An industrial unmanned aerial vehicle emergency closed-loop disposal method and system based on cloud edge cooperation, an edge computing node, and a storage medium

Abstract

The application relates to the field of unmanned aerial vehicle control and discloses an industrial unmanned aerial vehicle emergency closed-loop disposal method and system based on cloud edge cooperation, an edge computing node and a storage medium, the method comprising the following steps: receiving a video stream and performing real-time detection, outputting a composite detection tuple, performing cross-frame evidence accumulation and double-threshold decision, obtaining a confirmed event and confidence, and obtaining geodetic coordinates and a positioning uncertainty ellipse through back projection; obtaining an event type hierarchical code according to an abnormal category, generating a space-time-semantic fingerprint by combining the geodetic coordinates and a disposal semantic embedding vector, and performing merging decision; determining a severity level after normalizing the confidence and a diffusion trend index; performing hierarchical solving with backtracking based on a four-layer mapping rule library, generating a mapping track and encapsulating an instruction data packet; and updating a detection model and a rule library according to a returned execution state and a review conclusion, and completing closed-loop disposal. The application can reduce response delay, reduce bandwidth occupation, and realize disposal decision automation and interpretability.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to an emergency closed-loop handling method, system, edge computing node, and computer-readable storage medium for industrial UAVs based on cloud-edge collaboration. Background Technology

[0002] Industrial drones are increasingly being used in scenarios such as park inspections, power line patrols, and emergency rescue. As task complexity increases, the industry is adopting a three-layer collaborative architecture of cloud, edge, and endpoint. Data is collected by the drone, edge computing nodes are deployed near the site for real-time analysis, and the cloud platform handles global management and data archiving. These three components interact with low latency via 5G or a private network. Edge nodes typically use lightweight deep learning models to perform real-time target detection and abnormal behavior recognition on the transmitted video streams, eliminate duplicate reporting through event deduplication mechanisms, and map the recognition results to specific handling instructions based on a rule base.

[0003] However, this model has common problems: in traditional architectures, all video data must be uploaded back to the cloud, resulting in response latency of several seconds to minutes. Existing edge AI models typically only output categories and bounding boxes, lacking semantic information that can be directly used for decision-making, creating a semantic gap between perception and decision-making. Existing deduplication solutions often use a weighted sum of similarity across three dimensions (location, time, and type) and compare it with a threshold, requiring pairwise floating-point similarity calculations, with computational complexity increasing approximately quadratically with the number of concurrent events. Existing rule bases are mostly single-layer key-value mappings, with rules strongly bound to devices; if the device is unavailable or on-site constraints change, automatic alternative solutions cannot be found.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a cloud-edge collaborative emergency closed-loop handling method, system, edge computing node, and computer-readable storage medium for industrial drones. It aims to solve the technical problems in the prior art, such as long response delays of drones, bandwidth and computing power bottlenecks caused by centralized cloud processing, disconnect between edge model computing power allocation and business risks and lack of handling semantics in output, crude methods for suppressing repeated event reporting, poor scalability of single-layer flat mapping of handling rule base, and lack of automated closed-loop handling mechanism.

[0006] To achieve the above objectives, this invention provides a cloud-edge collaborative emergency closed-loop handling method for industrial drones, which includes the following steps: The system receives video streams collected by a drone, identifies the video streams based on a detection model, and outputs a composite detection tuple, which includes anomaly category, pixel bounding box, multidimensional processing semantic embedding vector, target true scale estimation, and diffusion trend index. Based on the composite detection tuple, cross-frame evidence accumulation and dual-threshold hysteresis decision are performed to obtain the confirmed event and the existence confidence of the confirmed event. After performing non-maximum suppression on the pixel bounding box, pixel-geometry back projection is performed in combination with the real-time pose of the UAV and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmed event. An event type hierarchy code is obtained based on the anomaly category. An event spatiotemporal-semantic fingerprint is generated based on the event type hierarchy code, the geodetic coordinates, and the multidimensional processing semantic embedding vector. The confirmed events are merged based on the event spatiotemporal-semantic fingerprint to obtain merged event objects. Based on the event type risk coefficient, regional sensitivity coefficient, existence confidence and diffusion trend index in the merged event object, a severity level is obtained, and a structured event message is created that includes the event type hierarchical code, the severity level, the geodetic coordinates, the multidimensional treatment semantic embedding vector, the target true scale estimate and the diffusion trend index; Based on a four-layer decoupled mapping rule base consisting of a semantic layer, a policy layer, a capability layer, and a parameter layer, the event type level code, the severity level, and the multi-dimensional handling semantic embedding vector in the structured event message are used as search keys to perform a backtracking layered mapping to obtain a mapping trajectory, and the mapping trajectory is encapsulated into an instruction data packet. The instruction data packet is sent to the UAV or ground equipment and the execution status is received back. The next action to be performed is determined based on the execution status. The structured event message and the mapping trajectory are reported to the cloud platform and the verification conclusion is received. The detection model and the four-layer decoupling mapping rule base are updated according to the verification conclusion.

[0007] Optionally, in the cloud-edge collaborative emergency closed-loop handling method for industrial drones, the detection model is a risk-adaptive, resilient, lightweight detection model, which includes: a backbone network, a neck, and a detection head. The backbone network includes multiple downsampling stages. After each downsampling stage, a risk situation gating unit is set up. The risk situation gating unit is used to take the channel statistics and the risk situation score of the current region at the previous moment as input, and output the channel mask through a fully connected and binarized operator. It dynamically determines the subset of channels participating in the convolution operation in the current stage, so that when the risk situation score is lower than the low risk threshold, the same set of model weights retains some channels and reduces the inference frame rate, and when the risk situation score exceeds the high risk threshold, it retains all channels and increases the inference frame rate. The risk situation score is obtained by weighting the highest severity level of confirmed events in the region at the previous moment, the number of active events, and the false alarm immunity table hit rate. The neck includes a feature pyramid and a path aggregation network, used to fuse multi-scale features selected by the channel mask. During the fusion process, the positions of channels closed by the channel mask are occupied with zero values. The detection head includes a classification branch, a bounding box regression branch, and a treatment attribute branch. The classification branch is used to output the anomaly category, the bounding box regression branch is used to output the pixel bounding box, and the treatment attribute branch is used to output the multidimensional treatment semantic embedding vector, the target true scale estimate, and the diffusion trend index.

[0008] Optionally, the aforementioned cloud-edge collaborative emergency closed-loop handling method for industrial drones, wherein the step of performing cross-frame evidence accumulation and dual-threshold hysteresis decision based on the composite detection tuple to obtain the confirmed event and the existence confidence of the confirmed event, performing non-maximum suppression on the pixel bounding box, and then performing pixel-to-geometry back projection combined with the drone's real-time pose and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmed event, specifically includes: For each composite detection tuple in the current frame of the video stream, the detection confidence of the current frame and the accumulated event presence confidence of the previous frame are weighted and recursively calculated using a forgetting factor, and the detection confidence is modulated using a frame quality factor to obtain the event presence confidence of the current frame; wherein, the forgetting factor is a preset constant, and the frame quality factor is obtained by normalizing motion blur, occlusion ratio, and target pixel area; When the confidence level of the event meets the activation threshold, the detection object corresponding to the composite detection tuple is determined as the confirmed event, and the confirmed event and the event existence confidence level of the confirmed event are output. When the confidence level of the event decreases to the closing threshold and continues for a preset frame, the confirmed event is determined to be resolved, wherein the closing threshold is less than the opening threshold; Based on the confidence level of the event, the overlapping pixel bounding boxes are sorted, the pixel bounding box with the highest confidence level of the event is retained and the other overlapping boxes are suppressed to obtain the target pixel bounding box after non-maximum suppression. An observation ray is constructed based on the real-time pose of the UAV, the gimbal pointing angle, and the camera intrinsic parameters. The midpoint of the bottom edge of the target pixel bounding box is projected onto the ground along the observation ray to obtain the projection point. The latitude and longitude of the midpoint of the bottom edge are intersected with the digital elevation model to obtain the elevation of the midpoint of the bottom edge. The latitude and longitude and the elevation are used as the geodetic coordinates of the confirmed event. Error propagation calculations are performed based on the pose angle error, positioning error, and ranging error of the UAV in real time to obtain the positioning uncertainty ellipse of the confirmed event.

[0009] Optionally, the aforementioned cloud-edge collaborative emergency closed-loop handling method for industrial drones, wherein obtaining an event type hierarchy code based on the anomaly category, generating an event spatiotemporal-semantic fingerprint based on the event type hierarchy code, the geodetic coordinates, and the multidimensional handling semantic embedding vector, and performing a merging decision on the confirmed events based on the event spatiotemporal-semantic fingerprint to obtain merged event objects, specifically includes: The event type hierarchy code is obtained based on the anomaly category mapping. The influence radius is determined based on the event type hierarchy code. The influence radius is used as the grid side length. The geodetic coordinates are partitioned into a hierarchical quadtree. The partitioning result is encoded into a fixed-length bit string. The fixed-length bit string is used as the spatial segment of the event spatiotemporal-semantic fingerprint. The neighborhood grid key of the grid is generated. The event type hierarchy code is used as the type segment of the event spatiotemporal-semantic fingerprint, and the multidimensional processing semantic embedding vector is binarized using local sensitive hashing to obtain the semantic segment of the event spatiotemporal-semantic fingerprint. By concatenating the spatial segment, the type segment, and the semantic segment, the event spatiotemporal-semantic fingerprint is obtained. Prefix filtering is performed according to the type segment, and hash retrieval is performed in the event index table using the spatial segment and the neighborhood raster key to obtain a set of candidate active events; For each existing event in the candidate active event set, determine whether the Hamming distance of the semantic segment is not greater than a preset threshold, whether the observation time of the confirmed event falls within the time active window of the existing event, and whether the distance between the geodetic coordinates of the confirmed event and the last position of the existing event is not greater than the product of the maximum moving speed and the time difference corresponding to the event type hierarchical code. If all three conditions are met, the confirmed event is determined to be mergeable evidence. The positions of the existing events are weighted and fused using the reciprocal of the positioning uncertainty ellipse as the weight, and the existing events are used as the merged event objects. Otherwise, a new event object is created based on the confirmed event, and the newly created event object is used as the merged event object. When performing prefix filtering according to the type segment, if the major categories of the type segments are different, they are directly determined to be different events and the search ends.

[0010] Optionally, the cloud-edge collaborative emergency closed-loop handling method for industrial drones, wherein obtaining the severity level based on the event type risk coefficient, regional sensitivity coefficient, existence confidence level, and diffusion trend index in the merged event objects specifically includes: Obtain the event type risk coefficient, regional sensitivity coefficient, event existence confidence level, and diffusion trend index from the merged event objects; The confidence level of the event is normalized to a preset interval to obtain the cumulative confidence coefficient, and the diffusion trend index is normalized to a preset interval to obtain the diffusion trend coefficient. The comprehensive urgency score is obtained by multiplying the event type risk coefficient, the regional sensitivity coefficient, the cumulative confidence coefficient, and the diffusion trend coefficient by their respective preset weights and then summing them. The severity level is determined based on the threshold range into which the comprehensive urgency score falls. When the comprehensive urgency score is not less than the first threshold, it is determined to be an emergency level. When the comprehensive urgency score is between the second threshold and the first threshold, it is determined to be a normal level. When the comprehensive urgency score is less than the second threshold, it is determined to be a warning level.

[0011] Optionally, the aforementioned cloud-edge collaborative emergency closed-loop handling method for industrial drones, wherein the method utilizes a four-layer decoupled mapping rule base consisting of a semantic layer, a policy layer, a capability layer, and a parameter layer, using the event type hierarchy code, the severity level, and the multi-dimensional handling semantic embedding vector in the structured event message as retrieval keys, to perform a backtracking-based hierarchical mapping to obtain a mapping trajectory, and encapsulates the mapping trajectory into an instruction data packet, specifically including: Through the semantic layer, the event type hierarchy code, the severity level, and the multidimensional handling semantic embedding vector are combined into a triplet retrieval key, and matching is performed in the explicit rule table of the semantic layer based on the triplet retrieval key; If a match is found, the set of abstract handling intentions that are found in the explicit rule table is output. If no match is found, the nearest neighbor search is performed in the intention prototype vector library based on the multidimensional handling semantic embedding vector, the set of abstract handling intentions is output, and the situation of initiating the nearest neighbor search after the explicit rule is not found, along with the event type level code, the severity level, and the multidimensional handling semantic embedding vector, is reported to the cloud. Through the strategy layer, each disposal intention in the set of abstract disposal intentions is expanded into a directed acyclic graph of action templates. The nodes of the directed acyclic graph of action templates are abstract disposal actions, and the edges of the nodes are execution relationships. Through the capability layer, for each node in the directed acyclic graph of the action template, subset matching is performed in the device capability profile according to the set of capability labels carried by the node to obtain a set of feasible executors. For each candidate executor in the set of feasible executors, the cost function value is calculated, and the candidate executor with the smallest cost function value is selected as the executor. The executor is bound to the corresponding node to obtain the directed acyclic graph of the target action template after binding the executor. The cost function is obtained by weighting the estimated arrival time, the remaining power margin after the task is completed, the task preemption conflict cost, and the equipment loss balancing cost. Through the parameter layer, the abstract action of each node in the directed acyclic graph of the target action template is instantiated into specific instruction parameters. The specific instruction parameters are input into the safety envelope constraint solver for verification. If the verification passes, an action instance with the specific instruction parameters is obtained. If the verification fails, the specific instruction parameters are converged to the safety boundary. When the capability layer cannot match a feasible executor or the parameter layer determines that there is no feasible instance of the action node, backtrack to the strategy layer, select the suboptimal branch and re-execute all steps after the strategy layer is expanded, until a complete executable mapping link is obtained or the number of backtracking times reaches the upper limit. During the layer-by-layer execution process, the execution results of each layer are recorded, a mapping trajectory is generated, and after the execution is completed, the action instance with the specific instruction parameters is encapsulated into an instruction data packet according to the topological order of the directed acyclic graph of the action template.

[0012] Optionally, the cloud-edge collaborative emergency closed-loop handling method for industrial drones, wherein updating the detection model and the four-layer decoupling mapping rule base according to the review conclusion specifically includes: Receive the review conclusion returned by the cloud platform, the review conclusion including effective handling, invalid handling, false alarm and missed detection; The event-related video clips, detection results, and mapping trajectory in the structured event message corresponding to the review conclusion are uploaded to the cloud platform to perform an update operation; Receive the updated detection model and adjusted rule base issued by the cloud platform, and redeploy the updated detection model and adjusted rule base; The spatiotemporal-semantic fingerprint of events that are judged as false alarms is written into the false alarm immunity table. When subsequent observations hit the false alarm immunity table and the counter exceeds the suppression threshold, the event existence confidence of subsequent observations is downweighted. The update operation performed by the cloud platform is as follows: using the cloud-based high-parameter model as the teacher model and the risk-adaptive elastic lightweight detection model as the student model, the data containing the review conclusion, the event-related video clips, the detection results, and the mapping trajectory are used as an incremental sample set with treatment labels. This data is then input into the teacher-student distillation framework to construct the total distillation loss. The detection model is updated with the goal of minimizing the total distillation loss. The updated detection model is then distributed to edge computing nodes. Additionally, based on the rule identifier dimension, the treatment success rate, average treatment time, backtracking trigger rate, and manual intervention rate are statistically analyzed. The rule base is adjusted based on the statistical results, and the adjusted rule base is distributed to edge computing nodes.

[0013] Furthermore, to achieve the above objectives, the present invention also provides an emergency closed-loop response system for industrial drones based on cloud-edge collaboration, wherein the emergency closed-loop response system for industrial drones based on cloud-edge collaboration includes: The edge real-time detection module is used to receive video streams collected by the UAV, identify the video streams based on the detection model, and output a composite detection tuple. The composite detection tuple includes anomaly category, pixel bounding box, multi-dimensional processing semantic embedding vector, target true scale estimation, and diffusion trend index. The confirmation event acquisition module is used to perform cross-frame evidence accumulation and double threshold hysteresis decision based on the composite detection tuple to obtain the confirmation event and the existence confidence of the confirmation event. After performing non-maximum suppression on the pixel bounding box, pixel-geometry back projection is performed in combination with the real-time pose of the UAV and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmation event. The merging and fingerprint generation module is used to obtain an event type hierarchy code based on the anomaly category, generate an event spatiotemporal-semantic fingerprint based on the event type hierarchy code, the geodetic coordinates, and the multidimensional processing semantic embedding vector, and perform a merging decision on the confirmed events based on the event spatiotemporal-semantic fingerprint to obtain merged event objects. The event message structuring module is used to obtain the severity level based on the event type risk coefficient, regional sensitivity coefficient, existence confidence and diffusion trend index in the merged event object, and create a structured event message including the event type hierarchical code, the severity level, the geodetic coordinates, the multidimensional treatment semantic embedding vector, the target true scale estimate and the diffusion trend index; The decoupling mapping and encapsulation module is used to perform a backtracking layered mapping to obtain a mapping trajectory based on a four-layer decoupling mapping rule base consisting of a semantic layer, a policy layer, a capability layer, and a parameter layer, using the event type level code, the severity level, and the multi-dimensional handling semantic embedding vector in the structured event message as the retrieval key, and encapsulate the mapping trajectory into an instruction data packet. The instruction issuance and execution module is used to issue the instruction data packet to the UAV or ground equipment and receive the returned execution status, and determine the next action to be executed based on the execution status; The closed-loop and update module is used to report the structured event message and the mapping trajectory to the cloud platform and receive the returned review conclusion, and update the detection model and the four-layer decoupled mapping rule base according to the review conclusion.

[0014] Furthermore, to achieve the above objectives, the present invention also provides an edge computing node, wherein the edge computing node includes: a memory, a processor, and a cloud-edge collaborative industrial drone emergency closed-loop handling program stored on the memory and executable on the processor. When the cloud-edge collaborative industrial drone emergency closed-loop handling program is executed by the processor, it implements the steps of the cloud-edge collaborative industrial drone emergency closed-loop handling method described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an emergency closed-loop handling program for industrial drones based on cloud-edge collaboration, and when the emergency closed-loop handling program for industrial drones based on cloud-edge collaboration is executed by a processor, it implements the steps of the emergency closed-loop handling method for industrial drones based on cloud-edge collaboration as described above.

[0016] In this invention, a video stream is received and real-time detection is performed, outputting a composite detection tuple. Based on this tuple, cross-frame evidence accumulation and dual-threshold decision-making are performed to obtain confirmed events and their confidence levels. Geodetic coordinates and a positioning uncertainty ellipse are obtained through back projection. An event type hierarchy code is obtained based on the anomaly category. A spatiotemporal-semantic fingerprint is generated by combining the geodetic coordinates and the semantic embedding vector for handling, and a merge decision is performed. After normalizing the confidence level and diffusion trend index, a weighted score is performed using risk and sensitivity coefficients to determine the severity level, creating a structured event message. A layered solution with backtracking is executed based on a four-layer mapping rule base to generate a mapping trajectory and encapsulate a command data packet. The command data packet is issued, and the detection model and rule base are updated based on the returned execution status and review conclusions to complete the closed-loop handling. This invention can reduce the response latency of emergency closed-loop handling of industrial drones, reduce bandwidth consumption, improve decision-making accuracy, and achieve automated and interpretable handling decisions. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the cloud-edge collaborative emergency closed-loop response method for industrial drones according to the present invention; Figure 2 This is a structural diagram of a preferred embodiment of the cloud-edge collaborative emergency closed-loop response system for industrial drones of the present invention; Figure 3 This is a structural diagram of a preferred embodiment of the edge computing node of the present invention. Detailed Implementation

[0018] This application provides a cloud-edge collaborative emergency closed-loop response method, system, and edge computing node for industrial drones. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0019] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0020] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0021] The preferred embodiment of the present invention describes an emergency closed-loop response method for industrial drones based on cloud-edge collaboration, such as... Figure 1 As shown, the cloud-edge collaborative emergency closed-loop response method for industrial drones includes the following steps: Step S10: Receive the video stream collected by the UAV terminal, identify the video stream based on the detection model, and output a composite detection tuple. The composite detection tuple includes anomaly category, pixel bounding box, multidimensional processing semantic embedding vector, target true scale estimation, and diffusion trend index.

[0022] The risk-adaptive, elastic, lightweight detection model is a core detection model modified from the general lightweight target detection framework. The backbone network of this model embeds risk situation gating units at each stage, which can dynamically select a subset of convolution channels based on the risk situation score of the region at the previous time. Under the same set of model weights, it can achieve elastic inference for low-risk, low-computing-power cruising and high-risk, full-precision detection. In addition to the classification and regression branches, the detection head adds a disposal attribute branch in parallel, which directly outputs physical and semantic quantities that are strongly related to disposal decisions.

[0023] In this embodiment, a risk-adaptive, resilient, lightweight detection model is deployed on edge computing nodes to perform real-time target detection and abnormal behavior recognition on video streams transmitted from drones. All AI analysis is completed in the near field, eliminating the need to transmit massive video streams back to the cloud, significantly reducing transmission latency and bandwidth consumption, and achieving millisecond-level triggering of emergency response. The model consists of three parts: a risk-situation-gated resilient backbone, a multi-scale feature fusion neck, and a three-branch heterogeneous detection head. The backbone network embeds risk-situation-gated units after each downsampling stage. Taking channel statistics and the risk-situation score of the current region at the previous moment as input, it outputs a 0 / 1 channel mask through a fully connected layer and a binarization operator, dynamically determining the subset of channels participating in convolution at this stage. When the risk-situation score is below a threshold, approximately 30% of channels are retained and the inference frame rate is reduced; when it is above the threshold, all channels are retained and the inference frame rate is increased. The neck uses a feature pyramid and path aggregation network to fuse multi-scale features, with the positions of gated channels occupied by zero values. The detection head adds a treatment attribute branch in parallel with the classification and regression branches. The classification branch outputs the anomaly category, the regression branch outputs the pixel bounding box and the detection confidence of the current frame, and the treatment attribute branch outputs a multi-dimensional treatment semantic embedding vector, the target's true scale estimate, and a diffusion trend index. The outputs of the three branches together constitute a composite detection tuple, providing standardized input for subsequent cross-frame evidence accumulation and event confirmation.

[0024] Specifically, the detection model is a risk-adaptive, resilient, lightweight detection model, which includes: a backbone network, a neck, and a detection head; The backbone network includes multiple downsampling stages. After each downsampling stage, a risk situation gating unit is set up. The risk situation gating unit is used to take the channel statistics and the risk situation score of the current region at the previous moment as input, and output the channel mask through a fully connected and binarized operator. It dynamically determines the subset of channels participating in the convolution operation in the current stage, so that when the risk situation score is lower than the low risk threshold, the same set of model weights retains some channels and reduces the inference frame rate, and when the risk situation score exceeds the high risk threshold, it retains all channels and increases the inference frame rate. The risk situation score is obtained by weighting the highest severity level of confirmed events in the region at the previous moment, the number of active events, and the false alarm immunity table hit rate. The neck includes a feature pyramid and a path aggregation network, used to fuse multi-scale features selected by the channel mask. During the fusion process, the positions of channels closed by the channel mask are occupied with zero values. The detection head includes a classification branch, a bounding box regression branch, and a treatment attribute branch. The classification branch is used to output the anomaly category, the bounding box regression branch is used to output the pixel bounding box, and the treatment attribute branch is used to output the multidimensional treatment semantic embedding vector, the target true scale estimate, and the diffusion trend index.

[0025] Understandably, edge computing nodes receive video streams captured by drones and utilize a locally deployed risk-adaptive, resilient, lightweight detection model to perform real-time target detection and abnormal behavior identification. This model consists of three parts: a backbone network, a neck network, and a detection head.

[0026] Specifically, the backbone network includes multiple downsampling stages, each followed by a risk situation gating unit. The risk situation gating unit takes channel statistics and the risk situation score of the region at the previous time step as input, outputs a 0 / 1 channel mask through a fully connected and binarization operator, and dynamically determines the subset of channels participating in the convolution operation at this stage. The risk situation score is obtained by weighting the highest severity level of confirmed events, the number of active events, and the false positive immunity table hit rate of the region at the previous time step. When the risk situation score is below the low-risk threshold, the channel mask retains approximately 30% of the channels and reduces the inference frame rate; when the risk situation score is in the medium-risk range, the channel mask retains approximately 60% of the channels; when the risk situation score exceeds the high-risk threshold, the channel mask retains all channels and increases the inference frame rate. Thus, the same set of model weights can dynamically balance computing power, power consumption, and accuracy without redeployment, achieving flexible inference with low-computing-power cruising during low-risk periods and full-accuracy detection during high-risk periods.

[0027] The false alarm immunity table is a local table deployed on edge nodes. It records fingerprints that have been verified as false alarms by the cloud, using event fingerprint truncation codes as keys and time-decayed counters as values. Subsequent observations that hit this table and whose counters exceed the suppression threshold are downweighted or silenced, achieving self-learning suppression of persistent false alarms.

[0028] Furthermore, the neck region includes a feature pyramid network and a path aggregation network for fusing multi-scale features selected by channel masks. During the fusion process, the positions of channels closed by channel masks are occupied with zero values, maintaining consistent feature map dimensions under different sparsity rates. This eliminates the need to switch operators or recompile the inference engine, thereby enhancing the model's ability to perceive anomalous targets at different scales.

[0029] Furthermore, the detection head includes a classification branch, a bounding box regression branch, and a disposal attribute branch. The classification branch outputs the anomaly category. The bounding box regression branch outputs the pixel bounding box and the detection confidence of the current frame. The disposal attribute branch outputs a K-dimensional disposal semantic embedding vector, a target true scale estimate, and a diffusion trend index, where K ranges from 32 to 128. During training, the K-dimensional disposal semantic embedding vector output by the disposal attribute branch is supervised by disposal intent labels in the rule base, ensuring that targets with similar disposal methods are grouped closer together in the vector space. This vector simultaneously carries two semantic meanings: "what the target is" and "how it should be disposed of." It subsequently serves as a semantic segment of the event spatiotemporal-semantic fingerprint for event merging and as a joint retrieval key for the four-layer mapping rule base semantic layer in solving the disposal intent. The outputs of the classification branch, bounding box regression branch, and disposal attribute branch together constitute a composite detection tuple, which includes the anomaly category, pixel bounding box, current frame detection confidence, K-dimensional disposal semantic embedding vector, target true scale estimate, and diffusion trend index.

[0030] This invention achieves dynamic matching of computing power and business risks through the aforementioned risk-based gating elastic backbone. By embedding disposal semantics into the model output through a three-branch heterogeneous detection head, it eliminates semantic gaps between the perception and decision-making layers. Model training employs a disposal cost-sensitive closed-loop distillation training method. Sample-level cost weights are constructed using cloud-based disposal verification conclusions and injected into the distillation loss, shifting the training objective from minimizing classification and localization errors to minimizing disposal costs. Relying on the disposal verification closed loop of edge computing nodes, it continuously optimizes for high-cost missed detections and persistent false alarms, forming a self-evolving closed loop of perception-disposal-verification-retraining.

[0031] Step S20: Based on the composite detection tuple, perform cross-frame evidence accumulation and dual-threshold hysteresis decision to obtain the confirmed event and the existence confidence of the confirmed event. After performing non-maximum suppression on the pixel bounding box, perform pixel-geometry back projection in combination with the UAV real-time pose and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmed event.

[0032] Specifically, for each composite detection tuple in the current frame of the video stream, the detection confidence of the current frame and the accumulated event presence confidence of the previous frame are weighted and recursively calculated using a forgetting factor, and the detection confidence is modulated using a frame quality factor to obtain the event presence confidence of the current frame; wherein, the forgetting factor is a preset constant, and the frame quality factor is obtained by normalizing motion blur, occlusion ratio, and target pixel area; When the confidence level of the event meets the activation threshold, the detection object corresponding to the composite detection tuple is determined as the confirmed event, and the confirmed event and the event existence confidence level of the confirmed event are output. When the confidence level of the event decreases to the closing threshold and continues for a preset frame, the confirmed event is determined to be resolved, wherein the closing threshold is less than the opening threshold; Based on the confidence level of the event, the overlapping pixel bounding boxes are sorted, the pixel bounding box with the highest confidence level of the event is retained and the other overlapping boxes are suppressed to obtain the target pixel bounding box after non-maximum suppression. An observation ray is constructed based on the real-time pose of the UAV, the gimbal pointing angle, and the camera intrinsic parameters. The midpoint of the bottom edge of the target pixel bounding box is projected onto the ground along the observation ray to obtain the projection point. The latitude and longitude of the midpoint of the bottom edge are intersected with the digital elevation model to obtain the elevation of the midpoint of the bottom edge. The latitude and longitude and the elevation are used as the geodetic coordinates of the confirmed event. Error propagation calculations are performed based on the pose angle error, positioning error, and ranging error of the UAV in real time to obtain the positioning uncertainty ellipse of the confirmed event.

[0033] Understandably, this step maintains an evidence chain for each candidate target (i.e., the detected object corresponding to the composite detection tuple output after target detection in each frame of the video stream using the risk-adaptive elastic lightweight detection model). The detection result of a single frame is recursively accumulated as a single piece of evidence, rather than being directly determined as an event. Specifically, let the detection confidence of a candidate target in frame t be s_t, and its event existence confidence P_t be recursively accumulated according to P_t = λ·P_(t-1) + (1-λ)·g(s_t,q_t), where λ is the forgetting factor ranging from 0.6 to 0.9, q_t is the frame quality factor obtained by normalizing motion blur, occlusion ratio, and target pixel area, g(·) is a monotonic mapping function, and P_(t-1) is the event existence confidence of the previous frame. Thus, the contribution of frames with lower quality to the event existence confidence is automatically reduced, avoiding interference from false detections in a single frame on the accumulated results.

[0034] After obtaining the event existence confidence level of the current frame, a dual-threshold hysteresis strategy is used for event confirmation and resolution decisions. When the event existence confidence level reaches and exceeds the opening threshold, the candidate target is confirmed as an event and output downstream. After confirmation, if the event existence confidence level drops below the closing threshold and remains below it for a preset frame, the event is determined to be resolved, where the opening threshold is greater than the closing threshold. A hysteresis interval is formed between the opening and closing thresholds to avoid repeated event generation and cancellation caused by confidence level fluctuations around a single threshold, thereby preventing oscillating issuance of handling instructions.

[0035] Furthermore, for behavioral events such as personnel intrusion and unauthorized operations, the system extracts the motion trajectory and posture sequence of the target in continuous video frames while accumulating evidence. Temporal features are extracted through a lightweight temporal network and fused with spatial features to participate in behavior classification. The classification results are also incorporated into the recursive accumulation process of the confidence level of the above-mentioned events in the form of evidence, thereby improving the recognition accuracy of behavioral events.

[0036] To meet the real-time requirements of edge nodes, this step adopts a pipelined parallel processing mechanism: video stream decoding and AI inference are executed asynchronously. When inference is performed in frame t, decoding and preprocessing are completed synchronously in frame t+1. Evidence accumulation and hysteresis decision are completed in a separate lightweight thread without occupying the main inference thread, ensuring millisecond-level triggering of the emergency response link.

[0037] Furthermore, for events confirmed by the dual-threshold hysteresis decision, before output, the overlapping pixel bounding boxes of the same type are sorted according to the confidence of the event existence. The pixel bounding box with the highest confidence of the event existence is retained and the remaining overlapping boxes are suppressed to obtain the target pixel bounding box after non-maximum suppression.

[0038] The midpoint of the bottom edge of the target pixel bounding box after nonmaximum suppression is used to construct an observation ray by combining the real-time pose of the UAV (latitude, longitude, altitude, and attitude angle), the gimbal pointing angle, and camera intrinsic parameters. This ray is then projected onto the ground to obtain the latitude and longitude of the midpoint. The intersection of the latitude and longitude of the midpoint with the digital elevation model yields the elevation of the midpoint, or it can be calculated using airborne laser ranging values. The latitude, longitude, and elevation are used together as the geodetic coordinates for confirming the event. Simultaneously, error propagation calculations are performed based on the pose angle error, positioning error, and ranging error of the UAV's real-time pose to obtain the positioning uncertainty ellipse for the confirmed event. This ensures that subsequent event merging, fingerprint generation, and decision-making are all conducted within a unified geodetic coordinate system.

[0039] As can be seen, this application effectively suppresses mishandling caused by single-frame false detections and eliminates instruction oscillations caused by detection jitter by accumulating evidence across frames and using dual-threshold hysteresis decision-making. Simultaneously, a pipelined parallel processing mechanism ensures that the main inference thread is not occupied by accumulated decisions, meeting the millisecond-level real-time response requirement. Through pixel-geometric backprojection, edge computing nodes map the midpoint of the bottom edge of the pixel bounding box to the latitude, longitude, and elevation of the target, and output a positioning uncertainty ellipse, ensuring that subsequent event merging, fingerprint generation, and handling decisions are all performed in a unified geodetic coordinate system. In multi-machine co-source observation, geometric consistency is further determined by three-dimensional intersection of observation rays, allowing the event positioning accuracy to continuously converge with the accumulation of observations. Compared with existing technologies, this mechanism replaces statistical similarity with geometric consistency evidence, improving the accuracy of target positioning and event merging in multi-machine collaborative scenarios.

[0040] Step S30: Obtain the event type hierarchy code according to the anomaly category; generate the event spatiotemporal-semantic fingerprint according to the event type hierarchy code, the geodetic coordinates, and the multidimensional processing semantic embedding vector; and perform a merging decision on the confirmed events according to the event spatiotemporal-semantic fingerprint to obtain merged event objects.

[0041] Specifically, an event type hierarchical code is obtained based on the anomaly category mapping, the influence radius is determined based on the event type hierarchical code, the influence radius is used as the grid side length, the geodetic coordinates are partitioned into a hierarchical quadtree, the partitioning result is encoded into a fixed-length bit string, the fixed-length bit string is used as the spatial segment of the event spatiotemporal-semantic fingerprint, and a neighborhood grid key of the grid is generated. The event type hierarchy code is used as the type segment of the event spatiotemporal-semantic fingerprint, and the multidimensional processing semantic embedding vector is binarized using local sensitive hashing to obtain the semantic segment of the event spatiotemporal-semantic fingerprint. By concatenating the spatial segment, the type segment, and the semantic segment, the event spatiotemporal-semantic fingerprint is obtained. Prefix filtering is performed according to the type segment, and hash retrieval is performed in the event index table using the spatial segment and the neighborhood raster key to obtain a set of candidate active events; For each existing event in the candidate active event set, determine whether the Hamming distance of the semantic segment is not greater than a preset threshold, whether the observation time of the confirmed event falls within the time active window of the existing event, and whether the distance between the geodetic coordinates of the confirmed event and the last position of the existing event is not greater than the product of the maximum moving speed and the time difference corresponding to the event type hierarchical code. If all three conditions are met, the confirmed event is determined to be mergeable evidence. The positions of the existing events are weighted and fused using the reciprocal of the positioning uncertainty ellipse as the weight, and the existing events are used as the merged event objects. Otherwise, a new event object is created based on the confirmed event, and the newly created event object is used as the merged event object. When performing prefix filtering according to the type segment, if the major categories of the type segments are different, they are directly determined to be different events and the search ends.

[0042] Understandably, edge computing nodes map event type hierarchical codes based on the anomaly categories of confirmed events. These codes employ a three-level encoding structure: "major category - subcategory - minor category." The edge computing nodes determine the influence radius based on the event type hierarchical code. Different event types correspond to different influence radii: 80 meters for open flames, 150 meters for smoke, 20 meters for personnel intrusion, and 5 meters for abnormal equipment overheating. Using this influence radius as the grid side length, the geographic coordinates of the confirmed event are partitioned into a hierarchical quadtree. The partitioning result is encoded into a fixed-length bit string as the spatial segment of the event's spatiotemporal-semantic fingerprint. Simultaneously, candidate keys are generated for the current grid's eight neighboring grids.

[0043] Edge computing nodes use the event type hierarchy code as the type segment of the event spatiotemporal-semantic fingerprint; they binarize the multidimensional processing semantic embedding vector using local sensitive hashing to obtain a 64-bit or 128-bit bit string as the semantic segment of the event spatiotemporal-semantic fingerprint; and they concatenate the spatial segment, type segment, and semantic segment to obtain the event spatiotemporal-semantic fingerprint.

[0044] Furthermore, the edge computing node searches the event index table using the event's spatiotemporal-semantic fingerprint. First, it performs prefix filtering on the type segments; if the major categories of the type segments are different, they are directly determined to be different events, and the search ends. Then, it performs a hash search on the event index table using the spatial segment and its 8-neighbor raster key to obtain a candidate active event set. For each existing event in the candidate active event set, it checks whether the Hamming distance of the semantic segment is not greater than a preset threshold, whether the observation time of the confirmed event falls within the time active window of the existing event, and whether the distance between the geodetic coordinates of the confirmed event and the last position of the existing event does not exceed the product of the maximum movement speed and time difference corresponding to the event type hierarchy code. If all three conditions are met, the edge computing node determines that the confirmed event is reducible evidence; otherwise, it is determined to be non-reducible.

[0045] If the evidence is determined to be mergingable, the edge computing node adds it to the existing event object's evidence chain. It then performs a weighted fusion update of the existing event's position using the reciprocal of the location uncertainty ellipse as weight. This update includes the last observation time, duration, and time series of the actual scale and diffusion trend of the existing event. The overall urgency score of the event is recalculated, and a level transition is triggered when the overall urgency score crosses a level threshold, requesting a rematch of the handling instruction. If the evidence is determined to be non-mergingable, the edge computing node creates a new event object based on the currently confirmed event, assigns a globally unique event ID, and sets it to the nascent state. Once the accumulated evidence reaches the confirmation condition, it migrates to the confirmed state.

[0046] When the type segment and semantic segment of two observations are compared but the back-projection position deviation exceeds their respective positioning uncertainty ellipse, the edge computing node constructs two observation rays based on the real-time pose and gimbal pointing of the two UAVs, and solves the skew distance and intersection point of the closest point pair of the two observation rays in three-dimensional space; when the skew distance is less than a preset threshold and the intersection point is located within the effective observation cone of the two UAVs, the edge computing node determines that the two observations originate from the same physical target and merges them, using the intersection point as the updated three-dimensional position of the physical target; when the skew distance is not less than a preset threshold or the intersection point is not within the effective observation cone of the two UAVs, the edge computing node determines them as two independent targets and establishes corresponding event objects for each.

[0047] Furthermore, edge computing nodes maintain a false alarm immunity table locally, using event fingerprint truncation codes as keys and time-decaying counters as values. When an edge computing node receives a false alarm judgment result from the cloud platform, it writes the corresponding event fingerprint into the false alarm immunity table and increments the counter. Subsequent observations, after completing a three-level cascaded comparison, if the fingerprint matches the false alarm immunity table and the counter exceeds the suppression threshold, the edge computing node proportionally reduces the event's confidence level or silences it, with the counter decaying exponentially over time. The edge computing node outputs the event's spatiotemporal-semantic fingerprint, along with the event lifecycle state transitions driven during the merging decision process, as the merged event object to the event hierarchy unit.

[0048] As can be seen, this application, through fixed-length composite encoding of event spatiotemporal-semantic fingerprints and three-level concatenated integer-bit decision-making, replaces the weighted summation of pairwise floating-point similarity calculations in existing technologies with integer-bit operations and hash lookups, reducing the matching complexity from approximately quadratic correlation with the number of concurrent events to constant level. By introducing semantic segments, edge computing nodes can distinguish between different events occurring sequentially at the same location and associate the same target moving across regions. Through the five-state transition of the event lifecycle state machine, repeated observations are no longer simply discarded but transformed into effective information to improve the accuracy of event location and level determination. Geometric consistency verification through the three-dimensional intersection of multi-machine observation rays replaces statistical similarity with geometric determination. A false alarm immunity table enables self-learning suppression of persistent false alarms. These mechanisms allow edge computing nodes to eliminate system redundancy and decision-making confusion caused by multiple-source repeated reporting from the source, providing accurate and non-redundant event object input for subsequent event classification and handling instruction matching.

[0049] Step S40: Based on the event type risk coefficient, regional sensitivity coefficient, existence confidence and diffusion trend index in the merged event object, the severity level is obtained, and a structured event message is created that includes the event type hierarchical code, the severity level, the geodetic coordinates, the multidimensional treatment semantic embedding vector, the target true scale estimate and the diffusion trend index.

[0050] Specifically, the event type risk coefficient, regional sensitivity coefficient, event existence confidence, and diffusion trend index are obtained from the merged event objects; The confidence level of the event is normalized to a preset interval to obtain the cumulative confidence coefficient, and the diffusion trend index is normalized to a preset interval to obtain the diffusion trend coefficient. The comprehensive urgency score is obtained by multiplying the event type risk coefficient, the regional sensitivity coefficient, the cumulative confidence coefficient, and the diffusion trend coefficient by their respective preset weights and then summing them. The severity level is determined based on the threshold range into which the comprehensive urgency score falls. When the comprehensive urgency score is not less than the first threshold, it is determined to be an emergency level. When the comprehensive urgency score is between the second threshold and the first threshold, it is determined to be a normal level. When the comprehensive urgency score is less than the second threshold, it is determined to be a warning level.

[0051] In this embodiment, the edge computing node normalizes the event existence confidence level to obtain a cumulative confidence coefficient; it also normalizes the diffusion trend index to obtain a diffusion trend coefficient. The edge computing node obtains the event type risk coefficient and regional sensitivity coefficient from the merged event objects. The event type risk coefficient is determined by a preset risk value corresponding to the event type hierarchy code, and the regional sensitivity coefficient is determined by the security sensitivity level of the region where the event occurred. The edge computing node multiplies the event type risk coefficient, regional sensitivity coefficient, cumulative confidence coefficient, and diffusion trend coefficient by their respective preset weights and then sums them to obtain a comprehensive urgency score. The sum of all preset weights is 1.

[0052] Edge computing nodes determine the severity level based on the threshold range that the comprehensive urgency score falls into: when the comprehensive urgency score is not less than the first threshold, it is determined to be an emergency level; when the comprehensive urgency score is between the second threshold and the first threshold, it is determined to be a normal level; when the comprehensive urgency score is less than the second threshold, it is determined to be a warning level. The second threshold is less than the first threshold. The first and second thresholds are dynamically configured by the system administrator according to the business scenario.

[0053] Furthermore, a structured event message is created in JSON format, containing the following fields: Event ID, generated from the edge node's unique identifier and timestamp hash; Event type hierarchy code, using a three-level encoding of "major category-subcategory-detail category"; Location, including latitude, longitude, and elevation coordinates along with the radius of location uncertainty; Occurrence time and last update time, using a Unix timestamp accurate to milliseconds; Severity level, enumerating values ​​as emergency, normal, or warning levels; Multidimensional handling semantic embedding vector, serving as the joint retrieval key for the semantic layer of the four-layer mapping rule base; Target true scale estimation and diffusion trend index; Cumulative evidence count and event lifecycle status; Event spatiotemporal-semantic fingerprint; Associated video clips, the video file path or URL (Uniform Resource Locator) stored in the edge node's local cache, with a length of 5 seconds before and after the event. After generating the structured event message, the edge computing node writes the message to its local event database and simultaneously pushes it to the cloud platform and instruction matching unit.

[0054] As the event object continuously absorbs new evidence throughout its lifecycle, the event's confidence level and diffusion trend index are updated over time. Edge computing nodes periodically recalculate the comprehensive urgency score. Once the score crosses a threshold range, a level transition is triggered, and a rematch of the handling instruction is requested, achieving dynamic hierarchical classification. The multi-dimensional handling semantic embedding vector carried in the structured event message allows the instruction matching unit to directly enter the semantic layer of the four-layer mapping rule base for solution, without the need for secondary feature extraction. The carried true scale and diffusion trend index are used by the parameter layer to solve specific parameters such as the number of deployments and the proximity distance.

[0055] Understandably, this application tightly couples hierarchical logic with structured generation through the aforementioned weighted scoring model, leveraging the near-field computing capabilities of edge nodes to achieve millisecond-level real-time processing. This eliminates the subjectivity of manual descriptions and provides a clear and computable data foundation for subsequent automated processing instruction matching. Compared to existing technologies, the hierarchical factors of edge computing nodes introduce cumulative confidence and diffusion trends to support dynamic hierarchical transitions. Furthermore, the hierarchical factors solidify processing semantic embedding vectors and video segments within structured event messages, providing direct input and evidence for subsequent four-layer mapping hierarchical solutions and root cause analysis.

[0056] Step S50: Based on the four-layer decoupled mapping rule base consisting of semantic layer, policy layer, capability layer and parameter layer, the event type level code, severity level and multi-dimensional handling semantic embedding vector in the structured event message are used as search keys to perform a backtracking layered mapping to obtain a mapping trajectory, and the mapping trajectory is encapsulated into an instruction data packet.

[0057] Specifically, through the semantic layer, the event type level code, the severity level, and the multidimensional handling semantic embedding vector are combined into a triplet retrieval key, and matching is performed in the explicit rule table of the semantic layer based on the triplet retrieval key; If a match is found, the set of abstract handling intentions that are found in the explicit rule table is output. If no match is found, the nearest neighbor search is performed in the intention prototype vector library based on the multidimensional handling semantic embedding vector, the set of abstract handling intentions is output, and the situation of initiating the nearest neighbor search after the explicit rule is not found, along with the event type level code, the severity level, and the multidimensional handling semantic embedding vector, is reported to the cloud. Through the strategy layer, each disposal intention in the set of abstract disposal intentions is expanded into a directed acyclic graph of action templates. The nodes of the directed acyclic graph of action templates are abstract disposal actions, and the edges of the nodes are execution relationships. Through the capability layer, for each node in the directed acyclic graph of the action template, subset matching is performed in the device capability profile according to the set of capability labels carried by the node to obtain a set of feasible executors. For each candidate executor in the set of feasible executors, the cost function value is calculated, and the candidate executor with the smallest cost function value is selected as the executor. The executor is bound to the corresponding node to obtain the directed acyclic graph of the target action template after binding the executor. The cost function is obtained by weighting the estimated arrival time, the remaining power margin after the task is completed, the task preemption conflict cost, and the equipment loss balancing cost. Through the parameter layer, the abstract action of each node in the directed acyclic graph of the target action template is instantiated into specific instruction parameters. The specific instruction parameters are input into the safety envelope constraint solver for verification. If the verification passes, an action instance with the specific instruction parameters is obtained. If the verification fails, the specific instruction parameters are converged to the safety boundary. When the capability layer cannot match a feasible executor or the parameter layer determines that there is no feasible instance of the action node, backtrack to the strategy layer, select the suboptimal branch and re-execute all steps after the strategy layer is expanded, until a complete executable mapping link is obtained or the number of backtracking times reaches the upper limit. During the layer-by-layer execution process, the execution results of each layer are recorded, a mapping trajectory is generated, and after the execution is completed, the action instance with the specific instruction parameters is encapsulated into an instruction data packet according to the topological order of the directed acyclic graph of the action template.

[0058] In this embodiment, a triplet search key is formed using the event type hierarchy code, severity level, and multidimensional handling semantic embedding vector from the structured event message. The search is then performed in the semantic layer of the four-layer decoupled mapping rule base. The edge computing node first performs an exact match between the event type hierarchy code and severity level in the explicit rule table. If a match is found, the set of abstract handling intentions that match the explicit rule table is directly output. If no match is found, the edge computing node performs a nearest neighbor search in the intention prototype vector library using the multidimensional handling semantic embedding vector, outputs the set of abstract handling intentions by distance weighting, and reports the situation of initiating the nearest neighbor search after the explicit rule mismatch, along with the event type hierarchy code, severity level, and multidimensional handling semantic embedding vector, to the cloud for rule completion. The abstract handling intentions are device-independent abstract enumerations, including verification, warning, handling, protection, and notification. They can be combined and output with sequential and urgency labels. The same set of semantic layer rules can be directly reused across different machine models, different parks, and different projects.

[0059] Furthermore, each disposal intention in the set of abstract disposal intentions is expanded into a directed acyclic graph (DAG) of action templates. The nodes of the DAG represent abstract disposal actions, including close-range observation, surrounding modeling, verbal warning, flashing lights to drive away, delivery of emergency supplies, deployment of fire extinguishing devices, target tracking, area containment guidance, and notification and reporting. The edges of the nodes represent execution relationships and support four types: sequential edges, parallel edges, conditional edges, and timeout edges. Each node carries three sets of attributes: a set of required capability labels, preconditions and termination conditions, and a maximum execution time. Edge computing nodes express arbitrarily programmable, branchable, and conditionally aborted disposal workflows according to the DAG of action templates, upgrading disposal plans from single or fixed sequence actions to flexible workflows adaptable to complex field conditions.

[0060] The action template, a directed acyclic graph G, is the rule representation of the four-layer mapping rule base strategy layer, denoted as G=(V,E). The node set V represents abstract handling actions such as close-in observation, verbal warning, flashing lights to drive away, delivering emergency supplies, deploying fire extinguishing devices, target tracking, area control guidance, and reporting. The directed edge set E supports four types: serial edges, parallel edges, conditional edges, and timeout edges. Each node carries three sets of attributes: the required capability label set, preconditions and termination conditions, and the maximum execution time.

[0061] Next, for each node in the directed acyclic graph of the action template, subset matching is performed in the device capability profile according to the set of capability labels carried by the node to obtain a set of feasible executors. The device capability profile includes load capacity labels, maximum load and endurance, wind resistance level, current battery level, current location and altitude, airspace permissions and task occupancy status, which are refreshed and registered by drones or ground emergency equipment connected to the edge computing node at a rate of seconds. For each candidate executor in the set of feasible executors, the edge computing node calculates a cost function value weighted by four factors: estimated arrival time, remaining battery margin after task completion, task preemption conflict cost, and equipment loss balancing cost. The candidate executor with the smallest cost function value is selected as the executor, and the executor is bound to the action node corresponding to the candidate executor, resulting in a directed acyclic graph of the action template after binding the executors. If the set of feasible executors is empty, cross-layer backtracking is triggered.

[0062] Furthermore, the edge computing node instantiates the abstract actions of each node in the directed acyclic graph of the action template bound to the executor into specific command parameters. During instantiation, the edge computing node calculates the approach waypoint sequence, approach altitude, and minimum safe distance based on the target's geodetic coordinates and real-world scale; calculates the deployment altitude and horizontal lead based on real-time weather conditions; and verifies whether the waypoints fall within no-fly zones or obstacle ranges based on airspace constraints and the terrain digital elevation model. The edge computing node sends the specific command parameters to the safety envelope constraint solver for verification. For cases where the altitude limit is exceeded, the no-fly zone is entered, the wind resistance level is insufficient, or the battery power is insufficient to complete the return, the specific command parameters are first converged to the safe boundary. If the verification still fails after convergence, it is determined that there is no feasible instance for the corresponding action node and cross-layer backtracking is triggered; if the verification passes, an action instance with specific command parameters is obtained.

[0063] When an edge computing node cannot find a feasible executor at the capability layer or determines that there is no feasible instance of the action node at the parameter layer, it backtracks to the strategy layer, selects a suboptimal branch, and re-executes the strategy layer's expansion operation, the capability layer's executor matching and binding operation, and the parameter layer's instruction instantiation and verification operation until a complete executable mapping link is obtained or the backtracking count reaches the limit. The reason for each backtracking is recorded. Specific backtracking paths include: when the feasible executor set at the capability layer is empty, backtracking to the strategy layer to select an alternative action branch; when the safety envelope at the parameter layer cannot be satisfied, backtracking to the capability layer to select an executor with a higher wind resistance level or backtracking to the strategy layer to use a long-distance zoom observation branch that does not require close proximity; when all branches at the strategy layer are infeasible, backtracking to the semantic layer to downgrade the handling intention from handling to protection plus notification and report a request for manual intervention. Edge computing nodes apply dynamic mutex locks to mutually exclusive actions within the same target area. Actions that fail to acquire the lock are queued according to priority or trigger backtracking. During the rule configuration phase, static detection of overlapping input patterns and priority conflicts at the same layer is performed; if a conflict occurs, submission is rejected and a correction is prompted.

[0064] Furthermore, during the solution process, the edge computing nodes record the execution results of each layer to generate a mapping trajectory. This trajectory includes the rule identifiers and version numbers of each of the four layers, the candidate sets and eliminated branches for each layer, all backtracking nodes and their backtracking reasons, the finally selected executor and parameters, and the convergence process constrained by the safety envelope. After the solution is completed, the edge computing nodes encapsulate the action instances with specific instruction parameters into instruction data packets conforming to the communication protocol between the edge computing nodes and the drones, according to the topological order of the directed acyclic graph of the action template. These data packets include the instruction ID, the associated event ID, the mapping trajectory ID, the target device ID, the action type, a parameter JSON object, preconditions and termination conditions, the maximum execution time, a timestamp, and a checksum. The edge computing nodes then distribute the instruction data packets to the corresponding drones or ground emergency equipment via 5G or a private network.

[0065] As can be seen, edge computing nodes, through the aforementioned four-layer decoupled mapping and backtracking hierarchical solution mechanism, reconstruct the existing single-layer tiling key-value mapping into a four-level progressive mapping from abstract handling intentions to specific instruction parameters. This achieves complete decoupling between rules and hardware. New models only need to register device capability profiles without modifying any upper-layer rules. When devices are unavailable, airspace is restricted, or weather is unfavorable, alternative solutions are automatically obtained through cross-layer backtracking, significantly improving the feasibility of handling and scenario adaptability. Simultaneously, the mapping trajectory generated during the solution process is reported to the cloud platform along with structured event messages and archived with the handling report. This allows each automatically issued instruction to be traced and audited layer by layer, solving the interpretability deficiency of single-layer rule matching, which only provides the result but not the reasoning behind it. This provides a complete basis for incident review and responsibility determination.

[0066] Step S60: Send the instruction data packet to the UAV or ground equipment and receive the returned execution status, and confirm the next action to be executed based on the execution status.

[0067] Specifically, edge computing nodes transmit command data packets to corresponding drones or ground emergency equipment via 5G or private networks. After receiving and parsing the command data packets, the drones or ground emergency equipment execute the response actions and simultaneously transmit the execution status back to the edge computing nodes in real time. The execution status includes whether the command has been received, is in progress, the termination condition has been met, the execution timed out, the execution failed and the reason code for the failure, as well as on-site observation information such as current coordinates, sensor readings and keyframe images.

[0068] Furthermore, the edge computing node determines the subsequent action to be executed based on the returned execution status: when the termination condition is met, it enters the successor node along the serial edge or condition edge; when the execution times out, it jumps to the alternative node along the timeout edge; when the execution fails, it triggers a retry at the same layer or a backtracking across layers to solve the problem again according to the failure reason code. If the failure is due to insufficient power, it backtracks to the capability layer to select a different execution body. If the failure is due to excessive wind speed, it backtracks to the strategy layer to select a different action branch.

[0069] Step S70: Report the structured event message and the mapping trajectory to the cloud platform and receive the returned review conclusion. Update the detection model and the four-layer decoupling mapping rule base according to the review conclusion.

[0070] Specifically, edge computing nodes report structured event messages and mapping trajectories to the cloud platform. The mapping trajectory includes the rule identifiers and version numbers of each of the four layers, the candidate sets and eliminated branches of each layer, all backtracking nodes and backtracking reasons, the finally selected executor and parameters, and the safety envelope constraint convergence process. It is reported along with the structured event message and archived with the handling report, and the node receives the review conclusion returned by the cloud platform.

[0071] Furthermore, updating the detection model and the four-layer decoupling mapping rule base based on the review conclusion specifically includes: Receive the review conclusion returned by the cloud platform, the review conclusion including effective handling, invalid handling, false alarm and missed detection; The event-related video clips, detection results, and mapping trajectory in the structured event message corresponding to the review conclusion are uploaded to the cloud platform to perform an update operation; Receive the updated detection model and adjusted rule base issued by the cloud platform, and redeploy the updated detection model and adjusted rule base; The spatiotemporal-semantic fingerprint of events that are judged as false alarms is written into the false alarm immunity table. When subsequent observations hit the false alarm immunity table and the counter exceeds the suppression threshold, the event existence confidence of subsequent observations is downweighted. The update operation performed by the cloud platform is as follows: using the cloud-based high-parameter model as the teacher model and the risk-adaptive elastic lightweight detection model as the student model, the data containing the review conclusion, the event-related video clips, the detection results, and the mapping trajectory are used as an incremental sample set with treatment labels. This data is then input into the teacher-student distillation framework to construct the total distillation loss. The detection model is updated with the goal of minimizing the total distillation loss. The updated detection model is then distributed to edge computing nodes. Additionally, based on the rule identifier dimension, the treatment success rate, average treatment time, backtracking trigger rate, and manual intervention rate are statistically analyzed. The rule base is adjusted based on the statistical results, and the adjusted rule base is distributed to edge computing nodes.

[0072] Understandably, edge computing nodes receive review conclusions returned by the cloud platform. These conclusions are categorized into four types: effective handling, invalid handling, false positives, and missed detections. The edge computing nodes upload the event-related video clips, detection results, and mapping trajectories from the structured event messages corresponding to the review conclusions to the cloud platform, triggering a model update operation. Using the cloud-based high-parameter model as the teacher and the risk-adaptive, elastic, lightweight detection model deployed on the edge computing nodes as the student, the data containing the review conclusions, event-related video clips, detection results, and mapping trajectories is used as an incremental sample set with handling labels. This data is input into the teacher-student distillation framework to construct the total distillation loss. The total distillation loss is obtained by weighted summation of three terms: the multi-task loss of the handling attribute branch, the feature distillation loss between the teacher and student models, the output distillation loss, and the channel sparsity regularization term of the detection model. The detection model is updated with the goal of minimizing the total distillation loss, and the updated detection model is then distributed to the edge computing nodes.

[0073] Furthermore, the edge computing node receives the updated detection model from the cloud platform and deploys it to the edge node. The edge computing node also receives the rule base adjusted and issued by the cloud platform after statistically analyzing the success rate, average processing time, backtracking trigger rate, and manual intervention rate according to the rule identifier dimension. The adjustment methods include lowering the priority of rules with a success rate below a preset threshold, reducing the weight or taking offline strategy layer branches with a backtracking trigger rate above a preset threshold, and completing the corresponding explicit rules for event type hierarchical codes in the semantic layer that cause frequent triggering of nearest neighbor retrieval. The edge computing node then deploys the adjusted rules to the local rule base.

[0074] Furthermore, the spatiotemporal-semantic fingerprint of events that are judged as false alarms is written into the false alarm immunity table, with the event fingerprint truncated code as the key and a counter with time decay as the value. When subsequent observations hit the false alarm immunity table and the counter exceeds the suppression threshold, the edge computing node reduces the confidence of the subsequent observation of the event or directly silences it. The counter decays exponentially over time to avoid permanent missed alarms.

[0075] As can be seen, this application, through the complete operation of receiving the cloud platform's review conclusion and updating the detection model, rule base, and false alarm immunity table respectively, enables the review results to flow back along two paths. First, the review conclusion is multiplied by the event type cost coefficient and the regional sensitivity cost coefficient to form a sample-level cost weight injected into the distillation loss, driving the detection model to optimize according to the minimum handling cost rather than the minimum classification error, forming a self-evolving closed loop of perception-handling-review-retraining. Second, statistical data based on the rule dimension drives the rule base priority adjustment, explicit rule completion, and the de-weighting or removal of low success rate branches, making the rule base a decision asset that continuously optimizes itself with operation. At the same time, the event fingerprints judged as false alarms are synchronously written into the false alarm immunity table to achieve self-learning suppression of stubborn false alarms. Compared with the existing technology where AI models are trained offline once, the rule base is manually statically configured, and the handling results are only used for post-event archiving, this mechanism makes both the perception model and decision rules assets that continuously evolve with the running data. It can perform targeted optimization for high-cost missed detections and stubborn false alarms in this scenario, and this capability can only be established by relying on the unique handling review closed loop of the edge computing node.

[0076] The most prominent beneficial effects of this invention are as follows: (1) This invention deploys a risk-adaptive, elastic, and lightweight detection model on edge computing nodes. By embedding risk status gating units after each downsampling stage in the backbone network, the channel mask is dynamically generated using the risk status score of the region at the previous moment as input. This allows only about 30% of the channels to be activated during low-risk periods to achieve low-power cruise, and all channels to be activated during high-risk periods to achieve full-precision detection. The same set of model weights automatically adapts to computing resources under different risk levels, solving the inherent problem of static model computing power allocation being disconnected from business risks. At the same time, the detection head adds a disposal attribute branch in addition to the classification and regression branches, directly outputting the target's true scale, diffusion trend index, regional accessibility marker, and disposal semantic embedding vector. This moves the disposal judgment, which originally required downstream rules or manual completion, to the model, allowing the model output to carry dual semantics simultaneously. The perception results can directly support decision-making without secondary transformation, eliminating the semantic gap between the perception layer and the decision layer.

[0077] (2) This invention abandons the existing judgment paradigm based on the weighted summation of floating-point similarity in three dimensions of location, time, and type. Instead, it adopts a fixed-length composite coding structure of event spatiotemporal-semantic fingerprint. The merging and judgment are completed through three levels of cascading: type prefix filtering, spatial grid hash retrieval, and semantic segment Hamming distance comparison. The entire process is based on integer bit operations and hash table lookup, making the judgment computational load independent of the number of concurrent events. This eliminates the approximately quadratic computational expansion caused by pairwise comparisons within the candidate set in traditional schemes. In this invention, repeated observations are not discarded as redundant information but are continuously merged into the existing event evidence chain as evidence. This drives the event location to converge successively, the level to jump dynamically, and the disposal instructions to be rematched in a timely manner. At the same time, it has the geometric consistency verification capability of three-dimensional intersection of multi-machine observation rays and the self-learning suppression capability of the false alarm immunity table. This transforms repeated observations from a system burden into an effective resource for improving the accuracy of event location and level judgment.

[0078] (3) This invention reconstructs the traditional rigid key-value mapping into a four-level progressive decoupling system consisting of a semantic layer, a strategy layer, a capability layer, and a parameter layer. The semantic layer outputs abstract handling intentions independent of the device, enabling the reuse of the same set of rules across different device models. The strategy layer expands the intentions into a directed acyclic graph of action templates supporting four edge types: serial, parallel, conditional, and timeout. This upgrades the handling scheme from a fixed sequence to a flexible workflow that can be programmed, branched, and conditionally terminated. The capability layer selects the executor through capability tag subset matching and multi-factor cost function optimization, achieving complete decoupling between rules and hardware. The parameter layer instantiates actions into specific parameters within the safety envelope by combining on-site contexts such as meteorology, airspace, and terrain. When any layer has no feasible solution, it automatically backtracks to the previous layer to select a suboptimal branch and re-expands the solution. When constraints such as device loss, meteorological exceedances, or airspace control change, it automatically solves for alternative paths instead of directly declaring no solution, significantly improving the handling achievability rate under complex on-site conditions.

[0079] Furthermore, such as Figure 2 As shown, based on the above-mentioned cloud-edge collaborative emergency closed-loop handling method for industrial drones, the present invention also provides a cloud-edge collaborative emergency closed-loop handling system for industrial drones, wherein the cloud-edge collaborative emergency closed-loop handling system for industrial drones includes: The edge real-time detection module 51 is used to receive video streams collected by the UAV, identify the video streams based on the detection model, and output composite detection tuples. The composite detection tuples include anomaly category, pixel bounding box, multi-dimensional processing semantic embedding vector, target true scale estimation and diffusion trend index. The confirmation event acquisition module 52 is used to perform cross-frame evidence accumulation and double threshold hysteresis decision based on the composite detection tuple to obtain the confirmation event and the existence confidence of the confirmation event. After performing non-maximum suppression on the pixel bounding box, pixel-geometry back projection is performed in combination with the real-time pose of the UAV and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmation event. The merging and fingerprint generation module 53 is used to obtain an event type hierarchy code based on the anomaly category, generate an event spatiotemporal-semantic fingerprint based on the event type hierarchy code, the geodetic coordinates and the multidimensional processing semantic embedding vector, and perform a merging decision on the confirmed events based on the event spatiotemporal-semantic fingerprint to obtain merged event objects. The event message structuring module 54 is used to obtain the severity level based on the event type risk coefficient, regional sensitivity coefficient, existence confidence and diffusion trend index in the merged event object, and create a structured event message including the event type hierarchical code, the severity level, the geodetic coordinates, the multidimensional treatment semantic embedding vector, the target true scale estimate and the diffusion trend index; The decoupling mapping and encapsulation module 55 is used to perform a backtracking layered mapping to obtain a mapping trajectory based on a four-layer decoupling mapping rule base consisting of a semantic layer, a policy layer, a capability layer and a parameter layer, using the event type level code, the severity level and the multi-dimensional handling semantic embedding vector in the structured event message as the retrieval key, and encapsulate the mapping trajectory into an instruction data packet; The instruction issuance and execution module 56 is used to issue the instruction data packet to the UAV or ground equipment and receive the returned execution status, and determine the next action to be executed based on the execution status; The closed-loop and update module 57 is used to report the structured event message and the mapping trajectory to the cloud platform and receive the returned review conclusion, and update the detection model and the four-layer decoupled mapping rule base according to the review conclusion.

[0080] Furthermore, such as Figure 3 As shown, based on the above-mentioned cloud-edge collaborative emergency closed-loop handling method and system for industrial drones, the present invention also provides an edge computing node, which includes a processor 10, a memory 20 and a display 30. Figure 3 Only a portion of the components of an edge computing node are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0081] In some embodiments, the memory 20 may be an internal storage unit of the edge computing node, such as a hard drive or memory of the edge computing node. In other embodiments, the memory 20 may be an external storage device of the edge computing node, such as a plug-in hard drive, Smart MediaCard (SMC), Secure Digital (SD) card, or Flash Card equipped on the edge computing node. Further, the memory 20 may include both internal and external storage units of the edge computing node. The memory 20 is used to store application software and various types of data installed on the edge computing node, such as the program code installed on the edge computing node. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a cloud-edge collaborative industrial drone emergency closed-loop handling program 40, which can be executed by the processor 10 to implement the cloud-edge collaborative industrial drone emergency closed-loop handling method of this application.

[0082] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other drone control chip, used to run program code stored in the memory 20 or process data, such as executing the cloud-edge collaborative industrial drone emergency closed-loop disposal method.

[0083] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the edge computing node and to display a visual user interface. Components of the edge computing node communicate with each other via a system bus.

[0084] In one embodiment, when the processor 10 executes the cloud-edge collaborative industrial drone emergency closed-loop handling program 40 stored in the memory 20, the following steps are implemented: The system receives video streams collected by a drone, identifies the video streams based on a detection model, and outputs a composite detection tuple, which includes anomaly category, pixel bounding box, multidimensional processing semantic embedding vector, target true scale estimation, and diffusion trend index. Based on the composite detection tuple, cross-frame evidence accumulation and dual-threshold hysteresis decision are performed to obtain the confirmed event and the existence confidence of the confirmed event. After performing non-maximum suppression on the pixel bounding box, pixel-geometry back projection is performed in combination with the real-time pose of the UAV and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmed event. An event type hierarchy code is obtained based on the anomaly category. An event spatiotemporal-semantic fingerprint is generated based on the event type hierarchy code, the geodetic coordinates, and the multidimensional processing semantic embedding vector. The confirmed events are merged based on the event spatiotemporal-semantic fingerprint to obtain merged event objects. Based on the event type risk coefficient, regional sensitivity coefficient, existence confidence and diffusion trend index in the merged event object, a severity level is obtained, and a structured event message is created that includes the event type hierarchical code, the severity level, the geodetic coordinates, the multidimensional treatment semantic embedding vector, the target true scale estimate and the diffusion trend index; Based on a four-layer decoupled mapping rule base consisting of a semantic layer, a policy layer, a capability layer, and a parameter layer, the event type level code, the severity level, and the multi-dimensional handling semantic embedding vector in the structured event message are used as search keys to perform a backtracking layered mapping to obtain a mapping trajectory, and the mapping trajectory is encapsulated into an instruction data packet. The instruction data packet is sent to the UAV or ground equipment and the execution status is received back. The next action to be performed is determined based on the execution status. The structured event message and the mapping trajectory are reported to the cloud platform and the verification conclusion is received. The detection model and the four-layer decoupling mapping rule base are updated according to the verification conclusion.

[0085] The detection model is a risk-adaptive, resilient, lightweight detection model, which includes a backbone network, a neck, and a detection head. The backbone network includes multiple downsampling stages. After each downsampling stage, a risk situation gating unit is set up. The risk situation gating unit is used to take the channel statistics and the risk situation score of the current region at the previous moment as input, and output the channel mask through a fully connected and binarized operator. It dynamically determines the subset of channels participating in the convolution operation in the current stage, so that when the risk situation score is lower than the low risk threshold, the same set of model weights retains some channels and reduces the inference frame rate, and when the risk situation score exceeds the high risk threshold, it retains all channels and increases the inference frame rate. The risk situation score is obtained by weighting the highest severity level of confirmed events in the region at the previous moment, the number of active events, and the false alarm immunity table hit rate. The neck includes a feature pyramid and a path aggregation network, used to fuse multi-scale features selected by the channel mask. During the fusion process, the positions of channels closed by the channel mask are occupied with zero values. The detection head includes a classification branch, a bounding box regression branch, and a treatment attribute branch. The classification branch is used to output the anomaly category, the bounding box regression branch is used to output the pixel bounding box, and the treatment attribute branch is used to output the multidimensional treatment semantic embedding vector, the target true scale estimate, and the diffusion trend index.

[0086] Specifically, the process of performing cross-frame evidence accumulation and dual-threshold hysteresis decision based on the composite detection tuple to obtain the confirmed event and its existence confidence, performing non-maximum suppression on the pixel bounding box, and then performing pixel-to-geometry back projection based on the UAV's real-time pose and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmed event includes: For each composite detection tuple in the current frame of the video stream, the detection confidence of the current frame and the accumulated event presence confidence of the previous frame are weighted and recursively calculated using a forgetting factor, and the detection confidence is modulated using a frame quality factor to obtain the event presence confidence of the current frame; wherein, the forgetting factor is a preset constant, and the frame quality factor is obtained by normalizing motion blur, occlusion ratio, and target pixel area; When the confidence level of the event meets the activation threshold, the detection object corresponding to the composite detection tuple is determined as the confirmed event, and the confirmed event and the event existence confidence level of the confirmed event are output. When the confidence level of the event decreases to the closing threshold and continues for a preset frame, the confirmed event is determined to be resolved, wherein the closing threshold is less than the opening threshold; Based on the confidence level of the event, the overlapping pixel bounding boxes are sorted, the pixel bounding box with the highest confidence level of the event is retained and the other overlapping boxes are suppressed to obtain the target pixel bounding box after non-maximum suppression. An observation ray is constructed based on the real-time pose of the UAV, the gimbal pointing angle, and the camera intrinsic parameters. The midpoint of the bottom edge of the target pixel bounding box is projected onto the ground along the observation ray to obtain the projection point. The latitude and longitude of the midpoint of the bottom edge are intersected with the digital elevation model to obtain the elevation of the midpoint of the bottom edge. The latitude and longitude and the elevation are used as the geodetic coordinates of the confirmed event. Error propagation calculations are performed based on the pose angle error, positioning error, and ranging error of the UAV in real time to obtain the positioning uncertainty ellipse of the confirmed event.

[0087] Specifically, the process of obtaining an event type hierarchy code based on the anomaly category, generating an event spatiotemporal-semantic fingerprint based on the event type hierarchy code, the geodetic coordinates, and the multidimensional processing semantic embedding vector, and performing a merging decision on the confirmed events based on the event spatiotemporal-semantic fingerprint to obtain merged event objects includes: The event type hierarchy code is obtained based on the anomaly category mapping. The influence radius is determined based on the event type hierarchy code. The influence radius is used as the grid side length. The geodetic coordinates are partitioned into a hierarchical quadtree. The partitioning result is encoded into a fixed-length bit string. The fixed-length bit string is used as the spatial segment of the event spatiotemporal-semantic fingerprint. The neighborhood grid key of the grid is generated. The event type hierarchy code is used as the type segment of the event spatiotemporal-semantic fingerprint, and the multidimensional processing semantic embedding vector is binarized using local sensitive hashing to obtain the semantic segment of the event spatiotemporal-semantic fingerprint. By concatenating the spatial segment, the type segment, and the semantic segment, the event spatiotemporal-semantic fingerprint is obtained. Prefix filtering is performed according to the type segment, and hash retrieval is performed in the event index table using the spatial segment and the neighborhood raster key to obtain a set of candidate active events; For each existing event in the candidate active event set, determine whether the Hamming distance of the semantic segment is not greater than a preset threshold, whether the observation time of the confirmed event falls within the time active window of the existing event, and whether the distance between the geodetic coordinates of the confirmed event and the last position of the existing event is not greater than the product of the maximum moving speed and the time difference corresponding to the event type hierarchical code. If all three conditions are met, the confirmed event is determined to be mergeable evidence. The positions of the existing events are weighted and fused using the reciprocal of the positioning uncertainty ellipse as the weight, and the existing events are used as the merged event objects. Otherwise, a new event object is created based on the confirmed event, and the newly created event object is used as the merged event object. When performing prefix filtering according to the type segment, if the major categories of the type segments are different, they are directly determined to be different events and the search ends.

[0088] Specifically, the step of obtaining the severity level based on the event type risk coefficient, regional sensitivity coefficient, existence confidence level, and diffusion trend index of the merged event objects includes: Obtain the event type risk coefficient, regional sensitivity coefficient, event existence confidence, and diffusion trend index from the merged event objects; The confidence level of the event is normalized to a preset interval to obtain the cumulative confidence coefficient, and the diffusion trend index is normalized to a preset interval to obtain the diffusion trend coefficient. The comprehensive urgency score is obtained by multiplying the event type risk coefficient, the regional sensitivity coefficient, the cumulative confidence coefficient, and the diffusion trend coefficient by their respective preset weights and then summing them. The severity level is determined based on the threshold range into which the comprehensive urgency score falls. When the comprehensive urgency score is not less than the first threshold, it is determined to be an emergency level. When the comprehensive urgency score is between the second threshold and the first threshold, it is determined to be a normal level. When the comprehensive urgency score is less than the second threshold, it is determined to be a warning level.

[0089] Specifically, the four-layer decoupled mapping rule base consisting of a semantic layer, a policy layer, a capability layer, and a parameter layer is used as the retrieval key. A backtracking layered mapping is performed to obtain a mapping trajectory, which is then encapsulated into an instruction data packet. This includes: Through the semantic layer, the event type hierarchy code, the severity level, and the multidimensional handling semantic embedding vector are combined into a triplet retrieval key, and matching is performed in the explicit rule table of the semantic layer based on the triplet retrieval key; If a match is found, the set of abstract handling intentions that are found in the explicit rule table is output. If no match is found, the nearest neighbor search is performed in the intention prototype vector library based on the multidimensional handling semantic embedding vector, the set of abstract handling intentions is output, and the situation of initiating the nearest neighbor search after the explicit rule is not found, along with the event type level code, the severity level, and the multidimensional handling semantic embedding vector, is reported to the cloud. Through the strategy layer, each disposal intention in the set of abstract disposal intentions is expanded into a directed acyclic graph of action templates. The nodes of the directed acyclic graph of action templates are abstract disposal actions, and the edges of the nodes are execution relationships. Through the capability layer, for each node in the directed acyclic graph of the action template, subset matching is performed in the device capability profile according to the set of capability labels carried by the node to obtain a set of feasible executors. For each candidate executor in the set of feasible executors, the cost function value is calculated, and the candidate executor with the smallest cost function value is selected as the executor. The executor is bound to the corresponding node to obtain the directed acyclic graph of the target action template after binding the executor. The cost function is obtained by weighting the estimated arrival time, the remaining power margin after the task is completed, the task preemption conflict cost, and the equipment loss balancing cost. Through the parameter layer, the abstract action of each node in the directed acyclic graph of the target action template is instantiated into specific instruction parameters. The specific instruction parameters are input into the safety envelope constraint solver for verification. If the verification passes, an action instance with the specific instruction parameters is obtained. If the verification fails, the specific instruction parameters are converged to the safety boundary. When the capability layer cannot match a feasible executor or the parameter layer determines that there is no feasible instance of the action node, backtrack to the strategy layer, select the suboptimal branch and re-execute all steps after the strategy layer is expanded, until a complete executable mapping link is obtained or the number of backtracking times reaches the upper limit. During the layer-by-layer execution process, the execution results of each layer are recorded, a mapping trajectory is generated, and after the execution is completed, the action instance with the specific instruction parameters is encapsulated into an instruction data packet according to the topological order of the directed acyclic graph of the action template.

[0090] Specifically, updating the detection model and the four-layer decoupling mapping rule base based on the review conclusion includes: Receive the review conclusion returned by the cloud platform, the review conclusion including effective handling, invalid handling, false alarm and missed detection; The event-related video clips, detection results, and mapping trajectory in the structured event message corresponding to the review conclusion are uploaded to the cloud platform to perform an update operation; Receive the updated detection model and adjusted rule base issued by the cloud platform, and redeploy the updated detection model and adjusted rule base; The spatiotemporal-semantic fingerprint of events that are judged as false alarms is written into the false alarm immunity table. When subsequent observations hit the false alarm immunity table and the counter exceeds the suppression threshold, the event existence confidence of subsequent observations is downweighted. The update operation performed by the cloud platform is as follows: using the cloud-based high-parameter model as the teacher model and the risk-adaptive elastic lightweight detection model as the student model, the data containing the review conclusion, the event-related video clips, the detection results, and the mapping trajectory are used as an incremental sample set with treatment labels. This data is then input into the teacher-student distillation framework to construct the total distillation loss. The detection model is updated with the goal of minimizing the total distillation loss. The updated detection model is then distributed to edge computing nodes. Additionally, based on the rule identifier dimension, the treatment success rate, average treatment time, backtracking trigger rate, and manual intervention rate are statistically analyzed. The rule base is adjusted based on the statistical results, and the adjusted rule base is distributed to edge computing nodes.

[0091] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an emergency closed-loop handling program for industrial drones based on cloud-edge collaboration, and when the emergency closed-loop handling program for industrial drones based on cloud-edge collaboration is executed by a processor, it implements the steps of the emergency closed-loop handling method for industrial drones based on cloud-edge collaboration as described above.

[0092] In summary, this invention provides a cloud-edge collaborative emergency closed-loop handling method, system, edge computing node, and storage medium for industrial drones. The method includes: receiving video streams and performing real-time detection, outputting composite detection tuples; performing cross-frame evidence accumulation and dual-threshold decision based on the tuples to obtain confirmed events and their confidence levels, and obtaining geodetic coordinates and positioning uncertainty ellipses through back projection; obtaining event type hierarchical codes according to anomaly categories, generating spatiotemporal-semantic fingerprints by combining geodetic coordinates and handling semantic embedding vectors, and performing merged decision; normalizing the confidence level and diffusion trend index, and performing weighted scoring by combining risk coefficients and sensitivity coefficients to determine the severity level, creating structured event messages; performing hierarchical solution with backtracking based on a four-layer mapping rule base, generating mapping trajectories and encapsulating instruction data packets; issuing instruction data packets, updating the detection model and rule base according to the returned execution status and review conclusions, and completing the closed-loop handling. This invention can reduce the response latency of emergency closed-loop handling of industrial drones, reduce bandwidth consumption, improve decision accuracy, and achieve automated and interpretable handling decisions.

[0093] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or edge computing node that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or edge computing node. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or edge computing node that includes that element.

[0094] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0095] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A closed-loop emergency response method for industrial drones based on cloud-edge collaboration, characterized in that, The aforementioned cloud-edge collaborative emergency closed-loop response method for industrial drones includes: The system receives video streams collected by a drone, identifies the video streams based on a detection model, and outputs a composite detection tuple, which includes anomaly category, pixel bounding box, multidimensional processing semantic embedding vector, target true scale estimation, and diffusion trend index. Based on the composite detection tuple, cross-frame evidence accumulation and dual-threshold hysteresis decision are performed to obtain the confirmed event and the existence confidence of the confirmed event. After performing non-maximum suppression on the pixel bounding box, pixel-geometry back projection is performed in combination with the real-time pose of the UAV and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmed event. An event type hierarchy code is obtained based on the anomaly category. An event spatiotemporal-semantic fingerprint is generated based on the event type hierarchy code, the geodetic coordinates, and the multidimensional processing semantic embedding vector. The confirmed events are merged based on the event spatiotemporal-semantic fingerprint to obtain merged event objects. Based on the event type risk coefficient, regional sensitivity coefficient, existence confidence and diffusion trend index in the merged event object, a severity level is obtained, and a structured event message is created that includes the event type hierarchical code, the severity level, the geodetic coordinates, the multidimensional treatment semantic embedding vector, the target true scale estimate and the diffusion trend index; Based on a four-layer decoupled mapping rule base consisting of a semantic layer, a policy layer, a capability layer, and a parameter layer, the event type level code, the severity level, and the multi-dimensional handling semantic embedding vector in the structured event message are used as search keys to perform a backtracking layered mapping to obtain a mapping trajectory, and the mapping trajectory is encapsulated into an instruction data packet. The instruction data packet is sent to the UAV or ground equipment and the execution status is received back. The next action to be performed is determined based on the execution status. The structured event message and the mapping trajectory are reported to the cloud platform and the verification conclusion is received. The detection model and the four-layer decoupling mapping rule base are updated according to the verification conclusion.

2. The emergency closed-loop response method for industrial drones based on cloud-edge collaboration according to claim 1, characterized in that, The detection model is a risk-adaptive, resilient, lightweight detection model, which includes: a backbone network, a neck, and a detection head; The backbone network includes multiple downsampling stages. After each downsampling stage, a risk situation gating unit is set up. The risk situation gating unit is used to take the channel statistics and the risk situation score of the current region at the previous moment as input, and output the channel mask through a fully connected and binarized operator. It dynamically determines the subset of channels participating in the convolution operation in the current stage, so that when the risk situation score is lower than the low risk threshold, the same set of model weights retains some channels and reduces the inference frame rate, and when the risk situation score exceeds the high risk threshold, it retains all channels and increases the inference frame rate. The risk situation score is obtained by weighting the highest severity level of confirmed events in the region at the previous moment, the number of active events, and the false alarm immunity table hit rate. The neck includes a feature pyramid and a path aggregation network, used to fuse multi-scale features selected by the channel mask. During the fusion process, the positions of channels closed by the channel mask are occupied with zero values. The detection head includes a classification branch, a bounding box regression branch, and a treatment attribute branch. The classification branch is used to output the anomaly category, the bounding box regression branch is used to output the pixel bounding box, and the treatment attribute branch is used to output the multidimensional treatment semantic embedding vector, the target true scale estimate, and the diffusion trend index.

3. The emergency closed-loop response method for industrial drones based on cloud-edge collaboration according to claim 1, characterized in that, The process involves cross-frame evidence accumulation and dual-threshold hysteresis decision based on the composite detection tuple to obtain the confirmed event and its existence confidence. After performing non-maximum suppression on the pixel bounding box, pixel-to-geometry back projection is performed using the UAV's real-time pose and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmed event. Specifically, this includes: For each composite detection tuple in the current frame of the video stream, the detection confidence of the current frame and the accumulated event presence confidence of the previous frame are weighted and recursively calculated using a forgetting factor, and the detection confidence is modulated using a frame quality factor to obtain the event presence confidence of the current frame; wherein, the forgetting factor is a preset constant, and the frame quality factor is obtained by normalizing motion blur, occlusion ratio, and target pixel area; When the confidence level of the event meets the activation threshold, the detection object corresponding to the composite detection tuple is determined as the confirmed event, and the confirmed event and the event existence confidence level of the confirmed event are output. When the confidence level of the event decreases to the closing threshold and continues for a preset frame, the confirmed event is determined to be resolved, wherein the closing threshold is less than the opening threshold; Based on the confidence level of the event, the overlapping pixel bounding boxes are sorted, the pixel bounding box with the highest confidence level of the event is retained and the other overlapping boxes are suppressed to obtain the target pixel bounding box after non-maximum suppression. An observation ray is constructed based on the real-time pose of the UAV, the gimbal pointing angle, and the camera intrinsic parameters. The midpoint of the bottom edge of the target pixel bounding box is projected onto the ground along the observation ray to obtain the projection point. The latitude and longitude of the midpoint of the bottom edge are intersected with the digital elevation model to obtain the elevation of the midpoint of the bottom edge. The latitude and longitude and the elevation are used as the geodetic coordinates of the confirmed event. Error propagation calculations are performed based on the pose angle error, positioning error, and ranging error of the UAV in real time to obtain the positioning uncertainty ellipse of the confirmed event.

4. The emergency closed-loop response method for industrial drones based on cloud-edge collaboration according to claim 1, characterized in that, The process of obtaining an event type hierarchy code based on the anomaly category, generating an event spatiotemporal-semantic fingerprint based on the event type hierarchy code, the geodetic coordinates, and the multidimensional processing semantic embedding vector, and performing a merging decision on the confirmed events based on the event spatiotemporal-semantic fingerprint to obtain merged event objects specifically includes: The event type hierarchy code is obtained based on the anomaly category mapping. The influence radius is determined based on the event type hierarchy code. The influence radius is used as the grid side length. The geodetic coordinates are partitioned into a hierarchical quadtree. The partitioning result is encoded into a fixed-length bit string. The fixed-length bit string is used as the spatial segment of the event spatiotemporal-semantic fingerprint. The neighborhood grid key of the grid is generated. The event type hierarchy code is used as the type segment of the event spatiotemporal-semantic fingerprint, and the multidimensional processing semantic embedding vector is binarized using local sensitive hashing to obtain the semantic segment of the event spatiotemporal-semantic fingerprint. By concatenating the spatial segment, the type segment, and the semantic segment, the event spatiotemporal-semantic fingerprint is obtained. Prefix filtering is performed according to the type segment, and hash retrieval is performed in the event index table using the spatial segment and the neighborhood raster key to obtain a set of candidate active events; For each existing event in the candidate active event set, determine whether the Hamming distance of the semantic segment is not greater than a preset threshold, whether the observation time of the confirmed event falls within the time active window of the existing event, and whether the distance between the geodetic coordinates of the confirmed event and the last position of the existing event is not greater than the product of the maximum moving speed and the time difference corresponding to the event type hierarchical code. If all three conditions are met, the confirmed event is determined to be mergeable evidence. The positions of the existing events are weighted and fused using the reciprocal of the positioning uncertainty ellipse as the weight, and the existing events are used as the merged event objects. Otherwise, a new event object is created based on the confirmed event, and the newly created event object is used as the merged event object. When performing prefix filtering according to the type segment, if the major categories of the type segments are different, they are directly determined to be different events and the search ends.

5. The emergency closed-loop response method for industrial drones based on cloud-edge collaboration according to claim 1, characterized in that, The severity level is determined based on the event type risk coefficient, regional sensitivity coefficient, existence confidence level, and diffusion trend index of the merged event objects, specifically including: Obtain the event type risk coefficient, regional sensitivity coefficient, event existence confidence level, and diffusion trend index from the merged event objects; The confidence level of the event is normalized to a preset interval to obtain the cumulative confidence coefficient, and the diffusion trend index is normalized to a preset interval to obtain the diffusion trend coefficient. The comprehensive urgency score is obtained by multiplying the event type risk coefficient, the regional sensitivity coefficient, the cumulative confidence coefficient, and the diffusion trend coefficient by their respective preset weights and then summing them. The severity level is determined based on the threshold range into which the comprehensive urgency score falls. When the comprehensive urgency score is not less than the first threshold, it is determined to be an emergency level. When the comprehensive urgency score is between the second threshold and the first threshold, it is determined to be a normal level. When the comprehensive urgency score is less than the second threshold, it is determined to be a warning level.

6. The emergency closed-loop response method for industrial drones based on cloud-edge collaboration according to claim 1, characterized in that, The four-layer decoupled mapping rule base, consisting of a semantic layer, a policy layer, a capability layer, and a parameter layer, uses the event type hierarchy code, the severity level, and the multi-dimensional handling semantic embedding vector in the structured event message as search keys to perform a backtracking layered mapping to obtain a mapping trajectory. The mapping trajectory is then encapsulated into an instruction data packet, specifically including: Through the semantic layer, the event type hierarchy code, the severity level, and the multidimensional handling semantic embedding vector are combined into a triplet retrieval key, and matching is performed in the explicit rule table of the semantic layer based on the triplet retrieval key; If a match is found, the set of abstract handling intentions that are found in the explicit rule table is output. If no match is found, the nearest neighbor search is performed in the intention prototype vector library based on the multidimensional handling semantic embedding vector, the set of abstract handling intentions is output, and the situation of initiating the nearest neighbor search after the explicit rule is not found, along with the event type level code, the severity level, and the multidimensional handling semantic embedding vector, is reported to the cloud. Through the strategy layer, each disposal intention in the set of abstract disposal intentions is expanded into a directed acyclic graph of action templates. The nodes of the directed acyclic graph of action templates are abstract disposal actions, and the edges of the nodes are execution relationships. Through the capability layer, for each node in the directed acyclic graph of the action template, subset matching is performed in the device capability profile according to the set of capability labels carried by the node to obtain a set of feasible executors. For each candidate executor in the set of feasible executors, the cost function value is calculated, and the candidate executor with the smallest cost function value is selected as the executor. The executor is bound to the corresponding node to obtain the directed acyclic graph of the target action template after binding the executor. The cost function is obtained by weighting the estimated arrival time, the remaining power margin after the task is completed, the task preemption conflict cost, and the equipment loss balancing cost. Through the parameter layer, the abstract action of each node in the directed acyclic graph of the target action template is instantiated into specific instruction parameters. The specific instruction parameters are input into the safety envelope constraint solver for verification. If the verification passes, an action instance with the specific instruction parameters is obtained. If the verification fails, the specific instruction parameters are converged to the safety boundary. When the capability layer cannot match a feasible executor or the parameter layer determines that there is no feasible instance of the action node, backtrack to the strategy layer, select the suboptimal branch and re-execute all steps after the strategy layer is expanded, until a complete executable mapping link is obtained or the number of backtracking times reaches the upper limit. During the layer-by-layer execution process, the execution results of each layer are recorded, a mapping trajectory is generated, and after the execution is completed, the action instance with the specific instruction parameters is encapsulated into an instruction data packet according to the topological order of the directed acyclic graph of the action template.

7. The emergency closed-loop response method for industrial drones based on cloud-edge collaboration according to claim 2, characterized in that, The step of updating the detection model and the four-layer decoupling mapping rule base based on the review conclusion specifically includes: Receive the review conclusion returned by the cloud platform, the review conclusion including effective handling, invalid handling, false alarm and missed detection; The event-related video clips, detection results, and mapping trajectory in the structured event message corresponding to the review conclusion are uploaded to the cloud platform to perform an update operation; Receive the updated detection model and adjusted rule base issued by the cloud platform, and redeploy the updated detection model and adjusted rule base; The spatiotemporal-semantic fingerprint of events that are judged as false alarms is written into the false alarm immunity table. When subsequent observations hit the false alarm immunity table and the counter exceeds the suppression threshold, the event existence confidence of subsequent observations is downweighted. The update operation performed by the cloud platform is as follows: using the cloud-based high-parameter model as the teacher model and the risk-adaptive elastic lightweight detection model as the student model, the data containing the review conclusion, the event-related video clips, the detection results, and the mapping trajectory are used as an incremental sample set with treatment labels. This data is then input into the teacher-student distillation framework to construct the total distillation loss. The detection model is updated with the goal of minimizing the total distillation loss. The updated detection model is then distributed to edge computing nodes. Additionally, based on the rule identifier dimension, the treatment success rate, average treatment time, backtracking trigger rate, and manual intervention rate are statistically analyzed. The rule base is adjusted based on the statistical results, and the adjusted rule base is distributed to edge computing nodes.

8. An emergency closed-loop response system for industrial drones based on cloud-edge collaboration, characterized in that, The cloud-edge collaborative industrial drone emergency closed-loop response system includes: The edge real-time detection module is used to receive video streams collected by the UAV, identify the video streams based on the detection model, and output a composite detection tuple. The composite detection tuple includes anomaly category, pixel bounding box, multi-dimensional processing semantic embedding vector, target true scale estimation, and diffusion trend index. The confirmation event acquisition module is used to perform cross-frame evidence accumulation and double threshold hysteresis decision based on the composite detection tuple to obtain the confirmation event and the existence confidence of the confirmation event. After performing non-maximum suppression on the pixel bounding box, pixel-geometry back projection is performed in combination with the real-time pose of the UAV and camera parameters to obtain the geodetic coordinates and positioning uncertainty ellipse of the confirmation event. The merging and fingerprint generation module is used to obtain an event type hierarchy code based on the anomaly category, generate an event spatiotemporal-semantic fingerprint based on the event type hierarchy code, the geodetic coordinates, and the multidimensional processing semantic embedding vector, and perform a merging decision on the confirmed events based on the event spatiotemporal-semantic fingerprint to obtain merged event objects. The event message structuring module is used to obtain the severity level based on the event type risk coefficient, regional sensitivity coefficient, existence confidence and diffusion trend index in the merged event object, and create a structured event message including the event type hierarchical code, the severity level, the geodetic coordinates, the multidimensional treatment semantic embedding vector, the target true scale estimate and the diffusion trend index; The decoupling mapping and encapsulation module is used to perform a backtracking layered mapping to obtain a mapping trajectory based on a four-layer decoupling mapping rule base consisting of a semantic layer, a policy layer, a capability layer, and a parameter layer, using the event type level code, the severity level, and the multi-dimensional handling semantic embedding vector in the structured event message as the retrieval key, and encapsulate the mapping trajectory into an instruction data packet. The instruction issuance and execution module is used to issue the instruction data packet to the UAV or ground equipment and receive the returned execution status, and determine the next action to be executed based on the execution status; The closed-loop and update module is used to report the structured event message and the mapping trajectory to the cloud platform and receive the returned review conclusion, and update the detection model and the four-layer decoupled mapping rule base according to the review conclusion.

9. An edge computing node, characterized in that, The edge computing node includes: a memory, a processor, and a cloud-edge collaborative industrial drone emergency closed-loop handling program stored in the memory and executable on the processor. When the cloud-edge collaborative industrial drone emergency closed-loop handling program is executed by the processor, it implements the steps of the cloud-edge collaborative industrial drone emergency closed-loop handling method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an emergency closed-loop handling program for industrial drones based on cloud-edge collaboration. When the emergency closed-loop handling program for industrial drones based on cloud-edge collaboration is executed by a processor, it implements the steps of the emergency closed-loop handling method for industrial drones based on cloud-edge collaboration as described in any one of claims 1-7.

Images