' 1 ,y' 1 ,z' 1 ),(x 2 ,y 2 ,z 2 ),...,(x n ,y n ,z n )] represents the position set of other n common nodes calculated by the positioning algorithm when the anchor node moves a small amount due to other factors (human movement, equipment damage). The location information of the sensors in the disaster area is obtained based on the proposed reliable positioning algorithm. By comparing the position changes of the sensors before and after the flash flood, the flash flood disaster area can be depicted.

[0030] As a further technical solution of the present invention, an emergency rescue method based on UAV remote sensing vision and dynamic path optimization:

[0031] Focusing on the difficulties in the emergency rescue process, the present invention plans to study an emergency rescue method based on UAV remote sensing vision and dynamic path optimization. This method combines anti-interference UAV remote sensing vision target recognition with a multi-target disaster area emergency rescue path optimization algorithm. The target type in the disaster area is obtained through the recognition algorithm, and the type and location of the target are determined in combination with the UAV position information. Based on the type and location of the target, the path optimization algorithm can formulate the optimal retreat route to ensure the efficiency and safety of the rescue operation;

[0032] The present invention designs a recognition model with Real Time Detection Transformer (RTDETR) as the main structure, combined with a soft threshold-residual network (Soft-Residual Network) and a cascaded attention (CGA) module, called SC-RTDETR. The remote sensing image of the disturbed disaster area is first filtered through a soft threshold-residual network to extract key features, and then input into an efficient hybrid encoder. The encoder combines a cross-scale feature fusion module with a cascaded attention module to ensure that feature information is fully integrated and focused at different scales. Finally, after precise query selection and overlap-based matching by the decoder, SC-RTDETR can accurately identify the affected houses and people.

[0033] SC-RTDETR eliminates interference in remote sensing images through a soft threshold-residual network module. First, the input feature map of the affected disaster area remote sensing image is extracted and enhanced through a standard residual block. The residual block directly passes the input to the output through a jump connection. At the same time, it is processed by multiple convolutional layers to enhance the feature expression ability. Next, the enhanced feature map dynamically adjusts the filter threshold according to the specific characteristics of each image signal. After the feature map is processed by soft threshold adaptive filtering, the injected interference signal is accurately identified and filtered. This adaptive filtering can ensure that the output feature map can remove interference to the greatest extent while maintaining important information. The principle is shown in Formula 7:

[0034] Formula 7: y=sign(F(x)+x)·max(|F(x)+x|-λ,0);

[0035] Where x is the interfered remote sensing image, λ is the threshold parameter, F(x) represents the nonlinear transformation in the residual block, and y is the remote sensing image with interference removed.

[0036] SC-RTDETR improves the accuracy of small target recognition through cascaded attention modules. Cascaded attention optimizes the model's attention mechanism through grouping and cascading strategies, thereby improving the target recognition ability in complex scenes. First, the input feature map is divided into multiple groups according to the spatial or channel dimension, and the feature map of each group is independently calculated for attention. This grouping mechanism ensures that the model can focus on fine-grained local features in the image and reduce the interference caused by global information. In each group, the self-attention mechanism is used to calculate the correlation between features within the group, thereby improving the expression ability of local features. Then, the attention results of each group are connected in series through a cascade mechanism and accumulated layer by layer. This cascade mechanism effectively integrates the attention information of different groups, allowing the model to form a more comprehensive feature representation in the global scope. Finally, the feature map that has been grouped and cascaded is output to the next layer of the network to obtain an enhanced feature representation. This structural design enables the cascaded attention module to reduce the computational complexity while enhancing the model's attention allocation to key areas. It is particularly suitable for small target recognition and target detection tasks in complex scenes. The principle of cascaded attention is shown in Formula 8:

[0037] Formula 8:

[0038] Where X L It is the feature representation of remote sensing images after multi-layer cascade attention mechanism, FC is the fully connected layer, is a linear transformation matrix, LayerNorm() is a normalization layer, and Y is the output target recognition result. The target recognition result is combined with the drone's own position information to obtain the information of safe points, disaster-stricken houses and disaster-stricken people, which is the key to planning emergency rescue routes in disaster areas.

[0039] As a further technical solution of the present invention, the emergency rescue method based on UAV remote sensing vision and dynamic path optimization also includes:

[0040] The safe point set is defined as S = {s 1 ,s 2 ,...,s m}, where m represents the number of safe points, s j represents the position of the jth safe point. The position of the victims is the most important consideration in path planning. The location set of the victims is P = {(p 1 ,N 1 ),(p 2 ,N 2 ),...,(p n ,N n )}, each location may have a different number of victims, where p i represents the location of the i-th disaster victim, N i Indicates that at position pi The number of disaster victims on the hazard-affected houses may contain unidentified disaster victims. The set of disaster-affected house locations is defined as H = {(h 1 ,a 1 ),(h 2 ,a 2 ),...,(h k ,a k )}, each element contains the location of the house and the number of possible unidentified disaster victims at that location;

[0041] In order to optimize the evacuation route, the objective function is designed to minimize the total time, and weight coefficients α and β are introduced to represent the relative priorities of disaster victims and houses, where the weight coefficient α is based on the number of disaster victims N. i Adjustment, while considering the risk factors of the path, the specific expression of the objective function is shown in Formula 9:

[0042] Formula 9:

[0043] Among them, x ij ·N i ·d ij Indicates that the victims are from position p i To a safe spot j The evacuation cost is based on the number of people affected N i and path distance d ij ,y ij ·a i ·d ij Indicates that from the affected houses i To a safe spot j Evacuation costs, based on the number of unidentified affected persons in potential homesa i and path distance d ij , w ij is the risk weight of the path;

[0044] Combined with the number of disaster victims, the path capacity constraint also needs to consider the number of disaster victims at each location. The constraint expression is shown in Formula 10:

[0045] Formula 10:

[0046] Where N i ·x ij Indicates that from position p i To a safe spot j Total number of disaster victims evacuated, a i ·y ij Indicates the location of the affected house h i To a safe spot j The number of unidentified disaster victims, C jrepresents the path carrying capacity;

[0047] According to the time and risk assessment of different paths, the multi-target disaster area emergency rescue path optimization algorithm optimizes the overall rescue route through phased decision-making to ensure the efficiency and safety of rescue operations under flash flood disasters.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] Federated learning framework: This framework combines the affinity matrix-guided feature extractor update mechanism with the maximum mean difference (MMD)-driven feature aggregation mechanism to achieve cluster collaborative learning and dynamically adjust the matrix weights to enhance the feature extraction capability of drones. By measuring the feature distribution differences through MMD, the framework generates a comprehensive feature representation, improves the generalization ability of the model and better adapts to complex flash flood disaster environments.

[0050] The optimal deployment algorithm for the wide-area disaster area cluster of drones based on the dual-objective multi-armed bandit: maximizes the coverage of the disaster area while minimizing the overall communication time to achieve the optimal deployment of the wide-area disaster area. It improves the response speed of the drone cluster and provides technical support for the rapid and accurate positioning of the drone in the disaster area.

[0051] Space-ground collaborative two-stage semi-definite relaxed disaster area positioning method: This reliable positioning technology can solve the difficulties of untimely flash flood monitoring and inaccurate disaster area positioning caused by the inability to normally apply positioning technologies such as GPS after a disaster occurs, while solving the misjudgment of flash floods.

[0052] Emergency rescue method based on UAV remote sensing vision and dynamic path optimization: This method combines anti-interference UAV remote sensing vision target recognition with a multi-target disaster area emergency rescue path optimization algorithm. The target type in the disaster area is obtained through the recognition algorithm, and the type and location of the target are determined in combination with the UAV position information. Based on this precise target information, the path optimization algorithm can formulate the optimal retreat route to ensure the efficiency and safety of the rescue operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The figure is a schematic diagram of the technical principle of a method for emergency rescue of mountain torrent disasters in coordination between heaven and earth.

[0054] Figure 2 A federated learning framework for drone clusters for emergency rescue is provided in this embodiment.

[0055] Figure 3 A two-stage semidefinite relaxed disaster area location map is provided for this embodiment.

[0056] Figure 4 The schematic diagram of the SC-RTDETR provided in this embodiment. DETAILED DESCRIPTION

[0057] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] See also Figure 1 The embodiment of the present invention provides a method for emergency rescue of mountain torrent disasters in coordination with heaven and earth, comprising the following steps:

[0059] Based on the federated learning framework, the following key emergency rescue technologies are deployed and trained; first, in response to the problem that the flash flood disaster area is vast and difficult to respond in time, a wide-area optimal deployment algorithm for drone disaster area clusters based on a dual-target multi-armed tiger machine is proposed to balance the coverage range and communication efficiency of the drone cluster; secondly, in response to the dilemma of GPS signal failure under flash flood disasters, a two-stage semi-definite relaxation disaster area positioning method for space-ground collaboration is developed to ensure that the disaster area can be accurately located; finally, in order to overcome the problems of interference with remote sensing images, low accuracy in small target recognition, and difficulty in emergency rescue and retreat under flash flood disasters, an emergency rescue method based on drone remote sensing vision and dynamic path optimization is studied to improve the efficiency of emergency rescue and retreat.

[0060] See also Figure 2 In this embodiment, the federated learning framework realizes the collaborative learning of clusters by combining the feature extractor update mechanism guided by affinity matrix and the feature aggregation mechanism driven by maximum mean difference (MMD);

[0061] First, the federated learning framework constructs an affinity matrix to quantify the data similarity between different drones, and dynamically adjusts the matrix weights according to the performance of each drone's local model, so that drones with similar feature distributions obtain higher weights; based on the weights, the federated learning framework enhances the aggregation of the drone's local feature extractors to improve the overall feature extraction capabilities of the cluster; then, the federated learning framework introduces MMD to measure the differences in feature distributions of different rounds, and performs weighted averaging of features based on the differences to generate comprehensive feature representations; these comprehensive features can better capture the global characteristics of data distribution and improve the model's generalization ability for unknown data. Based on these comprehensive features, the global classifier can more effectively adapt to the complex and changing flash flood disaster environment.

[0062] In this embodiment, in the deployment scenario of drone clusters in disaster areas, it is crucial to achieve full coverage and rapid response in the disaster area. However, there is a contradiction between maximizing the coverage of the disaster area and minimizing the drone communication convergence time. To solve this problem, the present invention proposes a wide-area optimal deployment algorithm for drone clusters in disaster areas based on a dual-objective multi-armed tiger machine:

[0063] The first optimization goal is to maximize the coverage to achieve full coverage of the disaster area. This process is shown in Formula 1:

[0064] Formula 1:

[0065] where θ = {θ 1 ,θ 2 ,…,θ U} is the beamwidth group assigned to the drone, N i (θ i ) is the beam width θ of drone i i The number of deployed environmental information sensors covered by OL It refers to the number of times multiple drones cover the sensor simultaneously;

[0066] The second optimization goal is to minimize the maximum overall communication time to achieve the fastest deployment under full coverage of the disaster area. This process is shown in Formula 2:

[0067] Formula 2:

[0068] in is the overall communication time, is the average communication time, T i,G is the global update time, T i,L It is the local update time;

[0069] The present invention intends to use the Pareto UCB1 algorithm to solve the above dual-objective optimization problem. The Pareto UCB1 algorithm directly maximizes the reward vector in the reward space of the two objectives. In the initial stage, the Pareto UCB1 algorithm pulls each arm once to complete the initialization. In each round of iteration, the selection process of the Pareto optimal arm set A' is shown in Formula 3:

[0070] Formula 3:

[0071] in is the empirical mean reward of arm α, is the empirical mean reward of arm k, n α is the number of times arm α is pulled, n k is the number of times arm k is pulled, n is the total number of arms, Q is the target number, |A * | is the number of Pareto optimal arm sets. After finding the optimal arm set A', the algorithm randomly selects an arm from it to pull and updates its reward value.

[0072] Finally, the Pareto UCB1 algorithm evaluates its performance by calculating the Pareto regret, which measures the distance of the suboptimal reward vector from the Pareto optimal reward. The combination of the BO-MAB algorithm and the Pareto UCB1 algorithm can effectively optimize the coverage and communication delay in complex drone cluster deployment problems, thereby improving the overall effectiveness and efficiency of emergency response in disaster areas.

[0073] The above design can maximize the coverage of the disaster area while minimizing the overall communication time, achieving the optimal deployment in the wide-area disaster area. It also improves the response speed of the drone cluster and provides technical support for the drone to quickly and accurately locate the disaster area.

[0074] See also Figure 3 In this embodiment, a two-stage semi-definite relaxed disaster area positioning method of earth-sky collaboration is as follows:

[0075] The anchor node (i.e., the node with known location information) is designed with IP68 / IP69K protection level to ensure that it has high reliability and accuracy in flash floods and their secondary disasters (such as mudslides and landslides), and its node location information can be obtained through GPS. However, in extreme scenarios of flash floods, GPS signals are easily interfered with or blocked, resulting in positioning failure. If effective positioning cannot be achieved through GPS, the positioning algorithm of the Internet of Things combined with drones can be used as the final positioning guarantee. In the Internet of Things positioning algorithm of heaven and earth collaboration, the positions of other ordinary nodes can be located according to the anchor nodes. If the ordinary nodes undergo large-scale displacement, it means that there is a high probability that a flash flood disaster has occurred at this time; at present, the positioning accuracy of a large number of Internet of Things positioning algorithms has extremely high requirements on the reliability of the anchor node position, and a small number of anchor nodes are generally affected by other factors (human movement, equipment damage) The interference will generally affect the position calculation of ordinary nodes. In response to this problem, the present invention intends to design a two-stage semi-definite relaxation positioning method of heaven and earth collaboration to deal with the misjudgment of flash floods caused by the unreliable position of a small number of anchor nodes. The algorithm uses drones as relay nodes and collaborates with ground-deployed flash flood situation awareness sensors to achieve reliable positioning. The drones move at a constant speed. Therefore, the modeling process of the instantaneous position of the mth UAV at time step k is shown in Formula 4:

[0076] Formula 4:

[0077] in is the starting position of the mth UAV, j = (k-1)M+m, m = 1, 2, ..., M, k = 1, ..., K;

[0078] Due to the changes in the positions of anchor nodes and the high nonlinearity between the distance measurement of drones and the positions of disaster areas, the numerical solution for disaster area positioning is prone to fall into local optimality, making it difficult to achieve accurate positioning. In order to solve this problem, the present invention introduces a two-stage semidefinite relaxation technology. This technology can eliminate the interference of a small number of unreliable anchor nodes to the positioning process from a global perspective, thereby achieving reliable positioning of other ordinary nodes in the disaster area. The semidefinite relaxation technology can be used to transform complex nonlinear positioning problems into linear problems, thereby ensuring that the positioning algorithm converges to the global optimal solution. The specific process is shown in Formula 5-6:

[0079] Formula 5:

[0080] Formula 6:

[0081] where s i is the position of the i-th ground sensor, s i T is the position of the ith sensor on the ground at time T, s i o is the initial position of the i-th ground sensor, is the receiver position, ε p is the additional noise, u o is the disaster area, d p is the measurement range of the propagation path,

[0082] In the proposed reliable positioning algorithm, P1 = [(x 1 ,y 1 ,z 1 ),(x 2 ,y 2 ,z 2 ),...,(x n ,y n ,z n )] represents the position set of other n common nodes calculated by the positioning algorithm when the anchor node position is completely accurate, P2 = [(x' 1 ,y' 1 ,z' 1 ),(x 2 ,y 2 ,z 2 ),...,(x n ,y n ,z n)] represents the position set of other n common nodes calculated by the positioning algorithm when the anchor node moves a small amount due to other factors (human movement, equipment damage). When no flash flood occurs, if the anchor node is moved artificially, there will be a large difference between the position results P1 and P2 obtained by the general positioning algorithm, which will lead to misjudgment of flash flood disasters. The reliable positioning algorithm proposed in the present invention can eliminate the interference of unreliable anchor nodes, so that the positioning results P1 and P2 are very close, so the misjudgment of flash floods can be prevented. Finally, the location information of the sensors in the disaster area is obtained based on the proposed reliable positioning algorithm, and the flash flood disaster area can be depicted by comparing the position changes of the sensors before and after the flash flood occurs. Under the premise of solving the misjudgment of flash floods, this reliable positioning technology can overcome the difficulties of untimely flash flood monitoring and inaccurate disaster area positioning caused by the inability to normally apply positioning technologies such as GPS after the disaster occurs.

[0083] See also Figure 4 In this embodiment, the emergency rescue method based on UAV remote sensing vision and dynamic path optimization is:

[0084] In flash flood disaster scenarios, drone remote sensing images are easily interfered by factors such as clouds, fog, and rainfall, which leads to a decrease in image quality and seriously affects the identification of targets such as safe spots, disaster-affected houses, and disaster-affected personnel. Assessing safe spots requires accurate remote sensing images and flash flood situation data, but the deterioration in image quality makes it impossible to obtain key image information of safe spots. The identification of disaster-affected houses also faces similar challenges, which in turn affects disaster assessment and resource allocation. The identification of disaster victims is even more difficult because their features are not obvious and are easily drowned out by background noise, and the recognition accuracy is low in interference environments.

[0085] The inaccuracy of these recognition tasks will directly affect the reliability of emergency rescue path planning. The planning of rescue paths depends on the accurate assessment of the locations of safe points, affected houses, and affected people. Once there is a deviation in the target location information, the accuracy and feasibility of path planning will be greatly reduced, causing rescuers to mistakenly enter dangerous areas, increase rescue risks, and delay the best rescue time.

[0086] Focusing on the difficulties in the emergency rescue process, the present invention plans to study an emergency rescue method based on UAV remote sensing vision and dynamic path optimization. This method combines anti-interference UAV remote sensing vision target recognition with a multi-target disaster area emergency rescue path optimization algorithm. The target type in the disaster area is obtained through the recognition algorithm, and the type and location of the target are determined in combination with the UAV position information. Based on the type and location of the target, the path optimization algorithm can formulate the optimal retreat route to ensure the efficiency and safety of the rescue operation;

[0087] The present invention designs a recognition model with Real Time Detection Transformer (RTDETR) as the main structure, combined with a soft threshold-residual network (Soft-Residual Network) and a cascaded attention (CGA) module, called SC-RTDETR. The remote sensing image of the disturbed disaster area is first filtered through a soft threshold-residual network to extract key features, and then input into an efficient hybrid encoder. The encoder combines a cross-scale feature fusion module with a cascaded attention module to ensure that feature information is fully integrated and focused at different scales. Finally, after precise query selection and overlap-based matching by the decoder, SC-RTDETR can accurately identify the affected houses and people.

[0088] SC-RTDETR eliminates interference in remote sensing images through a soft threshold-residual network module. First, the input feature map of the affected disaster area remote sensing image is extracted and enhanced through a standard residual block. The residual block directly passes the input to the output through a jump connection. At the same time, it is processed by multiple convolutional layers to enhance the feature expression ability. Next, the enhanced feature map dynamically adjusts the filter threshold according to the specific characteristics of each image signal. After the feature map is processed by soft threshold adaptive filtering, the injected interference signal is accurately identified and filtered. This adaptive filtering can ensure that the output feature map can remove interference to the greatest extent while maintaining important information. The principle is shown in Formula 7:

[0089] Formula 7: y=sign(F(x)+x)·max(|F(x)+x|-λ,0);

[0090] Where x is the interfered remote sensing image, λ is the threshold parameter, F(x) represents the nonlinear transformation in the residual block, and y is the remote sensing image with interference removed.

[0091] SC-RTDETR improves the accuracy of small target recognition through cascaded attention modules. Cascaded attention optimizes the model's attention mechanism through grouping and cascading strategies, thereby improving the target recognition ability in complex scenes. First, the input feature map is divided into multiple groups according to the spatial or channel dimension, and the feature map of each group is independently calculated for attention. This grouping mechanism ensures that the model can focus on fine-grained local features in the image and reduce the interference caused by global information. In each group, the self-attention mechanism is used to calculate the correlation between features within the group, thereby improving the expression ability of local features. Then, the attention results of each group are connected in series through a cascade mechanism and accumulated layer by layer. This cascade mechanism effectively integrates the attention information of different groups, allowing the model to form a more comprehensive feature representation in the global scope. Finally, the feature map that has been grouped and cascaded is output to the next layer of the network to obtain an enhanced feature representation. This structural design enables the cascaded attention module to reduce the computational complexity while enhancing the model's attention allocation to key areas. It is particularly suitable for small target recognition and target detection tasks in complex scenes. The principle of cascaded attention is shown in Formula 8:

[0092] Formula 8:

[0093] Where X L It is the feature representation of remote sensing images after multi-layer cascade attention mechanism, FC is the fully connected layer, is a linear transformation matrix, LayerNorm() is a normalization layer, and Y is the output target recognition result. The target recognition result is combined with the drone's own position information to obtain the information of safe points, disaster-stricken houses and disaster-stricken people, which is the key to planning emergency rescue routes in disaster areas.

[0094] As a further technical solution of the present invention, an emergency rescue method based on UAV remote sensing vision and dynamic path optimization:

[0095] In the emergency rescue path planning, the present invention will fully integrate the flash flood situation data and UAV remote sensing images, conduct a comprehensive analysis of the terrain of the disaster-stricken area, the spread of flash floods, etc., evaluate the emergency waiting rescue area of ​​the UAV, and plan the retreat path on this basis. The present invention intends to design a multi-target disaster area emergency rescue path optimization algorithm to optimize the emergency rescue path while considering multiple factors, such as the location of the safe point, the location of the affected house, the location of the affected personnel (including the number of people at different locations), the time cost of the path, and the risk assessment of the path. The set of safe points is defined as S = {s 1 ,s 2 ,...,s m}, where m represents the number of safe points, s jrepresents the position of the jth safe point. The position of the victims is the most important consideration in path planning. The location set of the victims is P = {(p 1 ,N 1 ),(p 2 ,N 2 ),...,(p n ,N n )}, each location may have a different number of victims, where p i represents the location of the i-th disaster victim, N i Indicates that at position p i The number of disaster victims on the hazard-affected houses may contain unidentified disaster victims. The set of disaster-affected house locations is defined as H = {(h 1 ,a 1 ),(h 2 ,a 2 ),...,(h k ,a k )}, each element contains the location of the house and the number of possible unidentified disaster victims at that location;

[0096] In order to optimize the evacuation route, the objective function is designed to minimize the total time, and weight coefficients α and β are introduced to represent the relative priorities of disaster victims and houses, where the weight coefficient α is based on the number of disaster victims N. i Adjustment, while considering the risk factors of the path, the specific expression of the objective function is shown in Formula 9:

[0097] Formula 9:

[0098] Among them, x ij ·N i ·d ij Indicates that the victims are from position p i To a safe spot j The evacuation cost is based on the number of people affected N i and path distance d ij ,y ij ·a i ·d ij Indicates that from the affected houses i To a safe spot j Evacuation costs, based on the number of unidentified affected persons in potential homesa i and path distance d ij , w ij is the risk weight of the path;

[0099] Combined with the number of disaster victims, the path capacity constraint also needs to consider the number of disaster victims at each location. The constraint expression is shown in Formula 10:

[0100] Formula 10:

[0101] Where N i ·x ij Indicates that from position p i To a safe spot j Total number of disaster victims evacuated, a i ·y ij Indicates the location of the affected house h i To a safe spot j The number of unidentified disaster victims, C j represents the path carrying capacity;

[0102] According to the time and risk assessment of different paths, the multi-target disaster area emergency rescue path optimization algorithm optimizes the overall rescue route through phased decision-making to ensure the efficiency and safety of rescue operations under flash flood disasters.

[0103] In summary: Federated learning framework: In the emergency rescue of flash flood disasters, the UAV cluster wide-area optimal deployment algorithm based on dual-objective multi-armed tiger machine, the two-stage semi-definite relaxation disaster area positioning algorithm based on ground-ground collaboration, and the emergency rescue retreat algorithm based on UAV remote sensing vision and dynamic path optimization will be trained and applied under this federated framework. The framework combines the feature extractor update guided by the affinity matrix with the feature aggregation mechanism driven by the maximum mean difference (MMD) to achieve cluster collaborative learning. First, the data similarity between UAVs is quantified by constructing an affinity matrix, and the weights are dynamically adjusted according to the performance of the local model of each UAV, so that UAVs with similar feature distributions obtain higher weights. Subsequently, the framework enhances the aggregation ability of local feature extractors based on these weights and improves the overall feature extraction effect of the cluster. At the same time, the framework uses MMD to measure the differences in feature distributions of different rounds, and weighted averages the features to generate more globally representative comprehensive features, enhancing the generalization ability of the model to unknown data. Finally, the global classifier is better adapted to the complex flash flood disaster environment based on these comprehensive features.

[0104] The two-stage semi-definite relaxed disaster area positioning of the space-ground collaboration can eliminate the interference of unreliable anchor nodes and prevent misjudgment of flash floods. The location information of the sensors in the disaster area is obtained based on the reliable positioning algorithm. By comparing the position changes of the sensors before and after the flash flood, the flash flood disaster area can be depicted. On the premise of solving the misjudgment of flash floods, this reliable positioning technology can overcome the difficulties of untimely flash flood monitoring and inaccurate disaster area positioning caused by the inability to normally apply positioning technologies such as GPS after the disaster.

[0105] SC-RTDETR: The remote sensing images of the affected disaster areas are first filtered through a soft threshold-residual network to extract key features, and then input into an efficient hybrid encoder. The encoder combines a cross-scale feature fusion module with a cascaded attention module to ensure that feature information is fully integrated and focused at different scales. Finally, after the decoder's precise query selection and overlap-based matching, SC-RTDETR can accurately identify affected houses and people.

[0106] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention.

[0107] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for emergency rescue of mountain torrent disasters in cooperation with heaven and earth, characterized in that: The following steps are involved: Based on the federated learning framework, the following key emergency rescue technologies are deployed and trained. First, a wide-area optimal deployment algorithm for drone clusters in disaster areas based on a dual-objective multi-armed slot machine is proposed to balance the coverage and communication efficiency of drone clusters. Secondly, a two-stage semi-definite relaxed disaster area positioning method based on space-ground collaboration is developed to ensure that the disaster area is accurately located; finally, an emergency rescue method based on UAV remote sensing vision and dynamic path optimization is proposed.

2. The method for emergency rescue of mountain torrent disasters in cooperation with heaven and earth according to claim 1, characterized in that: The federated learning framework achieves collaborative learning of clusters by combining the feature extractor update mechanism guided by affinity matrix and the feature aggregation mechanism driven by maximum mean difference (MMD). First, the federated learning framework constructs an affinity matrix to quantify the data similarity between different drones, and dynamically adjusts the matrix weights according to the performance of each drone’s local model, so that drones with similar feature distributions receive higher weights. Based on the weights, the federated learning framework performs enhanced aggregation on the local feature extractors of the drones, thereby improving the overall feature extraction capability of the cluster; Subsequently, the federated learning framework introduced MMD to measure the differences in feature distributions in different rounds, and performed weighted averaging of the features based on the differences to generate a comprehensive feature representation.

3. The method for emergency rescue of mountain torrent disasters in cooperation with heaven and earth according to claim 2 is characterized in that: Optimal deployment algorithm for UAV disaster area clusters based on dual-objective multi-armed bandit: The first optimization goal is to maximize the coverage to achieve full coverage of the disaster area. This process is shown in Formula 1: Formula 1: where θ={θ1,θ2,…,θ U } is the beamwidth group assigned to the drone, N i (θ i ) is the beam width θ of drone i i The number of deployed environmental information sensors covered by OL It refers to the number of times multiple drones cover the sensor simultaneously; The second optimization goal is to minimize the maximum overall communication time to achieve the fastest deployment under full coverage of the disaster area. This process is shown in Formula 2: Formula 2: in is the overall communication time, is the average communication time, T i,G is the global update time, T i,L It is the local update time; The Pareto UCB1 algorithm is used to solve the dual-objective optimization problem. The Pareto UCB1 algorithm directly maximizes the reward vector in the reward space of the two objectives. In the initial stage, the Pareto UCB1 algorithm pulls each arm once to complete the initialization. In each round of iteration, the selection process of the Pareto optimal arm set A' is shown in Formula 3: in is the empirical mean reward of arm α, is the empirical mean reward of arm k, n α is the number of times arm α is pulled, n k is the number of times arm k is pulled, n is the total number of arms, Q is the target number, |A * | is the number of Pareto optimal arm sets. After finding the optimal arm set A', the algorithm randomly selects an arm to pull and updates its reward value; Finally, the Pareto UCB1 algorithm evaluates its performance by calculating the Pareto regret.

4. The method for emergency rescue of mountain torrent disasters in cooperation with heaven and earth according to claim 1, characterized in that: Two-stage semidefinite relaxation disaster area positioning method for space-ground collaboration: Using drones as relay nodes, the drones collaborate with the flash flood situation awareness sensors deployed on the ground to achieve positioning. The drones move at a constant speed. Therefore, the modeling process of the instantaneous position of the mth UAV at time step k is shown in Formula 4: Formula 4: in is the starting position of the mth UAV, j = (k-1)M+m, m = 1, 2, ..., M, k = 1, ..., K; Due to the change of anchor node position and the high nonlinearity between drone ranging and disaster area location, the complex nonlinear positioning problem is transformed into a linear problem through semidefinite relaxation technology, thereby ensuring that the positioning algorithm converges to the global optimal solution. The specific process is shown in Formula 5-6: Formula 5: Formula 6: where s i is the position of the i-th ground sensor, s i T is the position of the ith sensor at time T on the ground, s i o is the initial position of the i-th ground sensor, is the receiver position, ε p is the additional noise, u o is the disaster area, d p is the measurement range of the propagation path, P1=[(x1,y1,z1),(x2,y2,z2),...,(x n ,y n ,z n )] represents the position set of other n common nodes calculated by the positioning algorithm when the anchor node position is completely accurate, P2 = [(x'1, y'1, z'1), (x2, y2, z2), ..., (x n ,y n ,z n )] represents the position set of other n common nodes calculated by the positioning algorithm when the anchor node moves due to other factors. The location information of the sensors in the disaster area is obtained based on the proposed positioning algorithm. By comparing the position changes of the sensors before and after the flash flood, the flash flood disaster area can be depicted.

5. The method for emergency rescue of mountain torrent disasters in cooperation with heaven and earth according to claim 1, characterized in that: Emergency rescue method based on UAV remote sensing vision and dynamic path optimization: The emergency rescue method based on UAV remote sensing vision and dynamic path optimization combines anti-interference UAV remote sensing vision target recognition with multi-target disaster area emergency rescue path optimization algorithm. The target type in the disaster area is obtained through the recognition algorithm, and the type and location of the target are determined in combination with the UAV position information. Based on the type and location of the target, the path optimization algorithm formulates the optimal retreat route to ensure the efficiency and safety of the rescue operation; The design uses Real Time Detection Transformer as the main structure, combined with a soft threshold-residual network and a cascaded attention module to create a recognition model called SC-RTDETR. The remote sensing images of the affected disaster area are first filtered through a soft threshold-residual network to extract key features, and then input into an efficient hybrid encoder. The encoder combines a cross-scale feature fusion module with a cascaded attention module to ensure that feature information is fully integrated and focused at different scales. Finally, after the decoder's precise query selection and overlap-based matching, SC-RTDETR accurately identifies affected houses and people. SC-RTDETR eliminates interference in remote sensing images through a soft threshold-residual network module. First, the input feature map of the affected disaster area remote sensing image is extracted and enhanced through a standard residual block. The residual block directly passes the input to the output through a jump connection. At the same time, it is processed by multiple convolutional layers to enhance the feature expression ability. Next, the enhanced feature map dynamically adjusts the filter threshold according to the specific characteristics of each image signal. After the feature map is processed by soft threshold adaptive filtering, the injected interference signal is accurately identified and filtered. This adaptive filtering can ensure that the output feature map can remove interference to the greatest extent while maintaining important information. The principle is shown in Formula 7: Formula 7: y=sign(F(x)+x)·max(|F(x)+x|-λ,0); Where x is the interfered remote sensing image, λ is the threshold parameter, F(x) represents the nonlinear transformation in the residual block, and y is the remote sensing image with interference removed; SC-RTDETR improves the recognition accuracy of small targets through the cascaded attention module. The cascaded attention optimizes the model's attention mechanism through grouping and cascading strategies, thereby improving the target recognition ability in complex scenes. First, the input feature map is divided into multiple groups according to the spatial or channel dimension. The feature map of each group is independently calculated for attention. In each group, the self-attention mechanism is used to calculate the correlation between the features within the group. Then, the attention results of each group are connected in series through the cascade mechanism and accumulated layer by layer. Finally, the feature map after grouping and cascading processing is output to the next layer of the network to obtain the enhanced feature representation. The principle of cascaded attention is shown in Formula 8: Formula 8: Where X L It is the feature representation of remote sensing images after multi-layer cascade attention mechanism, FC is the fully connected layer, is a linear transformation matrix, LayerNorm() is a normalization layer, and Y is the output target recognition result. The target recognition result is combined with the drone's own position information to obtain the information of safe points, disaster-stricken houses and disaster-stricken people.

6. The method for emergency rescue of mountain torrent disasters in cooperation with heaven and earth according to claim 1, characterized in that: Emergency rescue methods based on UAV remote sensing vision and dynamic path optimization also include: The safe point set is defined as S = {s1, s2, ..., s m }, where m represents the number of safe points, s j represents the position of the jth safe point. The position of the victims is the most important consideration in path planning. The position set of the victims is P = {(p1, N1), (p2, N2), ..., (p n ,N n )}, where p i represents the location of the i-th disaster victim, N i Indicates that at position p i The number of disaster victims on the , the location of the disaster-affected houses may contain unidentified disaster victims, and the location set of disaster-affected houses is defined as H = {(h1,a1),(h2,a2),...,(h k ,a k )}, each element contains the location of the house and the number of possible unidentified disaster victims at that location; In order to optimize the evacuation route, the objective function is designed to minimize the total time, and weight coefficients α and β are introduced to represent the relative priorities of disaster victims and houses, where the weight coefficient α is based on the number of disaster victims N. i Adjustment, while considering the risk factors of the path, the specific expression of the objective function is shown in Formula 9: Formula 9: Among them, x ij ·N i ·d ij Indicates that the victims are from position p i To a safe spot j The evacuation cost is based on the number of people affected N i and path distance d ij ,y ij ·a i ·d ij Indicates that from the affected houses h i To a safe spot j Evacuation costs, based on the number of unidentified affected persons in potential homesa i and path distance d ij , w ij is the risk weight of the path; Combined with the number of disaster victims, the path capacity constraint also needs to consider the number of disaster victims at each location. The constraint expression is shown in Formula 10: Formula 10: Where N i ·x ij Indicates that from position p i To a safe spot j Total number of disaster victims evacuated, a i ·y ij Indicates the location of the affected house h i To a safe spot j The number of unidentified disaster victims, C j represents the path carrying capacity; According to the time and risk assessment of different paths, the multi-target disaster area emergency rescue path optimization algorithm optimizes the overall rescue route through phased decision-making to ensure the efficiency and safety of rescue operations under flash flood disasters.

Images