Emergency dispatching method and system for high-altitude operation

Abstract

The application provides a kind of aerial work emergency scheduling method and system, first acquire the real-time video data set of aerial work area and pre-process to generate standardized video data stream, then call video analysis model to extract work environment feature set and work personnel behavior feature set, then based on work environment feature set, graph structure is constructed, environment state feature vector is generated by graph convolution network multi-level feature aggregation, then environment state feature vector and work personnel behavior feature set are combined to generate dynamic emergency scheduling strategy, finally, after the strategy is verified, emergency scheduling instruction set is generated and transmitted to work terminal equipment to trigger emergency response operation, whereby it can dynamically generate targeted emergency scheduling strategy according to actual situation, realize the automation and intelligentization of emergency scheduling, improve emergency response efficiency.

Description

Technical Field

[0001] This invention relates to the field of high-altitude operation safety assurance and emergency management technology, and more specifically, to a high-altitude operation emergency dispatch method and system. Background Technology

[0002] In the field of high-altitude operations, the working environment is complex and ever-changing, and the behavior of the workers directly affects the safety of the operation. In the event of emergencies, such as sudden worsening of severe weather or equipment failure, timely and effective emergency dispatch is crucial. Currently, emergency dispatch for high-altitude operations mainly relies on human experience and simple feedback from on-site personnel. Human experience is highly subjective and struggles to fully and accurately grasp the complex and changing environmental conditions, making it impossible to quickly and precisely formulate emergency dispatch strategies. Feedback from on-site personnel is often delayed and inaccurate, and it is difficult to integrate information about the entire working environment holistically.

[0003] Furthermore, existing emergency dispatch methods lack in-depth analysis of the characteristics of the working environment and the behavior of workers, and cannot dynamically generate targeted emergency dispatch strategies based on the actual situation. This results in slow emergency response and unreasonable dispatch when facing emergencies, increasing the risks of working at heights and seriously threatening the lives of workers. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an emergency dispatch method for high-altitude operations, the method comprising:

[0005] Acquire a set of real-time video data from the high-altitude work area, and perform preprocessing operations on the real-time video data set to generate a standardized video data stream;

[0006] The video analysis model is invoked to perform feature extraction on the standardized video data stream, generating a set of work environment features and a set of worker behavior features;

[0007] Based on the set of operational environment features, an operational environment graph structure is constructed. A graph convolutional network is then invoked to perform multi-level feature aggregation operations on the operational environment graph structure to generate an environmental state feature vector.

[0008] A dynamic emergency dispatch strategy is generated based on the environmental state feature vector and the set of operator behavior features.

[0009] The effectiveness verification operation of the dynamic emergency dispatch strategy is performed to obtain the strategy verification result, and an emergency dispatch instruction set is generated based on the strategy verification result. The emergency dispatch instruction set is then transmitted to the operation terminal equipment to trigger the emergency response operation.

[0010] In another aspect, embodiments of the present invention also provide a high-altitude operation emergency dispatch system, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this embodiment of the invention obtains and merges real-time video datasets of high-altitude work areas, performs preprocessing to generate standardized video data streams, and calls video analysis models to extract sets of work environment features and sets of worker behavior features. This comprehensively captures key information in high-altitude work scenarios. Based on the set of work environment features, a work environment graph structure is constructed, and multi-level feature aggregation is performed using a graph convolutional network. The generated environmental state feature vector accurately reflects the comprehensive state of the work environment. Combining the environmental state feature vector and the set of worker behavior features, a dynamic emergency dispatch strategy is generated, fully considering the interaction between the work environment and personnel behavior. This makes the dynamic emergency dispatch strategy more targeted and adaptable. The effectiveness of the dynamic emergency dispatch strategy is verified, and a set of emergency dispatch instructions is generated and transmitted to the work terminal equipment to trigger emergency response operations. This achieves automation and intelligence in emergency dispatch, greatly improves the speed and accuracy of emergency response, effectively reduces the risks of high-altitude work, and protects the lives of workers. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the high-altitude operation emergency dispatch method provided in the embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of exemplary hardware and software components of the high-altitude operation emergency dispatch system provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an embodiment of the emergency dispatch method for high-altitude operations provided by the present invention. The following is a detailed description of the emergency dispatch method for high-altitude operations.

[0015] Step S110: Obtain a set of real-time video data of the high-altitude operation area, and perform preprocessing operations on the set of real-time video data to generate a standardized video data stream.

[0016] In this embodiment, a high-altitude maintenance operation scenario for a large wind turbine generator is taken as an example. In this scenario, multiple video acquisition devices, such as cameras C1, C2, and C3, are deployed around the high-altitude operation area of ​​the wind turbine generator. These cameras capture the operation area in real time from different angles, and the video data they collect constitutes a real-time video data set. Due to differences in the performance and settings of different cameras, these video data may have inconsistencies in resolution, frame rate, color gamut, etc. Therefore, preprocessing is required to generate a standardized video data stream suitable for subsequent analysis.

[0017] Step S111: Perform resolution equalization processing on the real-time video data set to generate initial video data frames with a uniform resolution format.

[0018] Assume that the video resolution captured by camera C1 is a1×b1, the video resolution captured by camera C2 is a2×b2, and the video resolution captured by camera C3 is a3×b3. To achieve resolution equalization, a unified target resolution of A×B is first determined. For the video captured by camera C1, an image scaling algorithm is used. This algorithm processes each video frame. Taking horizontal pixels as an example, for each pixel position in the original resolution a1, based on its relationship with neighboring pixels and according to the set interpolation rules, the pixel value at the corresponding position under the target resolution A is calculated. For example, if a linear interpolation method is used, for the original pixel position i and the new pixel position j, the new pixel value is calculated based on the pixel values ​​of the original pixel positions i and i+1, according to the relative position relationship between i and j. The vertical pixels are calculated in the same way. Through this processing, the resolution of the video captured by camera C1 is adjusted to A×B. The same method is applied to the videos captured by cameras C2 and C3, ultimately generating initial video data frames with a unified resolution format of A×B.

[0019] Step S112: Call the moving target detection algorithm to perform noise filtering on the initial video data frame, remove the interfering pixel areas of non-targets, and obtain the filtered video data frame.

[0020] In this embodiment, the initial video data frame may contain various non-operational targets causing interference, such as debris drifting in the wind or images of other unrelated facilities in the distance. The moving target detection algorithm identifies interference by analyzing the motion characteristics of pixel regions in the initial video data frame, specifically by comparing the changes in pixel regions at the same location in adjacent frames frame by frame.

[0021] For example, if a pixel region's positional, color, or texture changes between adjacent frames do not conform to the motion patterns of the target object—such as abnormally fast or slow movement speed, or chaotic motion trajectory—then the pixel region is determined to be an interfering pixel region that is not part of the target object. Multiple processing methods are employed for pixel regions identified as interfering.

[0022] One approach is to set the pixel value to be similar to the surrounding background. This can be done by analyzing the color, brightness, and other characteristics of the surrounding background pixels to calculate a suitable pixel value to replace the pixel value of the interfering area. Another approach is to perform interpolation calculations based on the pixel values ​​of the surrounding work targets to generate a pixel value that blends more seamlessly with the surrounding work scene, thereby removing interfering pixel areas and obtaining filtered video data frames.

[0023] Step S113: Perform time alignment processing on the filtered video data frames to generate a standardized video data stream with continuous timestamps.

[0024] Because the startup time and frame rate of different cameras C1, C2, and C3 may differ, the captured video data will be inconsistent in time. Let the starting acquisition time of camera C1 be t1, the starting acquisition time of camera C2 be t2, and the starting acquisition time of camera C3 be t3, and their frame rates be f1, f2, and f3, respectively.

[0025] To generate a standardized video data stream with continuous timestamps, a unified start time T0 is first determined. For the video data captured by camera C1, the time difference Δt1n (where n represents the frame number) of each frame relative to T0 is calculated. Timing alignment is achieved by adjusting the frame rate or inserting / deleting frames. For example, if the frame rate f1 of camera C1 is higher than the expected unified frame rate F, and the calculated time difference Δt1n of a certain frame deviates significantly from the time point calculated according to the unified frame rate, some frames can be appropriately deleted. When deleting frames, frames with less impact on the overall video information are selected for deletion based on the content relationship between preceding and following frames. If the frame rate f1 is lower than the unified frame rate F, some intermediate frames are inserted using an interpolation algorithm. The interpolation algorithm calculates the pixel values ​​of the inserted frames based on the pixel values ​​and motion trends of preceding and following frames.

[0026] The same method is applied to the video data captured by cameras C2 and C3, and finally all the video data is arranged in chronological order, with each video frame corresponding to a continuous timestamp, forming a standardized video data stream with continuous timestamps.

[0027] Step S114: Perform a color gamut conversion operation on the standardized video data stream to map the original color gamut space to a preset standardized color gamut space, and perform a light compensation processing operation on the converted video data stream to eliminate the shadow area and overexposed area in the initial video data frame, obtaining a compensated video data stream.

[0028] In this embodiment, the original color gamut spaces of videos collected by different cameras may be different. Let the original color gamut space of camera C1 be G1, the original color gamut space of camera C2 be G2, the original color gamut space of camera C3 be G3, and the preset standardized color gamut space be G0. Using a color gamut conversion algorithm, convert the color value of each pixel in the standardized video data stream from its original color gamut space to the preset standardized color gamut space G0. For the red (R), green (G), and blue (B) color components of each pixel, calculate according to the conversion relationship between the original color gamut space and G0.

[0029] For example, for the video collected by camera C1, if the value of its red component in the original color gamut space G1 is R1, calculate the value R0 of the red component in G0 according to the conversion function from G1 to G0. After completing the color gamut conversion, perform light compensation processing. Because at the work site, some areas may be overexposed due to direct sunlight, and some areas may be in shadow due to equipment occlusion. The light compensation algorithm will analyze the brightness value L of each pixel. For pixels with too high brightness (overexposed), reduce their brightness. For example, through an adjustment coefficient k1 (0 < k1 < 1), multiply the R, G, and B component values of the overexposed pixels by k1 respectively. For pixels with too low brightness (shadow), increase their brightness. For example, through an adjustment coefficient k2 (k2 > 1), multiply the R, G, and B component values of the shadow pixels by k2 respectively, so as to eliminate the shadow area and overexposed area in the initial video data frame, obtaining a compensated video data stream.

[0030] Step S115: Perform a frame rate stability detection operation on the compensated video data stream to eliminate video jitter and generate a temporally coherent standardized video data stream.

[0031] In high-altitude wind turbine operation scenarios, camera shake may occur due to strong winds and equipment vibrations, leading to unstable video frame rates. A frame rate stability detection algorithm determines frame rate stability by analyzing the time intervals between adjacent frames and changes in image content within the compensated video data stream. Let the time interval between adjacent frames be Δt, and the average time interval be ΔT. If the deviation between the time interval Δt and the average time interval ΔT of a particular frame exceeds a set threshold, it indicates frame rate instability. Simultaneously, video jitter is detected by analyzing the displacement and deformation of image content in adjacent frames. For segments exhibiting frame rate instability and video jitter, appropriate processing measures are taken. If a segment of video has an excessively fast frame rate, some intermediate frames are inserted. The pixel values ​​of the inserted intermediate frames are obtained by interpolating the pixel values ​​of the preceding and following frames.

[0032] For example, bilinear interpolation is used to calculate the pixel value of the corresponding position in the intermediate frame based on the four neighboring pixel values ​​of the corresponding position in the preceding and following frames. If the frame rate of a certain video segment is too slow, some redundant frames are appropriately deleted. When deleting frames, priority is given to frames that have little impact on the continuity of the video content, such as frames with small image changes. After these operations, video jitter is eliminated, and a time-coherent, standardized video data stream is generated.

[0033] Step S120: Call the video analysis model to perform feature extraction on the standardized video data stream to generate a set of work environment features and a set of worker behavior features.

[0034] In the scenario of high-altitude maintenance of large wind turbine generators, a trained video analysis model is used to extract features from standardized video data streams to obtain feature information related to the working environment and personnel.

[0035] Step S121: Divide the standardized video data stream into multiple video segments of equal length, and perform a three-dimensional convolutional feature extraction operation on each video segment to generate a spatiotemporal correlation feature matrix.

[0036] The standardized video data stream is divided into multiple equal-length video segments S1, S2, S3, etc., according to a fixed length L. For each video segment Si, a three-dimensional convolutional feature extraction operation is performed. The three-dimensional convolution operation performs convolution calculations in the time dimension (let the length of the time dimension be t), the horizontal spatial dimension (let the length of the horizontal spatial dimension be x), and the vertical spatial dimension (let the length of the vertical spatial dimension be y) of the video segment. Let the convolution kernel be K, with a length of kt in the time dimension, kx in the horizontal spatial dimension, and ky in the vertical spatial dimension. For each frame in the video segment Si, the convolution kernel K slides at different positions, and the pixel value at each position is weighted and calculated. For example, for the nth pixel position in the mth frame of the video segment Si, the convolution kernel K covers the pixel at that position and its surrounding pixels. The new pixel value Pnew is obtained by multiplying the pixel value p at the covered position by the weight w corresponding to each element in the convolution kernel K, i.e., Pnew = Σ(w*p). In the time dimension, the convolution kernel K performs calculations across several adjacent frames, thereby capturing the spatiotemporal correlation information in the video clip.

[0037] Through the above three-dimensional convolution operation, each video segment Si generates a spatiotemporal correlation feature matrix Fi, which contains the feature information of the video segment in the spatiotemporal dimension.

[0038] Step S122: Perform attention weight allocation operation on the spatiotemporal correlation feature matrix to generate a visual focus distribution map with regional importance annotations, and perform edge tracking operation on the outline of the working equipment based on the visual focus distribution map to extract equipment operating status features and spatial displacement trajectory features.

[0039] In this embodiment, the attention weight allocation operation analyzes the importance of the region represented by each element in the spatiotemporal correlation feature matrix Fi to the work scenario. Regions related to the work equipment are assigned higher weights, while irrelevant background regions are assigned lower weights. For example, the importance weights are determined by analyzing elements in the feature matrix related to features such as the equipment's shape and color. In this way, a visual focus distribution map Vi with regional importance annotations is generated.

[0040] In high-altitude maintenance operations of wind turbine generators, edge tracking is performed on the outline of the equipment based on a visual focus distribution map Vi. The approximate location of the equipment is determined from Vi, and then, starting from that location, the outline of the equipment is tracked point-by-point along its edge, based on changes in pixel values. During tracking, the positional information of each point on the equipment outline is recorded to extract the equipment's operating status features, such as whether the equipment is stationary or rotating, and if rotating, its rotational speed. Simultaneously, by comparing the positional information of the equipment outline at different time points, the spatial displacement trajectory features of the equipment are extracted, for example, the equipment moving from position P1 to position P2 over a period of time.

[0041] Step S123: Perform skeletal key point detection on the worker's limb movements to generate a set of movement amplitude features and movement speed features.

[0042] In high-altitude maintenance operations of wind turbine generators, workers perform various tasks. A skeletal keypoint detection method is used to analyze images of the workers in video data and identify key skeletal points such as the wrist, elbow, and shoulder. Based on the positional changes of these key points in different frames, the range of motion of the workers' limbs is calculated.

[0043] For example, the change in distance between wrist and elbow keypoints in adjacent frames is calculated to obtain relevant data on the amplitude of movement. Simultaneously, the movement velocity of the worker's limbs is calculated based on the displacement of skeletal keypoints per unit time. Suppose a skeletal keypoint moves from position P1 to position P2 within a time interval Δt. By calculating the distance d between the two points and dividing it by the time interval Δt, the movement velocity v = d / Δt is obtained. These amplitude and velocity data are then processed to generate sets of amplitude and velocity feature data.

[0044] Step S124: Aggregate the equipment operating status features, spatial displacement trajectory features, motion amplitude features, and motion speed features to generate the set of worker behavior features.

[0045] The equipment operating status features and spatial displacement trajectory features extracted from the operating equipment are summarized with the motion amplitude features and movement speed features extracted from the operator's limb movements. These different types of features are then integrated according to the set rules.

[0046] For example, the equipment operating status characteristics are first classified according to different equipment types, and then spatial displacement trajectory characteristics are associated with them. For motion amplitude and speed characteristics, they are classified and organized according to different operational stages of the workers. Then, these classified and organized features are arranged in a set order, or weighted according to their correlation with operational safety, efficiency, etc., and then integrated to finally generate a set of worker behavior characteristics.

[0047] Step S125: Perform texture analysis on the fixed obstacles in the working environment to extract surface roughness features and geometric shape features.

[0048] In the high-altitude operation environment of wind turbine generators, there are some fixed obstacles, such as protruding parts of the tower and fixed supports of the equipment. Texture analysis is performed on these fixed obstacles. By analyzing the grayscale changes and color distribution of pixels on the obstacle surface, surface roughness features are extracted. For example, the standard deviation of the grayscale values ​​of surface pixels is calculated; the larger the standard deviation, the higher the surface roughness. Simultaneously, by identifying and analyzing the outline of the obstacle, its geometric shape features are determined, such as whether it is a cylinder, cuboid, etc., as well as relevant dimensional information such as length, width, and height.

[0049] Step S126: By fusing the surface roughness features, geometric features and equipment operating status features through a graph neural network, a dynamic interaction model of material properties and equipment kinematic parameters is constructed to generate a set of environmental interaction risk features.

[0050] This paper utilizes a graph neural network to fuse the extracted surface roughness and geometric features of fixed obstacles with the equipment operating status features of the working equipment. The graph neural network represents the relationships between different features through nodes and edges. Surface roughness, geometric features, and equipment operating status features are treated as different nodes, and the connection methods and weights of the edges are determined based on their actual correlations. For example, if the equipment operating status features and the geometric features of the obstacle are closely related in actual operation, the edge connecting these two feature nodes will have a higher weight. In this way, a dynamic interaction model is constructed between material properties (reflected by surface roughness, etc.) and equipment kinematic parameters (reflected by equipment operating status). In this dynamic interaction model, the interactions between different features are analyzed, such as how the equipment operating status is affected by the material and shape of the obstacle. Based on the analysis results, an environmental interaction risk feature set is generated. This set contains feature information related to the risks that may arise from the interaction between the equipment and obstacles in the working environment, such as the probability of collision risk and the degree of risk impact.

[0051] Step S130: Construct a work environment graph structure based on the set of work environment features, and call a graph convolutional network to perform multi-level feature aggregation operations on the work environment graph structure to generate an environment state feature vector.

[0052] In the scenario of high-altitude maintenance of large wind turbine generators, an operational environment graph structure is constructed based on the extracted set of operational environment features, and a graph convolutional network is used to perform multi-level feature aggregation operations to generate feature vectors that can reflect the environmental state.

[0053] Step S131: Set the work equipment node, environmental obstacle node, and personnel location node as the basic node set of the graph structure.

[0054] In this scenario, maintenance equipment for wind turbine generators, such as maintenance baskets and tool lifting platforms, are designated as work equipment nodes; protruding structures of the tower and surrounding fixed facilities are designated as environmental obstacle nodes; and the positions of workers at different times are designated as personnel position nodes. These nodes together constitute the basic node set of the graph structure.

[0055] Step S132: Construct a first type of edge connection relationship based on the actual distance between the device node and the obstacle node, and construct a second type of edge connection relationship based on the operational association between the personnel location node and the device node.

[0056] For equipment nodes and environmental obstacle nodes, the actual distance between them is determined through measurement or calculation based on video images, thus constructing a first-type edge connection relationship. For example, a laser rangefinder can be used, or the distance between equipment nodes and obstacle nodes can be calculated based on the proportional relationships in video images and the principles of triangulation. Edge connections between closer nodes have relatively higher weights, indicating a closer spatial relationship. For personnel location nodes and equipment nodes, a second-type edge connection relationship is constructed based on operational correlation factors such as the frequency and duration of personnel operation on the equipment. If a personnel frequently and for extended periods operate a particular piece of equipment, the edge connection weight between that personnel location node and the corresponding equipment node is higher, indicating a strong operational correlation between them.

[0057] Step S133: Construct a node attribute feature set based on environmental monitoring sensor data. The node attribute feature set includes equipment operating parameters, obstacle material properties, and personnel status parameters. The personnel status parameters include personnel protective equipment status parameters and personnel vital sign status parameters.

[0058] In high-altitude maintenance operations of wind turbine generators, various environmental monitoring sensors are deployed. These sensors acquire equipment operating parameters, such as the lifting speed of the maintenance scaffold and the load-bearing capacity of the tool lifting platform; they also acquire the material properties of obstacles, such as whether they are made of metal or composite materials; and they acquire personnel status parameters, including personal protective equipment status parameters such as the tightness of safety belts and whether safety helmets are worn correctly, and vital signs such as heart rate and blood pressure. This data is then integrated to construct a set of node attribute features.

[0059] Step S134: Combine the basic node set, edge connection relationship and node attribute feature set to generate the working environment graph structure.

[0060] Step S134-1: Perform vectorization encoding on the device operating parameters in the node attribute feature set to generate device attribute vectors.

[0061] For equipment operating parameters in the node attribute feature set, such as the lifting speed of the maintenance hoist and the load capacity of the tool lifting platform, a vectorized encoding method is adopted. Each parameter is mapped to a vector space, and a dimension is assigned to each parameter. For example, if the equipment operating parameters include lifting speed V and load capacity W, the lifting speed V is mapped to the first dimension of the vector, the load capacity W is mapped to the second dimension of the vector, and so on, to generate the equipment attribute vector.

[0062] Step S134-2: Perform texture feature extraction on the obstacle material attributes in the node attribute feature set to generate obstacle attribute vectors.

[0063] Texture features are extracted from the material properties of obstacles. Texture features are extracted by analyzing the grayscale variation patterns and color distribution rules of pixels on the obstacle surface. For example, a gray-level co-occurrence matrix is ​​calculated to describe the spatial relationships between pixels. Contrast, correlation, entropy, and other feature values ​​are extracted from the gray-level co-occurrence matrix, and these feature values ​​are then combined into a vector to generate the obstacle attribute vector.

[0064] Step S134-3: Perform vectorization encoding on the personnel status parameters in the node attribute feature set to generate personnel attribute vectors.

[0065] For personnel status parameters, including personal protective equipment status parameters and personnel vital sign status parameters, the personal protective equipment status parameters, such as whether the seat belt is fastened (which can be represented as 0 or 1, 0 indicating not fastened and 1 indicating fastened) and whether the safety helmet is worn correctly (also represented as 0 or 1), and the personnel vital sign status parameters, such as heart rate H and blood pressure B, are vectorized and encoded. These parameters are arranged in a set order. For example, the seat belt fastened status is set as the first element of the vector, the safety helmet worn correctly is set as the second element, the heart rate H is normalized (let's say the normalization function is Norm, normalizing the heart rate H to between 0 and 1, i.e., H_norm=Norm(H)) is set as the third element, the blood pressure B is normalized similarly (let's say B_norm=Norm(B)) is set as the fourth element, and so on, to generate a personnel attribute vector. The normalization process here is to ensure that parameters of different dimensions are processed on the same scale, ensuring the uniformity of dimensions and the matching of feature dimensions in subsequent calculations.

[0066] Step S134-4: Map the device attribute vector, obstacle attribute vector, and personnel attribute vector to a unified feature dimension space through a fully connected layer, and then concatenate them according to the channel dimension to generate a node feature matrix.

[0067] Equipment attribute vectors, obstacle attribute vectors, and personnel attribute vectors may have inconsistent initial dimensions due to the different parameters they contain. To effectively fuse them, they need to be mapped to a unified feature dimension space.

[0068] Taking the device attribute vector as an example, let its initial dimension be d1. It is passed through a fully connected layer containing multiple neurons, each connected to each element of the device attribute vector, and weighted using a weight matrix W1. For each element in the device attribute vector, it is multiplied by the corresponding weight in the weight matrix W1 and summed, then processed by an activation function (let's call it Act), such as Act(Σ(Wi1*ei1)), where Wi1 is the weight in the weight matrix W1 and ei1 is the element in the device attribute vector. This results in a new vector after processing by the fully connected layer, with a dimension of D. The same method is applied to the obstacle attribute vector and the personnel attribute vector, passing them through fully connected layers and using weight matrices W2 and W3 respectively, to obtain new vectors with the same dimension D.

[0069] Then, the three vectors processed by the fully connected layer are concatenated along the channel dimension. Assuming the processed device attribute vector is V1, the processed obstacle attribute vector is V2, and the processed personnel attribute vector is V3, they are arranged in order, such as [V1, V2, V3], to generate the node feature matrix.

[0070] Step S134-5: Calculate the spatial correlation degree between the device node and the obstacle node based on the actual distance value of the first type of edge connection relationship, and generate the first type of edge weight matrix.

[0071] For the actual distance between a device node and an obstacle node in a first-type edge connection relationship, let the device node be Ni, the obstacle node be Nj, and the distance between them be dij. To calculate the spatial correlation, a distance-related function (let's call it Func1) is used. This function returns a value representing the spatial correlation based on the distance. For example, Func1(dij) = 1 / (1+dij), where a smaller distance indicates a greater spatial correlation. For each pair of device nodes and obstacle nodes, the spatial correlation value is calculated using this function. These spatial correlation values ​​are arranged in the order of the node pairs to form a matrix, namely the first-type edge weight matrix. For example, if there are n device nodes and m obstacle nodes, the size of the first-type edge weight matrix is ​​n×m, and the element wij = Func1(dij) in the first-type edge weight matrix represents the spatial correlation weight between device node Ni and obstacle node Nj.

[0072] Step S134-6: Calculate the interaction strength value between the personnel location node and the device node based on the operation correlation parameter of the second type of edge connection relationship, and generate the second type of edge weight matrix.

[0073] For the operational correlation parameters between personnel location nodes and equipment nodes in the second type of edge connection relationship, let the personnel location node be Nk and the equipment node be Ni. The operational correlation parameters include operation frequency fki, operation duration tki, etc. A comprehensive function (denoted as Func2) is used to calculate the interaction strength value. For example, Func2(fki, tki) = α*fki + β*tki, where α and β are weight coefficients set according to the actual situation to balance the influence of operation frequency and operation duration on the interaction strength. For each pair of personnel location nodes and equipment nodes, the interaction strength value is calculated by this function. These interaction strength values ​​are arranged in the order of node pairs to form a matrix, namely the second type of edge weight matrix. Assuming there are p personnel location nodes and n equipment nodes, the size of the second type of edge weight matrix is ​​p×n, and the element zki = Func2(fki, tki) in the matrix represents the interaction strength weight between personnel location node Nk and equipment node Ni.

[0074] Step S134-7: Perform graph topology mapping between the node feature matrix and the first type of edge weight matrix and the second type of edge weight matrix to generate initial graph structure data.

[0075] Each element in the node feature matrix corresponds to a node in the graph structure, with each row of the node feature matrix representing the feature vector of a node. Elements in the first and second type edge weight matrices correspond to the edges between corresponding node pairs in the graph structure. For example, the i-th row of the node feature matrix corresponds to node Ni in the graph structure; the element wij in the i-th row and j-th column of the first type edge weight matrix corresponds to the weight of the edge between equipment node Ni and obstacle node Nj; and the element zki in the k-th row and i-th column of the second type edge weight matrix corresponds to the weight of the edge between personnel location node Nk and equipment node Ni. This correspondence maps the matrix data to the graph structure, generating the initial graph structure data. During this process, it is crucial to ensure the accuracy of the correspondence between node features and edge weights to reflect the actual relationships between elements in the working environment.

[0076] Step S134-8: Perform reachability verification between nodes on the initial graph structure data, delete edge connections that do not meet the preset connectivity threshold, supplement missing key edge connections, and generate a fully connected working environment graph structure.

[0077] The initial graph structure data undergoes node reachability verification, checking whether any node can be connected to all other nodes via a series of edges. A preset connectivity threshold T is set. For each edge connection, its weight is used to determine if it meets the connectivity requirements. For example, if an edge's weight is below the threshold T, the connection is considered weak and may contribute little to the overall graph connectivity, thus it is deleted. Simultaneously, by analyzing the logical relationships between nodes and the actual needs of the operational scenario, any missing critical edge connections are identified. For instance, in a wind turbine maintenance scenario, if a device node has no direct connection to another device node that significantly impacts its operation, but the operational logic dictates a connection, this edge connection needs to be added. By continuously adjusting edge connections, a fully connected operational environment graph structure is ultimately generated, ensuring that the graph accurately reflects the connectivity and interrelationships between elements in the operational environment.

[0078] Step S134-9: Perform dynamic incremental updates on the fully connected operation environment graph structure according to the real-time update frequency of the equipment operating parameters to generate a dynamic operation environment graph structure with a timestamp version identifier.

[0079] Equipment operating parameters change in real time as the operation progresses, with an update frequency set to f_update. Whenever equipment operating parameters are updated, the corresponding parameter values ​​in the node attribute feature set are updated first. For example, if the lifting speed of the maintenance hoist changes, the parameter value representing the lifting speed in the node attribute feature set is modified. Then, based on the updated equipment operating parameters, relevant quantities are recalculated. For instance, the actual distance to obstacle nodes is recalculated based on the new position of the equipment node, and the weights of the first-type edge connections are recalculated using the same method as before for calculating the first-type edge weight matrix. For a fully connected work environment graph structure, a dynamic incremental update approach is adopted, meaning only the affected parts are updated, rather than the entire graph structure is reconstructed. During the update process, a timestamp version identifier is added to each update, such as recording the update time t1, t2, etc., generating a dynamic work environment graph structure with timestamp version identifiers. This records the changes in the work environment graph structure over time, facilitating subsequent analysis and decision-making.

[0080] Step S135: In the initial convolutional layer of the graph convolutional network, feature aggregation operation is performed on the direct neighboring nodes of each node in the working environment graph structure to generate a primary aggregated feature vector.

[0081] In the initial convolutional layer of the graph convolutional network, for each node in the job environment graph structure, let's call it N. Its direct neighbors are nodes directly connected to it via edges. For node N, we collect the feature vectors of its direct neighbors, denoted as N1, N2, ..., Nx, with feature vectors V1, V2, ..., Vx respectively. Simultaneously, we obtain the weights of the edges between node N and these direct neighbors, denoted as w1, w2, ..., wx. An aggregation method (let's call it Aggregate1), such as weighted summation, is used to calculate the primary aggregated feature vector V_primary. That is, V_primary = Aggregate1(w1*V1 + w2*V2 + ... + wx*Vx). The weighting here considers the importance of different neighboring nodes to the current node; the larger the edge weight, the greater the contribution of the corresponding neighboring node's features in the aggregation. In this way, a primary aggregated feature vector is generated for each node in the job environment graph structure. This primary aggregated feature vector contains the comprehensive feature information of the node and its direct neighbors.

[0082] Step S136: Perform cross-level feature propagation operation on the indirectly adjacent nodes with two edge connections in the intermediate convolutional layer of the graph convolutional network to generate intermediate aggregated feature vectors.

[0083] In the intermediate convolutional layers of a graph convolutional network, we consider the indirect adjacency relationships between nodes in the working environment graph structure, specifically referring to indirect adjacent nodes separated by two edges. Let the current node be N, and its indirect adjacent nodes separated by two edges be M1, M2, ..., My.

[0084] First, determine the paths from node N to these indirectly adjacent nodes. For example, from node N, reach the indirectly adjacent node M through intermediate node N'. For each indirectly adjacent node M, obtain its feature vector VM and the edge weights on the path from node N to M. Let the edge weight from node N to intermediate node N' be w1, and the edge weight from intermediate node N' to indirectly adjacent node M be w2. Use a cross-level feature propagation method (denoted as Propagate1), for example, by multiplying weights and transforming feature vectors, to propagate the features of the indirectly adjacent nodes to the current node N. That is, for each indirectly adjacent node M, calculate the propagated feature vector VM' = Propagate1(w1*w2*VM).

[0085] Then, the feature vectors propagated from all indirectly adjacent nodes are aggregated (let's call the aggregation method Aggregate2), for example, through weighted summation or other suitable aggregation methods, to generate an intermediate aggregated feature vector V_intermediate. That is, V_intermediate = Aggregate2(VM1'+VM2'+...+VMy'). This intermediate aggregated feature vector contains the comprehensive influence feature information of indirectly adjacent nodes with two edge connections on the current node, enriching the feature representation of the node.

[0086] Step S137: In the final convolutional layer of the graph convolutional network, the primary aggregated feature vector and the intermediate aggregated feature vector are fused to generate an environment state feature vector containing global dependencies.

[0087] In the final convolutional layer of the graph convolutional network, the primary aggregated feature vector generated by the initial convolutional layer and the intermediate aggregated feature vector generated by the intermediate convolutional layer are fused. Let the primary aggregated feature vector be V_primary and the intermediate aggregated feature vector be V_intermediate. A fusion method (let's call it Fuse1) is used, such as concatenation or weighted addition (the weighting coefficients are determined according to the actual situation). If concatenation is used, the primary aggregated feature vector and the intermediate aggregated feature vector are concatenated together in sequence to form a new vector V_fused.

[0088] Then, the fused vector is further processed through a fully connected layer (with the weight matrix denoted as W_fuse). After passing through an activation function (Act_fuse), the environment state feature vector V_environment, which contains global dependencies, is obtained. That is, V_environment = Act_fuse(W_fuse * V_fused). This environment state feature vector integrates the feature information of the direct adjacent nodes and the indirect adjacent nodes connected by two edges, reflecting the global dependencies between nodes in the job environment graph structure.

[0089] Step S138: Perform dimensionality reduction and visualization processing on the spatial coordinate encoding information embedded in the environmental state feature vector to generate an environmental risk heat map and mark the coordinates of high-risk areas.

[0090] The environmental state feature vector contains the spatial coordinate encoding information of each element in the working environment, denoted as S. Since this information may have a high dimensionality, making it unsuitable for intuitive analysis, dimensionality reduction and visualization are required. A dimensionality reduction algorithm (denoted as Reduce1), such as a method similar to Principal Component Analysis (PCA), is used to convert the high-dimensional spatial coordinate encoding information S into a low-dimensional representation S_low.

[0091] Then, an environmental risk heatmap is generated based on the dimensionality-reduced information. For each location in the space, a color value is assigned to represent the risk level based on its corresponding dimensionality-reduced feature value; the darker the color, the higher the risk. Simultaneously, high-risk areas are identified by analyzing the color distribution in the heatmap. Let T_risk be the threshold for determining high-risk areas. For areas in the heatmap whose color values ​​exceed the threshold T_risk, their coordinates are determined and marked. For example, the top-left and bottom-right coordinates of the high-risk area are marked to clearly indicate its location within the work environment.

[0092] Step S139: Perform spatial clustering analysis on the coordinate data of the environmental risk heat map to identify risk clustering areas and generate area boundary labeling information.

[0093] Spatial clustering analysis is performed on the coordinate data of the environmental risk heatmap using a spatial clustering algorithm (denoted as Cluster1), such as a method similar to DBSCAN (Density-Based Spatial Clustering of Applications with Noise). This algorithm divides points in space into different clusters based on the distance and density relationships between coordinate data points. For each cluster, it is determined whether it is a risk clustering region. If the points in a cluster generally correspond to high risk values ​​in the heatmap, and the cluster size reaches a certain level (let's say the minimum cluster size threshold is Min_size), then the cluster is considered a risk clustering region. For each risk clustering region, its boundary is determined. For example, the boundary is determined by calculating the convex hull of the outermost points in the cluster, generating boundary labeling information. The labeling information includes the vertex coordinates of the boundary and a shape description of the boundary, to accurately represent the range and shape of the risk clustering region.

[0094] Step S1390: Update the node attribute feature set of the operation environment map structure based on the region boundary annotation information to generate an optimized environment state feature vector.

[0095] Based on the boundary labeling information of the risk cluster area, update the node attribute feature set of the work environment map structure. For nodes located within the risk cluster area, or nodes closely related to the risk cluster area, adjust their node attribute features. For example, if a device node is located within the risk cluster area, its risk-related attribute values ​​may be increased. Let the node attribute feature set be Attributes. For a related node Ni, based on the situation of the risk cluster area, adjust its attributes through an adjustment function (let's call it Adjust1), i.e., Attributes[Ni] = Adjust1(Attributes[Ni], boundary labeling information).

[0096] Then, based on the updated node attribute feature set, the relevant operations in the graph convolutional network are recalculated, such as recalculating node feature vectors and edge weights. Through the same graph convolutional network processing flow as before, from the initial convolutional layer to the final convolutional layer, an optimized environment state feature vector is generated. This optimized environment state feature vector more accurately reflects the impact of risk distribution in the working environment on each node.

[0097] Step S140: Generate a dynamic emergency dispatch strategy based on the environmental state feature vector and the set of operator behavior features.

[0098] In the scenario of high-altitude maintenance of large wind turbine generators, a dynamic emergency dispatch strategy is generated based on the obtained environmental state feature vector and the set of operator behavior features to adapt to the actual situation at the work site.

[0099] Step S141: After aligning the environmental state feature vector and the set of worker behavior features to the same feature dimension by inputting them into the linear projection layer, input them into the pre-trained spatiotemporal association model, and perform feature interaction verification on the node attribute features in the environmental state feature vector and the dynamic parameters in the set of worker behavior features through the cross attention mechanism to generate an interaction matrix.

[0100] The environmental state feature vector and the set of worker behavior features may have different feature dimensions due to their different sources and compositions. First, let's denote the environmental state feature vector as V_env and the set of worker behavior features as Set_behave. Then, input them into a linear projection layer. For the environmental state feature vector V_env, the linear projection layer performs a weighted calculation using a weight matrix W1, i.e., V_env_projected = W1 * V_env, making its dimension the same as the dimension D of the processed set of worker behavior features.

[0101] For the set of behavioral features of workers, Set_behave, which contains multiple feature vectors such as V1_behave, V2_behave, ..., Vn_behave, each feature vector is passed through a linear projection layer and weighted by a weight matrix W2 to obtain V1_behave_projected=W2*V1_behave, V2_behave_projected=W2*V2_behave, ..., Vn_behave_projected=W2*Vn_behave. These projected feature vectors are then combined into a vector V_behave_projected with the same dimensions as the projected environmental state feature vectors.

[0102] Then, the projected environmental state feature vector (V_env_projected) and the projected worker behavior feature set (V_behave_projected) are input into the pre-trained spatiotemporal correlation model. This model validates the feature interaction between the two through a cross-attention mechanism. For each node attribute feature in the environmental state feature vector, denoted as Attr_env, and each dynamic parameter in the worker behavior feature set, denoted as Param_behave, the spatiotemporal correlation model calculates the attention weights between them. For example, using an attention calculation function (denoted as Attention1), the attention weight w = Attention1(Attr_env, Param_behave) between Attr_env and Param_behave is calculated. The attention weights between all node attribute features and dynamic parameters are combined into a matrix, generating the interaction matrix Matrix_interaction. This interaction matrix reflects the degree of correlation between the environmental state and worker behavior.

[0103] Step S142: Based on the association weight value between each node and behavioral feature in the interaction matrix, extract the risk correlation parameter between high-risk nodes and abnormal personnel behavior features in the work environment graph structure.

[0104] From the interaction matrix Matrix_interaction, for each node in the work environment graph structure, denoted as Nj, find its corresponding association weight value. These association weight values ​​reflect the degree of association between the node and the behavioral characteristics of the workers. For those nodes marked as high-risk nodes (e.g., nodes in high-risk areas identified in the previously generated environmental risk heatmap and subsequent analysis), denoted as Set_high-risk, extract their association weight values ​​with each behavioral characteristic in the set of worker behavioral characteristics from the interaction matrix.

[0105] For each behavioral feature in the set of worker behavioral features, denoted as Behave_attr, the association weight value between each node Nj in the high-risk node set Set_high-risk and Behave_attr is obtained from the interaction matrix, denoted as wNj_Behave_attr. By analyzing and calculating these association weight values, risk correlation parameters are extracted. For example, a comprehensive calculation method (denoted as Compute1) can be used to summarize and calculate the association weight values ​​between all nodes in the high-risk node set Set_high-risk and a certain behavioral feature Behave_attr, such as calculating their average or weighted average (if different weights are assigned based on factors such as the importance of the node). Let the weight of node Nj be αNj. Then, for each node Nj in Set_high-risk, that is, for each node Nj in Set_high-risk, Risk_correlation_Behave_attr=Compute1 (Σ(αNj*wNj_Behave_attr) , we obtain the set of risk correlation parameters between high-risk nodes and abnormal behavior characteristics of personnel in the work environment graph structure. Each parameter corresponds to a behavior characteristic of the worker, reflecting the degree of risk correlation between high-risk nodes and corresponding behavior characteristics.

[0106] Step S143: Perform an adaptive scaling operation on the node spatial coordinate features in the environmental state feature vector according to the risk correlation parameter to generate a risk-weighted spatial distribution feature vector.

[0107] For the environmental state feature vector V_env, which contains node spatial coordinate features, denoted as Coords, an adaptive scaling operation is performed on the spatial coordinates of each node in Coords based on the previously calculated risk correlation parameter set. Let the risk correlation parameter set be Set_Risk_correlation, and for each node coordinate in Coords, denoted as Coord_i, find its corresponding risk correlation parameter Risk_correlation_i (e.g., determined by node identifier or some kind of correspondence).

[0108] A scaling function (denoted as Scale1) is used to scale the node coordinates based on the risk correlation parameter. For example, for two-dimensional coordinates (x, y), the new coordinates (x_new, y_new) = Scale1(x, y, Risk_correlation_i). The scaling function Scale1 may scale the coordinates to different degrees in different directions depending on the value of the risk correlation parameter. If the risk correlation parameter is large, it indicates a high risk correlation between the area where the node is located and abnormal human behavior characteristics, and the coordinates may be scaled more significantly to highlight the importance of the area in the spatial distribution. By performing this operation on the coordinates of all nodes in Coords, a risk-weighted spatial distribution feature vector V_risk_weighted is generated. This spatial distribution feature vector reflects the relationship between different areas and risk correlation in spatial distribution.

[0109] Step S144: Input the risk-weighted spatial distribution feature vector and the edge connection relationship of the work environment graph structure into the graph attention network, and generate a topology optimization feature map containing emergency obstacle avoidance paths between nodes through the edge weight dynamic adjustment algorithm.

[0110] The risk-weighted spatial distribution feature vector V_risk_weighted and the edge connections of the work environment graph structure (including the complete edge connection information consisting of the first type of edge connection and the second type of edge connection) are input into the graph attention network. The graph attention network dynamically adjusts the edge weights in the graph structure based on the input information.

[0111] For each edge in the graph structure, denoted as Edge_ij, its initial edge weight is w_ij. The graph attention network recalculates the edge weights using a dynamic edge weight adjustment algorithm (denoted as Adjust_weight1), combining information from nodes in the risk-weighted spatial distribution feature vector V_risk_weighted with the characteristics of the nodes at both ends of the edge. For example, the algorithm Adjust_weight1 considers factors such as the positional relationship of the nodes at both ends of the edge in the risk-weighted spatial distribution, the node's own attribute characteristics, and its association with other nodes, to calculate a new edge weight w_ij_new = Adjust_weight1(V_risk_weighted, Edge_ij, w_ij).

[0112] A new graph structure is generated by adjusting all edge weights. Emergency obstacle avoidance paths between nodes are then sought within this new graph structure. A path search algorithm (let's call it Search_path1), such as one similar to the A* algorithm, is used to search for the shortest or optimal emergency obstacle avoidance path from one node to other nodes based on the adjusted edge weights and connectivity between nodes. This path information is then integrated into a feature map to generate a topology-optimized feature map, Feature_map_topology, containing emergency obstacle avoidance paths between nodes. This topology-optimized feature map demonstrates the emergency obstacle avoidance path information between nodes in the operational environment graph structure after considering the risk-weighted spatial distribution and dynamic adjustment of edge weights.

[0113] Step S145: Based on the path connectivity parameters in the topology optimization feature graph, Dijkstra's algorithm is called in combination with the edge weights of the graph attention network to calculate the shortest safe movement path sequence of the work terminal equipment in the work environment graph structure, and the equipment motion parameter constraints corresponding to the shortest safe movement path sequence are generated.

[0114] Path connectivity parameters are extracted from the Feature_map_topology, which describes the connectivity, length, and risk of each path in the graph structure. Let the set of path connectivity parameters be Set_connectivity.

[0115] The Dijkstra algorithm is invoked, combined with edge weights adjusted using a graph attention network (let the edge weight matrix be W_edges), to calculate the shortest safe movement path sequence from the current location to the target location (e.g., a safe zone or a location where emergency operations need to be performed) within the operational environment graph structure. The Dijkstra algorithm starts from the initial node and progressively explores the nodes in the graph structure, calculating the distance to each node based on the edge weights (this distance can be understood as the path cost from the initial node to that node, which is related to factors such as risk and distance).

[0116] For each node in the graph structure, denoted as Node_k, during execution, the shortest path distance from the starting node to Node_k (denoted as Distance_k) and the predecessor node on the path (denoted as Predecessor_k) are recorded. By continuously updating these values, the shortest safe movement path sequence from the starting node to the target node is finally determined. Let the shortest safe movement path sequence be Sequence_path, which consists of a series of nodes, for example, Sequence_path=[Node_start, Node_1, Node_2, ..., Node_end].

[0117] Based on the positional relationship of the nodes in the shortest safe movement path sequence Sequence_path and the actual situation of the operation environment, generate the constraint conditions for the device movement parameters corresponding to this path sequence. For example, consider factors such as the distance between nodes, the angle change, and the physical limitations of the device. Let the device movement parameters include speed V, acceleration A, steering angle θ, etc. For adjacent nodes Node_i and Node_i+1 in the path sequence, according to their spatial positional relationship, calculate the approximate speed range, acceleration range, and steering angle range required for the device to move from Node_i to Node_i+1. For example, by calculating the distance d = Distance(Node_i, Node_i+1) between two points and the desired movement time t, the speed range V_min <= V <= V_max can be obtained, where V_min and V_max are determined based on factors such as the distance d, time t, and device performance. The acceleration range and steering angle range are also obtained through similar calculation methods based on the node positional relationship and device characteristics. Organize these constraint conditions for the device movement parameters into a set, denoted as Set_constraints.

[0118] Step S146: Perform dynamic compatibility verification on the constraint conditions of the device movement parameters and the real-time device operation parameters, and eliminate the path segments that exceed the physical limits of the device according to the verification results, and generate a dynamic emergency scheduling strategy including the feasible path coordinates and the device control instruction set.

[0119] Obtain the real-time device operation parameters, denoted as Set_real-time, which includes parameters such as the current speed V_real, acceleration A_real, and steering angle θ_real of the device. Perform dynamic compatibility verification on the set of device movement parameter constraints Set_constraints and the real-time device operation parameters Set_real-time.

[0120] For each device movement parameter constraint in Set_constraints, let the speed constraint be V_min <= V <= V_max, the acceleration constraint be A_min <= A <= A_max, and the steering angle constraint be θ_min <= θ <= θ_max. Compare the speed V_real, acceleration A_real, and steering angle θ_real in the real-time device operation parameters with these constraints respectively.

[0121] If a certain parameter in the real-time device operation parameters exceeds the corresponding constraint range, for example, V_real < V_min or V_real > V_max, it means that the current path segment may pose risks to the device or exceed the physical limits of the device. For the path segments that exceed the physical limits of the device, eliminate them from the shortest safe movement path sequence Sequence_path.

[0122] After verifying and eliminating all path segments, a path sequence containing the coordinates of feasible paths is obtained, denoted as `Sequence_feasible_path`. Based on this feasible path sequence and the equipment motion parameter constraints, a set of equipment control instructions is generated. For example, for each path segment in the feasible path sequence, corresponding speed control instructions, acceleration control instructions, and steering angle control instructions are generated according to their corresponding equipment motion parameter constraints. Let the equipment control instruction set be `Set_commands`, which contains a series of control instructions for different motion parameters of the equipment.

[0123] By combining feasible path coordinates and equipment control command sets, a dynamic emergency dispatch strategy, Strategy_emergency, is generated that includes feasible path coordinates and equipment control command sets. This dynamic emergency dispatch strategy takes into account the real-time operating status and physical limits of the equipment, and provides specific guidance for the safe movement of work terminal equipment in emergency situations.

[0124] Step S150: Perform an effectiveness verification operation on the dynamic emergency dispatch strategy to obtain the strategy verification result, generate an emergency dispatch instruction set based on the strategy verification result, and transmit the emergency dispatch instruction set to the work terminal equipment to trigger the emergency response operation.

[0125] In the scenario of high-altitude maintenance of large wind turbine generators, the effectiveness of the generated dynamic emergency dispatch strategy Strategy_emergency is verified to ensure that the dynamic emergency dispatch strategy can be effectively executed and is safe and reliable in practical applications.

[0126] Step S151: Extract the feasible path coordinate sequence and equipment motion parameter constraints from the dynamic emergency dispatch strategy, and construct a three-dimensional dynamic scene reconstruction model based on the spatial coordinate system of the operation environment map structure.

[0127] Extract the feasible path coordinate sequence Sequence_feasible_path and the equipment motion parameter constraint set Set_constraints from the dynamic emergency dispatch strategy Strategy_emergency. Based on the spatial coordinate system defined by the operational environment graph structure (let the spatial coordinate system be Coordinate_system, which determines the position representation of each point in the operational environment), construct a three-dimensional dynamic scene reconstruction model Model_reconstruction.

[0128] When constructing the model, the nodes in the work environment diagram structure (including work equipment nodes, environmental obstacle nodes, and personnel location nodes) are first arranged according to their coordinate positions in the spatial coordinate system. For work equipment nodes, a corresponding 3D model representation is created in space based on their corresponding equipment attributes (such as equipment size and shape). For example, if it is a maintenance suspended platform equipment node, a 3D suspended platform model is created at the corresponding coordinate position based on the actual size and shape of the platform. Similarly, for environmental obstacle nodes, a corresponding 3D obstacle model is created based on their geometric features and material properties.

[0129] Then, based on the feasible path coordinate sequence (Sequence_feasible_path), the expected movement path of the terminal equipment is marked in the 3D scene. For each coordinate point in the path sequence, the corresponding position in the 3D scene is found, and they are connected by lines or other visualization methods to form a representation of the equipment's movement path. Simultaneously, the equipment motion parameter constraints (Set_constraints) are associated with the equipment model in the 3D scene, providing parameter basis for subsequent simulation operations. In this way, a 3D dynamic scene reconstruction model (Model_reconstruction) that reflects the working environment and the expected movement of the equipment is constructed, providing a visual and operable virtual environment for subsequent strategy verification.

[0130] Step S152: Input the feasible path coordinate sequence and device motion parameter constraints into the preloaded physics engine simulator, segment the simulation task based on the path curvature key points, execute the virtual motion simulation operation in the segmented parallel computing mode, and generate simulation motion trajectory data and virtual device state parameter set.

[0131] The extracted feasible path coordinate sequence (Sequence_feasible_path) and the device motion parameter constraint set (Set_constraints) are input into the preloaded physics engine simulator (let's call the physics engine simulator "Simulator"). The physics engine simulator can simulate the motion of objects in the physical environment.

[0132] First, the feasible path coordinate sequence is analyzed to identify key points of path curvature. Let Curvature1 be the path curvature calculation function; this function calculates the curvature value for each point on the path. Points with significant changes in curvature values ​​are identified as key points of path curvature. Let the set of key points of path curvature be Set_key_points.

[0133] The simulation task is segmented based on the path curvature keypoints (Set_key_points). The path is divided into multiple segments, designated as Segment_1, Segment_2, ..., Segment_n, with each segment bounded by a path curvature keypoint.

[0134] Using a segmented parallel computing model, virtual motion simulation is performed on each path segment in the physics engine simulator. For each path segment Segment_i, the initial motion state and motion parameters of the device are set in the physics engine simulator according to the parameters in the device motion parameter constraints Set_constraints, such as velocity, acceleration, and steering angle. For example, for the velocity constraint V_min<=V<=V_max, the initial velocity of the device is set to V_start at the starting point of the path segment, where V_min<=V_start<=V_max.

[0135] In the physics engine simulator, the motion of the device along the path segment is simulated according to physical laws and set parameters. During the motion, the device's position, velocity, acceleration, and other state information are recorded to generate simulated motion trajectory data (Trajectory_data). Simultaneously, other relevant state parameters of the device during motion, such as its attitude and forces, are recorded to generate a virtual device state parameter set (Set_virtual_state). Through parallel simulation operations on all path segments, the simulated motion trajectory data and the virtual device state parameter set for the entire feasible path are obtained.

[0136] Step S153: Perform a spatial collision detection operation on the simulated motion trajectory data and the obstacle node coordinates in the working environment map structure to generate a set of trajectory conflict point coordinates and a conflict timestamp sequence.

[0137] Spatial collision detection is performed between the simulated motion trajectory data (Trajectory_data) and the coordinates of obstacle nodes in the operational environment graph structure. For each point in the simulated motion trajectory data, designated as Point_trajectory, its coordinate position in the spatial coordinate system is obtained. For each obstacle node in the operational environment graph structure, designated as Obstacle_node, its coordinate position and obstacle size information (based on the obstacle's geometric features and size information determined during the previous construction of the operational environment graph structure) are obtained.

[0138] For example, a spatial collision detection algorithm (denoted as Collision_detect1) can be used to determine whether a point collides with an obstacle based on the point's coordinates and the obstacle's position and size information. For instance, if the point is located within the space occupied by the obstacle's geometry, a collision is determined to have occurred.

[0139] For each point where a collision occurs, its coordinates are recorded, generating a set of conflict point coordinates called `Set_conflict_points`. Simultaneously, the time of the collision is recorded (based on time information from the simulated motion trajectory data), generating a conflict timestamp sequence called `Sequence_conflict_timestamps`. The set of conflict point coordinates and the conflict timestamp sequence accurately pinpoint the specific location and time of the collision between the device's motion trajectory and the obstacle during the simulation.

[0140] Step S154: Perform a dynamic comparison operation between the set of virtual device status parameters and the physical limit threshold of the device operating parameters to generate an identifier of the device's out-of-limit operating parameters and the corresponding time interval.

[0141] The parameters in the virtual device state parameter set `Set_virtual_state` are dynamically compared with the physical limit thresholds of the device's operating parameters. These physical limit thresholds are determined based on the device's design specifications and safety standards; for example, the physical limit threshold for speed is `V_limit`, for acceleration is `A_limit`, and for steering angle is `θ_limit`, etc.

[0142] For the velocity parameter V_virtual, acceleration parameter A_virtual, and steering angle parameter θ_virtual in the virtual device state parameter set, each is compared with its corresponding physical limit threshold. If V_virtual > V_limit, the velocity parameter is marked as exceeding the limit, an over-limit operation parameter flag (denoted as Flag_V_overlimit) is generated, and the time interval of the velocity over-limit occurrence is recorded (based on the time information in the simulation motion trajectory data, the start and end times of the velocity over-limit are determined and denoted as Time_start_V_overlimit and Time_end_V_overlimit). Similarly, similar comparison and recording operations are performed for parameters such as acceleration and steering angle.

[0143] Through this dynamic comparison operation, a set of flags indicating that the equipment is operating beyond its limits (Set_flags_overlimit) and a set of time intervals (Set_time_intervals) corresponding to each flag are generated. These flags and time intervals can clearly reflect whether the equipment's operating parameters exceed physical limits during the simulation process and the specific time range in which the exceedance occurs.

[0144] Step S155: Integrate the set of coordinates of the trajectory conflict points, the conflict timestamp sequence, and the equipment over-limit operation parameter identifier to generate a strategy verification anomaly report, and extract the coordinates of the abnormal path segment and the abnormal equipment control command.

[0145] The set of trajectory conflict point coordinates (Set_conflict_points), the sequence of conflict timestamps (Sequence_conflict_timestamps), and the set of equipment overlimit operation parameter identifiers (Set_flags_overlimit) are merged to generate a policy verification anomaly report (Report_abnormality). The report details the coordinates of the trajectory conflict points, the timestamps of the conflicts, and the identifiers of the equipment overlimit operation parameters.

[0146] Based on the set of coordinates of trajectory conflict points and the sequence of conflict timestamps, abnormal path segments are identified. For example, if a path segment between two trajectory conflict points poses a collision risk, then that path segment is identified as an abnormal path segment. The coordinate information of the abnormal path segments is extracted, and the set of abnormal path segment coordinates is denoted as Set_abnormal_path_coords.

[0147] Simultaneously, based on the equipment's out-of-limit operation parameter identifiers and corresponding time intervals, as well as the relationship between equipment motion parameter constraints and equipment control command sets, abnormal equipment control commands that may lead to equipment out-of-limit operation are extracted. Let the set of abnormal equipment control commands be denoted as `Set_abnormal_commands`. The strategy verification anomaly report, the set of abnormal path segment coordinates, and the set of abnormal equipment control commands provide specific problem information for subsequent adjustments and optimizations to the dynamic emergency dispatch strategy.

[0148] Step S156: Perform piecewise smooth interpolation on the feasible path coordinate sequence based on the coordinates of the abnormal path segment to generate an alternative path coordinate sequence and recalculate the device motion parameter constraints.

[0149] Based on the set of abnormal path segment coordinates, `Set_abnormal_path_coords`, the feasible path coordinate sequence `Sequence_feasible_path` is processed by segmented smooth interpolation. For each abnormal path segment, its starting point coordinates are set as `Coord_start` and its ending point coordinates as `Coord_end`.

[0150] A piecewise smoothing interpolation algorithm (denoted as Interpolation1) is employed to insert a series of new coordinate points between the start and end points of abnormal path segments to smooth the path. For example, the coordinates of the inserted points are calculated using spline interpolation based on the positions of the start and end points and the overall trend of the path. Let the set of inserted coordinate points be Set_inserted_points.

[0151] These insertion points are combined with the parts of the original feasible path coordinate sequence excluding abnormal path segments to generate the alternative path coordinate sequence Sequence_alternative_path.

[0152] Then, the device motion parameter constraints are recalculated based on the alternative path coordinate sequence Sequence_alternative_path. The geometric characteristics of the alternative path, such as path length and curvature variations, are considered, as well as the actual working environment, such as obstacle distribution and space constraints.

[0153] For adjacent coordinate points on the path, denoted as Coord_i and Coord_i+1, the approximate speed range is determined by calculating the distance between the two points, d = Distance(Coord_i, Coord_i+1), and the expected travel time, t. Assuming the equipment needs to safely and reasonably traverse this path within a certain time, based on the relationship between distance and time, a lower speed limit V_min and an upper speed limit V_max can be obtained, i.e., V_min <= V <= V_max. Here, V_min and V_max are determined comprehensively based on factors such as path length, equipment performance, and safety requirements.

[0154] For acceleration constraints, the curvature variation of the path is analyzed. A large curvature variation means the device requires significant acceleration to adjust its direction of motion when traversing that path segment. Assuming the path curvature is Curvature, a curvature-related calculation method (denoted as Acceleration_compute) is used to calculate the lower limit A_min and upper limit A_max of the acceleration based on the curvature value and parameters such as the device's steering capability. That is, A_min <= A <= A_max, ensuring the device can meet path requirements without exceeding its physical limits.

[0155] The steering angle constraint is also determined based on the geometry of the path. For the path direction change formed by adjacent coordinate points, the required steering angle range of the equipment is calculated. Let the path direction change angle from Coord_i to Coord_i+1 be Delta_theta. Based on the equipment's steering performance and safety requirements, the lower limit θ_min and the upper limit θ_max of the steering angle are determined, i.e., θ_min <= θ <= θ_max.

[0156] These recalculated constraints on device motion parameters, such as velocity, acceleration, and steering angle, are compiled into a new set, denoted as New_Set_constraints. This new set of device motion parameter constraints is generated based on the alternative path coordinate sequence and is more suitable for the adjusted path conditions.

[0157] Step S157: Input the alternative path coordinate sequence and the recalculated device motion parameter constraints into the physics engine simulator to perform a secondary verification operation. If the set of trajectory conflict point coordinates is empty and the device over-limit operation parameter identifier is not activated, then it is marked as a verification pass strategy.

[0158] The alternative path coordinate sequence Sequence_alternative_path and the recalculated set of device motion parameter constraints New_Set_constraints are then fed back into the physics engine simulator Simulator.

[0159] In the physics engine simulator, the process is similar to the initial verification. Based on the alternative path coordinate sequence, the new expected movement path of the device in the 3D scene is marked. The initial motion state and motion parameters of the device are set according to the recalculated device motion parameter constraints. For example, at the starting point of the path, a suitable initial velocity V_start is selected based on the lower velocity limit V_min and the upper velocity limit V_max, where V_min <= V_start <= V_max. Initial values ​​for acceleration and steering angle are also set to satisfy the corresponding constraints.

[0160] Then, the physics engine simulator simulates the movement of the equipment along the alternative path based on physical laws and set parameters. During the movement, a spatial collision detection operation is performed again to check whether the equipment's trajectory collides with obstacle nodes in the working environment map structure. At the same time, the operating parameters of the equipment are monitored to determine whether they exceed physical limit thresholds.

[0161] If, during the entire simulation, the set of coordinates for trajectory conflict points is empty (meaning no collisions were detected between the equipment's trajectory and obstacles), and the equipment's out-of-limit operating parameter flags are not activated (i.e., the equipment's speed, acceleration, steering angle, and other operating parameters do not exceed their physical limit thresholds), then the dynamic emergency dispatch strategy is marked as a verified successful strategy. This indicates that the adjusted dynamic emergency dispatch strategy, considering the operating environment and equipment physical limitations, can effectively guide the equipment to safely complete emergency operation path planning, demonstrating its feasibility for practical application.

[0162] Step S158: Generate a device control instruction encoding sequence based on the device motion parameter constraints in the verification strategy, and synchronize the device control instruction encoding sequence with the alternative path coordinate sequence by timestamp synchronization to generate an emergency dispatch instruction set.

[0163] For the set of device motion parameter constraints `New_Set_constraints` in the validated strategy, it is converted into a device control command encoding sequence. For the speed constraint `V_min<=V<=V_max`, the corresponding speed control command code is generated according to the device's control protocol and command encoding rules. For example, if the device's speed control command code uses a predefined binary encoding method, the encoding value representing the speed range under that encoding method is determined based on `V_min` and `V_max`. Similarly, for the acceleration constraint `A_min<=A<=A_max` and the steering angle constraint `θ_min<=θ<=θ_max`, corresponding acceleration control command codes and steering angle control command codes are generated respectively. These codes are arranged in a set order to form the device control command encoding sequence `Set_command_codes`.

[0164] Simultaneously, for the alternative path coordinate sequence `Sequence_alternative_path`, timestamp information is added to each coordinate point. Based on factors such as the device's expected speed along the path and the length of the path segment, the estimated time for the device to reach each coordinate point is calculated. Let the average speed of the device along the path be `V_avg`. For the `i`th coordinate point `Coord_i` on the path, the timestamp `T_i` of `Coord_i` is determined by calculating the distance `d_i` from the path's starting point to `Coord_i`, and the estimated travel time `t_i` = `d_i / V_avg` obtained from `V_avg`.

[0165] The device control command encoding sequence `Set_command_codes` and the alternative path coordinate sequence `Sequence_alternative_path` are synchronized and aligned according to timestamps. For each command code in the device control command encoding sequence, the path coordinate points within the corresponding timestamp range are found. For example, a speed control command code corresponds to the speed control of the device within a certain time period; the set of coordinate points on the device's movement path within that time range is found. These corresponding command codes and path coordinate points are combined to generate the emergency dispatch command set `Set_emergency_commands`. This emergency dispatch command set not only contains the control commands that the device should execute at different locations, but also ensures the temporal consistency between the commands and the path coordinates.

[0166] Step S159: Perform an encapsulation operation on the emergency dispatch instruction set, add the work terminal device address identifier and instruction priority tag, generate the target transmission data packet and send it to the work terminal device.

[0167] The emergency dispatch command set `Set_emergency_commands` is encapsulated. First, a job terminal device address identifier is added to each command in the set. This identifier uniquely identifies the job terminal device receiving the command; it could be the device's MAC address or a specific device number. Let's call this identifier `Device_address`. For each command `Command_i` in the emergency dispatch command set, `Device_address` is associated with `Command_i`. For example, a field is added to the command's data structure to store `Device_address`.

[0168] Next, a priority tag is added to each instruction. The priority tag is determined based on the urgency and importance of the instruction. For example, instructions involving emergency braking of equipment or avoiding major dangers have higher priority, while some routine adjustment instructions have relatively lower priority. Assuming instruction priorities are divided into multiple levels, such as high, medium, and low, a priority determination algorithm (denoted as Priority_determine) is used to determine the priority of each instruction based on factors such as the type of instruction, its impact on equipment safety and emergency operations, and other factors, and a corresponding priority tag (Priority_tag_i) is added to each instruction.

[0169] The emergency dispatch instructions, each with an added terminal device address identifier and instruction priority tag, are packaged to generate a target transmission data packet (Packet_target). The format of the data packet is defined according to the communication protocol and may include fields such as a header, instruction data, and a checksum. The header may contain information such as the data packet length and version number. The instruction data stores the emergency dispatch instructions with added identifiers and tags, and the checksum is used to verify the integrity of the data packet during transmission.

[0170] Finally, the target data packet (Packet_target) is sent to the work terminal device via wireless or wired network communication. During transmission, the corresponding communication protocol is followed to ensure accurate delivery of the data packet to the designated work terminal device, enabling the work terminal device to correctly receive and execute emergency dispatch instructions.

[0171] Step S1510: Update the feasible path coordinate sequence in the dynamic emergency dispatch strategy according to the instruction confirmation signal returned by the operation terminal device and the real-time location feedback data, and trigger the periodic strategy effectiveness verification operation.

[0172] After receiving the target data packet (Packet_target) and successfully parsing and executing the emergency dispatch command, the task terminal device will return a command confirmation signal to the system. Simultaneously, the task terminal device will acquire its own location information in real time through its own positioning system or other location detection methods and send the real-time location feedback data back to the system.

[0173] Upon receiving the instruction confirmation signal and real-time location feedback data from the operational terminal equipment, the feasible path coordinate sequence in the dynamic emergency dispatch strategy is first updated based on the real-time location feedback data. Assuming the real-time location feedback from the operational terminal equipment is Current_position, Current_position is compared with the feasible path coordinate sequence in the dynamic emergency dispatch strategy. If the deviation between Current_position and a coordinate point in the path sequence exceeds a set threshold (denoted as Deviation_threshold), it indicates that the actual location of the operational terminal equipment has deviated from the expected path.

[0174] A path adjustment algorithm (denoted as Path_adjust) is used to adjust the feasible path coordinate sequence based on the Current_position and the actual conditions of the working environment (such as obstacle positions, the status of other equipment, etc.). For example, if the working terminal equipment deviates from the expected path because it encounters a newly appeared obstacle, the path adjustment algorithm will replan the path, bypass the obstacle, and add new coordinate points to the feasible path coordinate sequence to guide the working terminal equipment back to a reasonable emergency operation path.

[0175] After updating the feasible path coordinate sequence, a periodic strategy effectiveness verification operation is triggered. Following the previous strategy effectiveness verification process, starting with building a 3D dynamic scene reconstruction model, virtual motion simulation, collision detection, and parameter comparison are performed sequentially to verify the effectiveness of the dynamic emergency dispatch strategy. This ensures that even if positional deviations occur during the actual execution of the emergency dispatch strategy by the operational terminal equipment, the strategy can be adjusted promptly and its effectiveness verified, guaranteeing that the emergency dispatch strategy always adapts to the actual conditions of the work site and ensuring the safe conduct of high-altitude operations.

[0176] Furthermore, in the above scenarios, personal safety emergency dispatch and work guidance functions for high-altitude operations can be provided, such as configurable smart safety helmets, vital sign monitoring wristbands, and management systems. The vital sign monitoring wristband collects heart rate, blood pressure, and blood oxygen data from the smart safety helmet wearer. The smart safety helmet includes video, voice intercom, SOS alarm, fall warning, and proximity detection modules. Before high-altitude operations, data collected by the vital sign monitoring wristband and the relevant detection modules of the smart safety helmet are used to periodically check whether the worker is suitable for high-altitude work. For example, if the heart rate collected by the vital sign monitoring wristband exceeds the normal range, or the proximity detection module of the smart safety helmet detects a strong electric field nearby, determining that the worker is not suitable for the current high-altitude operation, an alert is issued directly through the management system backend, and emergency handling is carried out using the voice and video functions of the smart safety helmet, such as notifying the worker to stop work and guiding them safely back to the ground. During high-altitude operations, on the one hand, the worker's status is continuously monitored; on the other hand, the video capture function of the smart safety helmet facilitates remote work guidance. The collected video data, as part of the real-time video data set, is used together with data collected by other cameras for subsequent processing.

[0177] Figure 2 The illustration shows exemplary hardware and software components of an aerial work emergency dispatch system 100 that can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the aerial work emergency dispatch system 100 and to perform the functions in this application.

[0178] The high-altitude operation emergency dispatch system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the high-altitude operation emergency dispatch method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0179] For example, the high-altitude work emergency dispatch system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the high-altitude work emergency dispatch system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The high-altitude work emergency dispatch system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0180] For ease of explanation, only one processor is described in the high-altitude work emergency dispatch system 100. However, it should be noted that the high-altitude work emergency dispatch system 100 of this application may also include multiple processors, and therefore the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the high-altitude work emergency dispatch system 100 performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0181] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned high-altitude operation emergency dispatch method is implemented.

[0182] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. An aerial work emergency dispatching method, characterized by, The method comprises: acquiring a real-time video data set of an aerial work area, performing a preprocessing operation on the real-time video data set to generate a standardized video data stream; calling a video analysis model to perform feature extraction on the standardized video data stream to generate a work environment feature set and a work personnel behavior feature set; constructing a work environment graph structure based on the work environment feature set, calling a graph convolution network to perform a multi-level feature aggregation operation on the work environment graph structure to generate an environment state feature vector; generating a dynamic emergency dispatch strategy according to the environment state feature vector and the work personnel behavior feature set; performing an effectiveness verification operation on the dynamic emergency dispatch strategy to obtain a strategy verification result, and generating an emergency dispatch instruction set based on the strategy verification result, and transmitting the emergency dispatch instruction set to a work terminal device to trigger an emergency response operation; the method comprises: setting a work device node, an environmental obstacle node and a personnel position node as a basic node set of the graph structure; constructing a first type of edge connection relationship according to the actual distance between the device node and the obstacle node, and constructing a second type of edge connection relationship according to the operation association between the personnel position node and the device node; constructing a node attribute feature set based on environmental monitoring sensor data, the node attribute feature set including device operating parameters, obstacle material properties and personnel state parameters, the personnel state parameters including personnel protective equipment state parameters and personnel vital sign state parameters; combining the basic node set, the edge connection relationship and the node attribute feature set to generate the work environment graph structure; performing a topological integrity check operation on the work environment graph structure, supplementing missing edge connection relationships and deleting redundant nodes; updating the corresponding parameter values in the node attribute feature set according to the dynamic changes of the device operating parameters, and recalculating the actual distance from the device node to the obstacle node based on the real-time position of the device node, dynamically adjusting the first type of edge connection relationship, and generating a dynamically updated work environment graph structure.

2. The aerial emergency dispatch method of claim 1, wherein, The method comprises: performing a resolution equalization processing operation on the real-time video data set to generate an initial video data frame of a uniform resolution format; calling a moving target detection algorithm to perform a noise filtering operation on the initial video data frame to remove interference pixel regions of non-work targets, to obtain a filtered video data frame; performing a time sequence alignment processing operation on the filtered video data frame to generate a standardized video data stream with continuous timestamps; performing a color gamut conversion operation on the standardized video data stream to map the original color gamut space to a pre-set standardized color gamut space, and performing an illumination compensation processing operation on the converted video data stream to eliminate shadow regions and overexposure regions in the initial video data frame, to obtain a compensated video data stream; performing a frame rate stability detection operation on the compensated video data stream to eliminate video jitter and generate a time-continuous standardized video data stream.

3. The aerial emergency dispatch method of claim 1, wherein, The calling video analysis model performs feature extraction on the standardized video data stream to generate a job environment feature set and a job personnel behavior feature set, including: The standardized video data stream is segmented into multiple equal-length video segments, and a three-dimensional convolution feature extraction operation is performed on each video segment to generate a spatio-temporal correlation feature matrix; An attention weight distribution operation is performed on the spatio-temporal correlation feature matrix to generate a visual focus distribution map with region importance annotations, and an edge tracking operation is performed on the job equipment contour based on the visual focus distribution map to extract equipment operation state features and spatial displacement trajectory features; A skeletal key point detection operation is performed on the job personnel limb action to generate a set of action amplitude features and motion speed features; The equipment operation state features, spatial displacement trajectory features, action amplitude features, and motion speed features are aggregated to generate the job personnel behavior feature set; A texture analysis operation is performed on the fixed obstacles in the job environment to extract surface roughness features and geometric shape features; A dynamic interaction model of material properties and equipment kinematics parameters is constructed by fusing the surface roughness features, geometric shape features, and equipment operation state features through a graph neural network, and an environment interaction risk feature set is generated.

4. The aerial emergency dispatch method of claim 1, wherein, The combination of the base node set, edge connection relationship, and node attribute feature set to generate the job environment graph structure includes: A vectorization encoding operation is performed on the equipment operation parameters in the node attribute feature set to generate an equipment attribute vector; A texture feature extraction operation is performed on the obstacle material property in the node attribute feature set to generate an obstacle attribute vector; A vectorization encoding operation is performed on the personnel state parameters in the node attribute feature set to generate a personnel attribute vector; The equipment attribute vector, obstacle attribute vector, and personnel attribute vector are respectively mapped to a unified feature dimension space through a fully connected layer and then concatenated by channel dimension to generate a node feature matrix; The spatial correlation degree between the equipment node and the obstacle node is calculated according to the actual distance value of the first type of edge connection relationship to generate a first type of edge weight matrix; The interaction intensity value between the personnel position node and the equipment node is calculated according to the operation correlation parameter of the second type of edge connection relationship to generate a second type of edge weight matrix; The node feature matrix, the first type of edge weight matrix, and the second type of edge weight matrix are graph topological mapped to generate initial graph structure data; The inter-node reachability of the initial graph structure data is verified, the edge connection relationship that does not meet the preset connectivity threshold is deleted, the missing key edge connection relationship is supplemented, and a fully connected job environment graph structure is generated; According to the real-time update frequency of the equipment operation parameters, a dynamic incremental update is performed on the fully connected job environment graph structure to generate a dynamic job environment graph structure with a timestamp version identifier.

5. The aerial emergency dispatch method of claim 1, wherein, The calling graph convolution network performs multi-level feature aggregation operation on the job environment graph structure to generate an environment state feature vector, including: In the initial convolution layer of the graph convolution network, a feature aggregation operation is performed on the direct adjacent nodes of each node in the job environment graph structure to generate a primary aggregation feature vector; The indirect adjacent nodes spaced two edge connection relationships are executed cross-level feature propagation operation between the intermediate convolution layers of the graph convolution network, and a middle-level aggregated feature vector is generated; The preliminary aggregated feature vector and the middle-level aggregated feature vector are fused in the final convolution layer of the graph convolution network to generate an environment state feature vector containing global dependency relationship; Dimension reduction visualization processing is performed on the spatial coordinate encoding information embedded in the environment state feature vector to generate an environment risk heat map and label high-risk area coordinates; Spatial clustering analysis is performed on the coordinate data of the environment risk heat map to identify risk aggregation areas and generate regional boundary labeling information; The node attribute feature set of the work environment graph structure is updated based on the regional boundary labeling information to generate an optimized environment state feature vector.

6. The aerial emergency dispatch method of claim 1, wherein, The dynamic emergency scheduling strategy is generated based on the environment state feature vector and the work personnel behavior feature set, which includes: After the environment state feature vector and the work personnel behavior feature set are respectively input into a linear projection layer to align to the same feature dimension, they are input into a pre-trained spatio-temporal correlation model, and feature interaction verification is performed on the node attribute features in the environment state feature vector and the dynamic parameter in the work personnel behavior feature set through cross-attention mechanism to generate an interaction matrix; Based on the correlation weight value of each node and behavior feature in the interaction matrix, a risk correlation degree parameter between high-risk nodes in the work environment graph structure and personnel behavior abnormal features is extracted; An adaptive scaling operation is performed on the node spatial coordinate features in the environment state feature vector based on the risk correlation degree parameter to generate a risk-weighted spatial distribution feature vector; The risk-weighted spatial distribution feature vector and the edge connection relationship of the work environment graph structure are input into a graph attention network to generate a topological optimization feature map containing the emergency obstacle avoidance path between nodes through an edge weight dynamic adjustment algorithm; Based on the path connectivity parameter in the topological optimization feature map, a Dijkstra algorithm is called to calculate the shortest safe movement path sequence of the work terminal device in the work environment graph structure in combination with the edge weight of the graph attention network, and a device motion parameter constraint condition corresponding to the shortest safe movement path sequence is generated; The device motion parameter constraint condition and the real-time device running parameter are executed dynamic compatibility verification, and the path segment exceeding the physical limit of the device is eliminated according to the verification result to generate a dynamic emergency scheduling strategy containing feasible path coordinates and device control instruction set.

7. The aerial emergency dispatch method of claim 1, wherein, The strategy verification result is obtained by performing effectiveness verification operation on the dynamic emergency scheduling strategy, which includes: The feasible path coordinate sequence and the device motion parameter constraint condition in the dynamic emergency scheduling strategy are extracted, and a three-dimensional dynamic scene reconstruction model is constructed according to the spatial coordinate system of the work environment graph structure; The feasible path coordinate sequence and the device motion parameter constraint condition are input into a preloaded physics engine simulator, a path curvature key point segmentation simulation task is performed based on the path curvature key point segmentation simulation task, a virtual motion simulation operation is performed in a segmented parallel computing mode, and simulation motion trajectory data and virtual device state parameter set are generated; Perform a spatial collision detection operation on the simulated motion trajectory data and the obstacle node coordinates in the work environment graph structure, generating a set of trajectory conflict point coordinates and a sequence of conflict timestamps; Perform a dynamic comparison operation on the set of virtual device state parameters and the physical limit thresholds of device operating parameters, generating device over-limit operating parameter identifiers and corresponding time intervals; Fuse the set of trajectory conflict point coordinates, the sequence of conflict timestamps, and the device over-limit operating parameter identifiers to generate a strategy verification exception report, extract abnormal path segment coordinates and abnormal device control instructions; Perform segmented smoothing interpolation processing on the set of trajectory conflict point coordinates according to the abnormal path segment coordinates, generate an alternative path coordinate sequence and recalculate the device motion parameter constraints; Input the alternative path coordinate sequence and the recalculated device motion parameter constraints into the physics engine simulator to perform a secondary verification operation. If the set of trajectory conflict point coordinates is empty and the device over-limit operating parameter identifier is not activated, it is marked as a verified strategy.

8. The aerial emergency dispatch method of claim 7, wherein, Generate a set of emergency dispatch instructions based on the strategy verification result, and transmit the set of emergency dispatch instructions to the work terminal device to trigger an emergency response operation, including: Generate a device control instruction code sequence based on the device motion parameter constraints in the verified strategy, and synchronize and align the device control instruction code sequence with the alternative path coordinate sequence with timestamps to generate a set of emergency dispatch instructions; Perform an encapsulation operation on the set of emergency dispatch instructions, add a work terminal device address identifier and an instruction priority label, generate a target transmission data packet and send it to the work terminal device; Update the set of feasible path coordinates in the dynamic emergency dispatch strategy according to the instruction confirmation signal and real-time position feedback data returned by the work terminal device, and trigger a periodic strategy validity verification operation.

9. An aerial emergency dispatch system, comprising: A device including a processor and a memory, the memory and the processor being connected, the memory being used to store programs, instructions or codes, and the processor being used to execute the programs, instructions or codes in the memory to implement the aerial work emergency dispatch method of any one of claims 1-8.

Images