A multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making

Abstract

This invention relates to the field of logistics park management technology, specifically to a multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making. The method includes: pre-dividing multiple inspection units according to the geographic spatial structure of the target area and defining the path connection relationships between each inspection unit; acquiring candidate anomalies collected by multiple inspection robots performing inspection tasks along preset paths, and constructing candidate anomalies that meet preset spatiotemporal merging conditions into collaborative inspection segments; determining the transmission necessity based on the coverage association status and the degree of evidence complementarity of the collaborative inspection segments; determining a data upload strategy based on the transmission necessity, and sending the anomaly data corresponding to the collaborative inspection segments to the cloud based on the data upload strategy; and issuing disposal task instructions to the disposal robot in response to the disposal needs fed back from the cloud. This invention improves the data transmission efficiency and reliability of anomaly handling in multi-robot inspections.

Description

Technical Field

[0001] This invention relates to the field of logistics park management technology, specifically to a multi-robot intelligent inspection method that combines edge-side perception and cloud-side decision-making. Background Technology

[0002] With the rapid development of the logistics industry, logistics parks are constantly expanding in scale, encompassing various functional areas such as road access, loading and unloading platforms, cargo yards, conveyor belt sections, and equipment areas, leading to increasingly complex inspection needs. To ensure park safety and continuous equipment operation, multiple inspection robots are typically deployed, each moving along a pre-set path to inspect its assigned area and collect multi-source sensing data such as video images and thermal imaging. Because a single robot is limited by its fixed path and sensor configuration, it often only covers a portion of the park and its perspective, making it difficult to form a complete observation of the same abnormal object. Therefore, multiple robots need to work collaboratively to improve overall inspection coverage.

[0003] In existing technologies, edge-cloud collaborative inspection systems typically use lightweight edge models to initially screen data collected by robots, uploading suspected anomalies to the cloud for verification. However, when transmitting and processing anomaly candidate results collected by multiple robots, each robot usually uploads independently, lacking effective analysis of data redundancy and complementarity among multiple robots. This approach leads to multiple robots repeatedly uploading similar data for the same anomaly, consuming significant communication bandwidth and cloud computing resources. Furthermore, some anomalies requiring multi-view, multi-sensor combinations for confirmation are missed or misjudged due to a lack of supplementary data from different robots, thus affecting the overall efficiency and reliability of the inspection system. Summary of the Invention

[0004] To address the technical problems of high redundancy in edge-cloud collaborative data transmission and insufficient reliability in anomaly handling during multi-robot inspection, this invention aims to provide a multi-robot intelligent inspection method that combines edge-side perception and cloud-side decision-making. The specific technical solution adopted is as follows: Firstly, a multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making is provided. This method includes: pre-dividing multiple inspection units according to the geospatial structure of the target area and defining the path connection relationships between each inspection unit; acquiring anomaly candidate results collected by multiple inspection robots performing inspection tasks along preset paths; constructing collaborative inspection segments based on the inspection units pointed to by each anomaly candidate result and the path connection relationships, provided that the anomaly candidate results meet preset spatiotemporal merging conditions; determining the transmission necessity based on the coverage association state and the degree of evidence complementarity of the collaborative inspection segments; the coverage association state characterizes the degree of spatial collaborative coverage of the same anomaly object by multiple robots, and the degree of evidence complementarity characterizes the scale of non-redundant evidence among each anomaly candidate result; determining a data upload strategy based on the transmission necessity, and sending the anomaly data corresponding to the collaborative inspection segments to the cloud based on the data upload strategy, the data upload strategy including combinations of uploaded content; and issuing disposal task instructions to the disposal robot in response to the disposal needs fed back from the cloud.

[0005] In one possible design, multiple inspection units are pre-divided according to the geospatial structure of the target area, and the path connection relationship between each inspection unit is defined, including: dividing the target area into multiple inspection units according to the spatial boundaries of each functional area in the target area, with each inspection unit covering a functional area or a part of a functional area; and defining adjacent inspection units with continuous observation relationship on the inspection path as inspection units with path connection relationship.

[0006] In one possible design, multiple inspection robots collect abnormal candidate results along a preset path while performing inspection tasks. Based on the inspection unit pointed to by each abnormal candidate result and the path connection relationship, abnormal candidate results that meet preset spatiotemporal merging conditions are constructed into a collaborative inspection segment. This includes: acquiring abnormal candidate results output by each inspection robot after initial screening of the collected images through a lightweight edge model during the execution of the inspection task; the abnormal candidate results include anomaly category, confidence level, collection timestamp, robot identifier, and anomaly location; mapping each abnormal candidate result to the corresponding inspection unit based on the anomaly location, and determining the inspection unit identifier corresponding to each abnormal candidate result; identifying abnormal candidate results whose mapped inspection unit is the target inspection unit or an inspection unit with a path connection relationship to the target inspection unit, and whose collection time interval is less than a preset time threshold, as abnormal candidate results that meet preset spatiotemporal merging conditions, and merging them into the same collaborative inspection segment, where the target inspection unit is any inspection unit pointed to by any abnormal candidate result.

[0007] In one possible design, determining the coverage association status of a collaborative inspection segment includes: identifying candidate anomaly pairs from the collaborative inspection segment that point to the same anomaly object through anomaly category comparison and anomaly location comparison, as valid observation pairs, and determining the path acceptance ratio based on the number of valid observation pairs; marking the visible surface categories of the anomaly objects covered by each candidate anomaly result, counting the number of different visible surface categories as the number of filler surfaces, and determining the filler coverage rate based on the number of filler surfaces and the maximum number of visible surfaces allowed under the inspection unit to which the same anomaly object belongs, where visible surface categories include front side, side surface, back filler surface, and thermal filler surface; and determining the coverage association status based on the total number of candidate anomaly results contained in the collaborative inspection segment, the path acceptance ratio, and the filler coverage rate.

[0008] In one possible design, determining the degree of evidence complementarity for a collaborative inspection segment includes: selecting the anomaly candidate result with the highest confidence and image quality meeting preset conditions in the collaborative inspection segment as the representative result; for other anomaly candidate results in the collaborative inspection segment besides the representative result, determining a first number of anomaly candidate results that have new evidence relative to the representative result, the new evidence including new visible surface categories, thermal imaging information, or local details; determining a second number of anomaly candidate results whose anomaly category and anomaly location are consistent with the anomaly object pointed to by the representative result; determining the ratio of the first number to the number of other anomaly candidate results as the evidence complementation gain; determining the ratio of the second number to the number of other anomaly candidate results as the consistency ratio; and determining the degree of evidence complementarity based on the evidence complementation gain and the consistency ratio.

[0009] In one possible design, determining the transmission necessity includes: determining the segment anomaly category corresponding to the collaborative inspection segment based on the anomaly category corresponding to each anomaly candidate result in the collaborative inspection segment; determining the anomaly urgency of the collaborative inspection segment based on the segment anomaly category and the pre-defined correspondence between anomaly category and urgency; and determining the transmission necessity based on the coverage association status, the degree of evidence complementarity, and the anomaly urgency.

[0010] In one possible design, the data upload strategy is determined based on the necessity of transmission, including: when the necessity of transmission is greater than or equal to a first threshold, determining the combination of uploaded content included in the data upload strategy as the representative result and the abnormal candidate result with newly added evidence; when the necessity of transmission is less than the first threshold but greater than a second threshold, determining the combination of uploaded content included in the data upload strategy as the representative result, and retaining other abnormal candidate results to be uploaded after cloud-side request, wherein the second threshold is less than the first threshold; when the necessity of transmission is less than or equal to the second threshold, determining the abnormal category and abnormal location of the combination of uploaded content included in the data upload strategy as the representative result.

[0011] In one possible design, in response to the handling requirements fed back from the cloud side, a handling task instruction is issued to the handling robot, including: receiving the handling requirements fed back from the cloud side after reviewing the uploaded data, the handling requirements including the handling action category and the handling inspection unit identifier; matching the handling robot according to the handling requirements, and issuing the handling task instruction to the matched handling robot, the handling task instruction including the handling action category and the handling inspection unit identifier.

[0012] In one possible design, matching a disposal robot according to the disposal requirements includes: determining candidate disposal robots based on the disposal action type and robot type in the disposal requirements; and determining the disposal robot from the candidate disposal robots based on the current position, remaining battery power, and current task load of each candidate disposal robot.

[0013] In one possible design, the above method further includes: receiving post-disposal images collected by the disposal robot after the disposal robot executes the disposal task instruction and returns disposal completion information; comparing the post-disposal images with the abnormal candidate results corresponding to the disposal requirements to confirm the disposal results; adding the image data and annotation information corresponding to the confirmed disposal cases to the edge-side incremental optimization dataset, and using the edge-side incremental optimization dataset to periodically perform incremental training and updates on the edge-side lightweight model deployed on the inspection robot.

[0014] The present invention has the following beneficial effects: In the multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making provided by this invention, the target area is pre-divided into multiple inspection units and the path connection relationship between the inspection units is defined. This allows anomaly candidate results scattered across different robots to be grouped into collaborative inspection segments according to spatial and temporal conditions, rather than being treated as isolated events. This establishes a structural association between multi-robot observations at the edge stage. Based on this, the spatial collaborative coverage of the same anomaly object by multiple robots is measured by the coverage association state, and the scale of non-redundant evidence among anomaly candidate results is measured by the degree of evidence complementarity. This distinguishes between redundant and necessary complementary observations in multi-robot collaboration, and determines the transmission necessity based on both the coverage association state and the degree of evidence complementarity. This ensures that the upload decision reflects both the structural completeness of the collaborative coverage and the complementary value of the evidence information. Furthermore, a data upload strategy is determined based on the hierarchy of transmission necessity. Different combinations of uploaded content correspond to different transmission necessity intervals. This ensures that communication bandwidth and cloud computing resources prioritize anomalies with sufficient collaborative coverage, significant complementary evidence, and high urgency for handling. This avoids the occupation of link and computing resources by invalid and repetitive data transmission, while ensuring that anomaly candidate results with multi-view and multi-sensor complementary value can be completely delivered to the cloud for verification. Upon receiving the handling request from the cloud, a handling task instruction is issued to the handling robot, forming a complete closed-loop link between edge screening, cloud verification, and handling execution. This improves the overall data transmission efficiency and anomaly handling response capability of the multi-robot inspection system in the logistics park. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a multi-robot intelligent inspection method that combines edge-side perception and cloud-side decision-making, as provided in one embodiment of the present invention. Detailed Implementation

[0016] The following description, in conjunction with the accompanying drawings, details a specific scheme for a multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making provided by the present invention.

[0017] Please see Figure 1 The diagram illustrates a flowchart of a multi-robot intelligent inspection method combining edge perception and cloud-side decision-making according to an embodiment of the present invention, including the following steps S101-S105.

[0018] S101. Based on the geospatial structure of the target area, pre-divide into multiple inspection units and define the path connection relationship between each inspection unit.

[0019] In this embodiment of the invention, the target area can be a logistics park. As the core scenario for cargo distribution and continuous equipment operation, the logistics park's internal geographical structure includes various functional areas, such as road access, loading and unloading platforms, cargo yards, conveyor belt sections, and equipment areas. Each functional area has clearly defined physical boundaries, such as the isolation zone between the platform and the yard, the fence between the conveyor belt and the equipment area, and the corner structure between the access road and the yard. These physical boundaries provide a natural basis for the division of inspection units.

[0020] As one possible approach, the target area can be divided into multiple inspection units based on the spatial boundaries of each functional area within the target area.

[0021] Each inspection unit covers a functional area or a portion of a functional area. For functional areas with a large spatial span, such as long conveyor belt sections or large cargo yards, they can be further subdivided into multiple local inspection units based on their spatial distribution characteristics. For functional areas with a relatively compact spatial range, such as independent equipment areas or single loading and unloading platforms, they are considered as a complete inspection unit. Due to structural obstructions in areas such as cargo yards and loading and unloading platforms within logistics parks, abnormal objects within the same inspection unit may present different visible information when viewed from different path directions.

[0022] After the inspection units are divided, the path connection relationships between the inspection units are defined based on their spatial location within the target area and the direction of the inspection path. Optionally, adjacent inspection units that have a continuous observation relationship on the inspection path are defined as inspection units with a path connection relationship.

[0023] In this context, continuous observation relationship refers to the ability of a robot to observe different visible surfaces of the same abnormal object as it moves along an inspection path from one inspection unit to another. This means the two inspection units are physically connected, and their respective observation perspectives complement each other for the same abnormal object. For example, there is a path continuity relationship between an inspection unit located in the outer aisle of a cargo yard and an inspection unit located on the back platform of the cargo yard. An abnormality at the edge of the yard observed by the robot on the aisle side can be observed again from a different angle after the robot on the platform side reaches the back inspection unit, thus forming multi-faceted coverage of the same abnormal object. Similarly, there is a path continuity relationship between an inspection unit above a conveyor belt and an inspection unit below the side of the conveyor belt. The robot in the upper inspection unit can observe abnormalities on the conveyor belt surface, while the robot in the lower side inspection unit can obtain visual information and thermal imaging results of the conveyor belt support structure. Both cover different observation surfaces and different sensor information of the same conveyor belt section, thus forming a continuous inspection unit relationship.

[0024] In one example, taking a logistics park covering approximately 120,000 square meters as an example, the logistics park is pre-divided into several inspection units according to the above rules. Ten inspection robots and five handling robots are deployed within the logistics park, including cleaning robots, handling robots, and verification robots. The logistics park is fully covered by a 5G mobile communication network, and edge nodes are deployed with inference servers with pre-set computing power to quickly aggregate and forward the results of lightweight model inference on the edge side. After receiving an inspection task, each inspection robot moves autonomously along a pre-set path. The inspection task includes a task identifier, a list of robot identifiers, a set of inspection path coordinates, a key anomaly category, and an execution time period. The key anomaly categories include garbage accumulation, personnel intrusion, foreign objects on equipment surfaces, liquid leaks, and equipment overheating.

[0025] S102. Obtain the abnormal candidate results collected by multiple inspection robots performing inspection tasks along preset paths. Based on the inspection unit pointed to by each abnormal candidate result and the path connection relationship, construct the abnormal candidate results that meet the preset spatiotemporal merging conditions into a collaborative inspection segment.

[0026] In this embodiment of the invention, each inspection robot performs inspection tasks along a preset path and simultaneously collects multi-source sensing data.

[0027] The multi-source perception data includes real-time video streams from visual cameras, temperature distribution images from infrared thermal imagers, obstacle data from lidar, and position and orientation information from the integrated navigation module. The visual camera is used to identify visible anomalies such as garbage accumulation, foreign objects, and intruders; the infrared thermal imager is used to acquire temperature distribution images of equipment surfaces or conveyor belt areas to determine if the equipment is overheating or experiencing localized thermal anomalies; the lidar is used to acquire information about obstacles ahead and support obstacle avoidance; and the integrated navigation module is used to record the inspection robot's current position, orientation, and time information.

[0028] Each inspection robot is equipped with a lightweight edge model that performs real-time anomaly screening of real-time video streams and thermal imaging results from multi-source sensing data, outputting anomaly candidate results. The anomaly candidate results include anomaly category, confidence level, acquisition timestamp, robot identifier, and anomaly location.

[0029] It should be noted that the edge-side lightweight model can be pre-trained using an anomaly recognition neural network from existing technologies. The model's input consists of real-time video streams and thermal imaging results. The neural network identifies candidate anomalies through processing. Specifically, when detecting real-time video streams, the edge-side lightweight model identifies visible anomalies such as garbage accumulation, personnel intrusion, or foreign objects on equipment surfaces, and outputs corresponding anomaly boxes and anomaly categories upon detection. When detecting thermal imaging results, the edge-side lightweight model identifies areas with abnormal temperatures, and outputs corresponding anomaly boxes and anomaly categories upon detection. The edge-side lightweight model also provides a confidence score for each output anomaly candidate, such as the probability of the output result in a fully connected layer of the neural network, to characterize the reliability of the edge-side lightweight model's judgment of the anomaly candidate. The acquisition timestamp and robot identifier are information inherent in the frame image. Anomaly localization can be obtained by coordinate transformation based on the robot's positioning information and the position of the anomaly box in the image when the anomaly candidate is acquired. The specific training and deployment methods of the model are well-known techniques to those skilled in the art and will not be elaborated upon here.

[0030] Among them, the anomaly category is used to characterize the anomaly category to which the anomaly candidate result belongs. The anomaly categories include garbage accumulation, personnel intrusion, foreign objects on the equipment surface, liquid leakage, and equipment overheating; the confidence level is used to characterize the reliability of the end-side lightweight model's judgment of the anomaly candidate result; the collection timestamp is used to characterize the precise time when the anomaly candidate result was collected; the robot identifier is used to uniquely identify the inspection robot that generated the anomaly candidate result; and the anomaly location is used to characterize the geographical coordinate information corresponding to the anomaly candidate result.

[0031] In some embodiments, when each inspection robot outputs an anomaly candidate result, it also simultaneously records the robot's device orientation angle parameter when the anomaly candidate result is generated. The device orientation angle parameter is used to characterize the body orientation of the inspection robot when collecting the anomaly candidate result.

[0032] In some embodiments, anomalous candidate results with a confidence level higher than a preset baseline threshold (e.g., an empirical value of 0.6) are standardized into anomalous data. Anomalous data includes packet identifiers, robot identifiers, anomalous location coordinates, acquisition timestamps, anomalous information, image information, and sensor information. The anomalous information includes the anomalous category, confidence level, anomalous bounding box coordinates, and severity; the image information includes the image format, compression quality factor, data size, and encoded data; and the sensor information includes the visual resolution and the highest thermal imaging temperature. The anomalous region image is cropped from the original image based on the anomalous bounding box coordinates and then compressed using a preset compression quality factor, significantly reducing the data size of a single anomalous image.

[0033] After collecting and standardizing the candidate anomalies, the anomaly location of each candidate anomaly is determined by judging whether the anomaly location falls within the spatial range of a certain inspection unit, thus identifying the inspection unit to which the candidate anomaly points. Then, each candidate anomaly is mapped to its corresponding inspection unit, and the identifier of the inspection unit corresponding to each candidate anomaly is determined.

[0034] After mapping abnormal candidate results to inspection units, abnormal candidate results that meet the preset spatiotemporal merging conditions are constructed into collaborative inspection segments.

[0035] In some embodiments, the preset spatiotemporal merging condition can be understood as jointly filtering each abnormal candidate result according to spatial and temporal conditions. Specifically, an abnormal candidate result whose mapped inspection units are all target inspection units or inspection units with a path connection to the target inspection unit, and whose collection time interval is less than a preset time threshold, is considered to satisfy the preset spatiotemporal merging condition. The target inspection unit is the inspection unit pointed to by any abnormal candidate result.

[0036] Optionally, the preset time threshold value depends on the area of ​​each inspection unit and the moving speed of the inspection robot, to ensure that inspection robots with path connections can reach adjacent observation positions of the same abnormal object successively within the time window. For example, when the area of ​​the inspection unit is 2000 square meters and the moving speed of the inspection robot is 1.5 meters per second, the preset time threshold can be set to 60 seconds.

[0037] In one example, assuming any two candidate abnormal results are denoted as the first candidate abnormal result and the second candidate abnormal result, their corresponding collection timestamps are the first timestamp and the second timestamp, respectively, and the mapped inspection units are the first inspection unit and the second inspection unit, respectively. When the first inspection unit and the second inspection unit are the same inspection unit, or the first inspection unit and the second inspection unit have a path connection relationship, and the time interval between the first timestamp and the second timestamp is less than a preset time threshold, the first candidate abnormal result and the second candidate abnormal result meet the preset spatiotemporal merging condition and are merged into the same collaborative inspection segment.

[0038] In another example, when there are any two or more abnormal candidate results that map to the target inspection unit or inspection units that have a path connection with the target inspection unit, the median of the collection timestamps of multiple abnormal candidate results that meet the aforementioned spatial conditions can be taken as the reference timestamp for the collection time interval. The time interval between the collection timestamp of each abnormal candidate result and the reference timestamp can be determined, and then the abnormal candidate results with a time interval less than the preset time threshold can be determined as meeting the preset spatiotemporal merging conditions and included in the same collaborative inspection segment.

[0039] In some embodiments, after constructing a collaborative inspection segment, the anomaly categories of each anomaly candidate result contained in the collaborative inspection segment are further compared. If different anomaly categories are contained, the anomaly candidate results corresponding to different anomaly categories are determined to be mutually independent anomaly events. Based on this determination, the collaborative inspection segment constructed above is divided according to each anomaly category to obtain multiple collaborative inspection segments with different anomaly categories for subsequent processing.

[0040] In some embodiments, if there are no abnormal candidate results that meet the preset spatiotemporal merging conditions, collaborative inspection segments cannot be constructed, and the abnormal candidate results are directly uploaded to the cloud side independently.

[0041] It should be noted that the logic behind setting the pre-defined spatiotemporal merging conditions is as follows: In logistics parks, the same anomalous object is often not fully observed by a single robot. For example, a garbage pile at the edge of a storage yard might be partially seen by a robot on the aisle side first, while a robot on the platform side might only see the other side, obscured by cargo, at a later time. Similarly, a liquid leak under a conveyor belt might only appear as ground reflection in ordinary video, while thermal imaging is needed to further determine if it is accompanied by equipment overheating. Therefore, only by grouping temporally similar and spatially connected anomaly candidates into the same collaborative inspection segment can a complete data foundation be provided for subsequent coverage correlation analysis and evidence complementarity analysis.

[0042] S103. Determine the necessity of transmission based on the coverage association status and the degree of evidence complementarity of the collaborative inspection segments.

[0043] Among them, the coverage association state is used to characterize the degree of spatial collaborative coverage of the same abnormal object by multiple robots, and the degree of evidence complementarity is used to characterize the scale of non-redundant evidence among the candidate results of each abnormality.

[0044] As one possible approach, the coverage association status of the collaborative inspection segments is first determined.

[0045] It should be noted that in multi-robot inspection scenarios in logistics parks, the same abnormal object is often observed by robots on different paths from different angles and with different visibility. Relying solely on isolated observations by a single robot makes it difficult to determine whether the abnormal object has been adequately covered. Therefore, a quantitative calculation method for coverage association is needed. This method assesses whether collaborative coverage of the same abnormal object has been achieved by statistically analyzing the continuity of observations by multiple robots within a collaborative inspection segment at the inspection unit level and the complementarity at the visibility level. Only when multiple robots around the same inspection unit have completed observations of the same abnormal object from different visibility perspectives does that abnormal object possess high collaborative analysis value.

[0046] In some embodiments, the process for determining the associated status is as follows.

[0047] From the collaborative inspection segments, the candidate result pairs of anomalies that are identified as the same anomaly object through anomaly category comparison and anomaly location comparison are taken as valid observation pairs, and the path acceptance ratio is determined based on the number of valid observation pairs.

[0048] The determination of the same anomalous object through anomaly category comparison and anomaly location comparison involves comparing the consistency of the anomaly categories of two candidate anomalies and comparing the spatial proximity of their anomaly locations. If the anomaly categories are consistent and the distance between the anomaly locations is less than a preset threshold for determining the same object (e.g., an empirical value of 10 meters), then the two candidate anomalies are determined to point to the same anomalous object. For example, in a cargo yard inspection scenario, a channel-side inspection robot collects an anomaly candidate with the anomaly category of garbage accumulation and its anomaly location at the outer edge of the yard. At approximately the same time, a platform-side inspection robot collects another anomaly candidate with the same anomaly category of garbage accumulation and its anomaly location at the back of the yard near the loading / unloading area. After comparison, the anomaly categories of the two are consistent, and the distance between their anomaly locations is less than the preset threshold for determining the same object; therefore, the two candidate anomalies are determined to point to the same garbage anomaly object in the yard. This determination allows for the association of observations of the same anomalous object from different robots and perspectives, avoiding the erroneous merging of observations of different anomalies or the omission of complementary observations of the same anomaly.

[0049] The ratio of the number of effective observation pairs to the number of all possible observation pairs is then determined as the path connection ratio. This path connection ratio is used to measure the proportion of observation pairs that constitute effective connection relationships within a collaborative inspection segment among all possible observation pairs.

[0050] Subsequently, the visible surface categories of the abnormal object covered by each candidate anomaly result are labeled, and the number of different visible surface categories is counted as the number of filler surfaces. Visible surface categories include front side, side, back filler surfaces, and thermal filler surfaces. Specifically, the front side refers to the surface covered when the robot observes the abnormal object from directly in front; the side side refers to the surface covered when the robot observes the abnormal object from the side; the back filler surface refers to the surface covered when the robot observes the abnormal object from the back; and the thermal filler surface refers to the observation surface corresponding to the temperature distribution of the abnormal object obtained through a thermal imaging sensor. Then, based on the number of filler surfaces and the maximum allowed number of visible surfaces under the same inspection unit to which the abnormal object belongs, the filler coverage rate is determined. The filler coverage rate is used to measure the completeness of coverage of different visible surfaces of the abnormal object by each observation within a collaborative inspection segment.

[0051] It should be noted that before labeling the visible object categories covered by each anomaly candidate result, it is necessary to calculate and determine the visible object categories covered by each anomaly candidate result based on the robot pose information and anomaly localization information corresponding to each anomaly candidate result. Optionally, the robot position coordinates and device orientation angle parameters synchronously recorded by the integrated navigation module on each inspection robot when collecting each anomaly candidate result are invoked. The robot position coordinates and device orientation angle parameters are then used to perform relative orientation spatial geometric calculations with the anomaly localization coordinates in the corresponding anomaly candidate result. By calculating the angle between the direction vector pointing from the robot's current position to the anomaly localization coordinates and the robot's device orientation angle, the observation geometric angle between the robot body pointing and the anomaly object is obtained. According to the preset angle range to which this observation geometric angle belongs, it is converted into the corresponding observation viewpoint label and assigned to the corresponding visible object category. Specifically, when the observed geometric angle is within a first preset angle range, the abnormal candidate result is marked as covering the front side; when the observed geometric angle is within a second preset angle range, the abnormal candidate result is marked as covering the lateral surface; and when the observed geometric angle is within a third preset angle range, the abnormal candidate result is marked as covering the back replacement surface. If the abnormal candidate result originates from thermal imaging data acquired by an infrared thermal imager, regardless of the visible surface category calculated above, the abnormal candidate result is additionally marked as covering the thermal replacement surface, meaning that the abnormal candidate result simultaneously covers both the calculated visible surface category and the thermal replacement surface.

[0052] The first preset angle range is from -45 degrees to 45 degrees, corresponding to the robot's observation from directly in front of the abnormal object. Within this range, the observed geometric angle indicates that the robot's orientation is roughly opposite to the abnormal object, and the candidate abnormality is marked as covering the front side. The second preset angle range is from 45 degrees to 135 degrees and from -135 degrees to -45 degrees, corresponding to the robot's observation from the side of the abnormal object. Within this range, the observed geometric angle indicates that the robot's orientation is roughly perpendicular to the abnormal object, and the candidate abnormality is marked as covering the side surface. The third preset angle range is from 135 degrees to 180 degrees and from -180 degrees to -135 degrees, corresponding to the robot's observation from the back of the abnormal object. Within this range, the observed geometric angle indicates that the robot's orientation is roughly in the same direction as or away from the abnormal object, and the candidate abnormality is marked as covering the back side. The above preset angle ranges are defined with the robot's orientation as the reference zero-degree direction; positive angles indicate clockwise deviation, and negative angles indicate counterclockwise deviation. The boundary values ​​of each preset included angle range can be adjusted according to the actual direction of the inspection path within the target area and the definition of the observation surface of the inspection unit. This embodiment of the invention does not impose specific limitations on this.

[0053] To more clearly describe the quantitative calculation method of coverage association state in the embodiments, a calculation formula for coverage association state is constructed. Since coverage association state needs to simultaneously reflect the basis of the number of observations and the degree of supplementary coverage of visible surfaces, and as the supplementary coverage rate increases, the coverage association state should exhibit a gradual transition from local observations to effective collaborative coverage, a calculation structure can be adopted that uses the unit basis quantity as a product factor and constructs an adjustment factor using the supplementary coverage rate through an exponential decay function. For example, the calculation formula for coverage association state is as follows: In the formula, Indicates the coverage association status of collaborative inspection segments. A larger value indicates a higher degree of spatial collaborative coverage of the same abnormal object by multiple robots; Indicates the amount of foundation support for the unit. The result can be obtained by multiplying the total number of abnormal candidate results contained in the collaborative inspection segment and the path acceptance ratio, and then normalizing the result. For example, the normalization process can be performed by dividing the product value by the maximum allowed number of abnormal candidate results in the collaborative inspection segment. The maximum number of abnormal candidate results can be the total number of inspection robots in the target area. This is used to reflect whether the observations of multiple robots form a basis around the same anomalous object; This indicates the coverage rate, used to reflect the completeness of coverage of different visible surfaces; This represents an exponential function with the natural constant as its base. The meaning of this formula is that when... At lower levels, even if multiple robots observe the same anomalous object, they are merely repeating the observation of the same visible surface. Approaching a smaller value, covering associated states It remains at a low level; with Increased size leads to improved coverage and filling of gaps. Approaching 1, covering associated states Gradually approaching This indicates that collaborative coverage across different paths and different visible surfaces has been achieved.

[0054] After determining the coverage association status, determine the degree of evidence complementarity of the collaborative inspection segments.

[0055] It should be noted that in collaborative inspection segments, the establishment of coverage association does not mean that all anomaly candidate results are worth uploading. If multiple anomaly candidate results only provide duplicate observation information, their contribution to cloud-side verification is limited. Therefore, it is necessary to construct a quantitative calculation method for the degree of evidence complementarity. By selecting representative results and comparing the amount of new evidence and the consistency of anomaly pointing to other anomaly candidate results relative to the representative results, the scale of non-redundant evidence within the collaborative inspection segment can be determined.

[0056] In some embodiments, the process for determining the degree of complementarity of evidence is as follows.

[0057] The anomaly candidate results with the highest confidence level and image quality meeting preset conditions in the collaborative inspection segments are selected as representative results. Meeting the preset image quality conditions can mean that the size of the anomaly region in the image data corresponding to the anomaly candidate result is not less than a preset minimum pixel size; for example, the width and height of the anomaly box must both be no less than 32 pixels to avoid insufficient image information to support fine-grained cloud-side verification due to anomaly regions being too small. Alternatively, meeting the preset image quality conditions can also mean that the sharpness index of the image data corresponding to the anomaly candidate result is not lower than a preset sharpness threshold. The sharpness index can be the Laplacian variance of the image, which measures the richness of edge and texture information in an image; a higher Laplacian variance indicates a sharper image. The preset sharpness threshold can be pre-calibrated based on the image acquisition equipment parameters of the inspection robot and the lighting environment of the park.

[0058] For any anomalous candidate results within a collaborative inspection segment, excluding the representative result, a first number of anomalous candidate results that provide additional evidence relative to the representative result is determined. This additional evidence includes new visible surface categories, thermal imaging information, or local details. The ratio of this first number to the number of other anomalous candidate results is defined as the supplementary evidence gain. The supplementary evidence gain measures the richness of the additional evidence provided by the other anomalous candidate results relative to the representative result.

[0059] A second number of outlier candidate results is determined, whose outlier category and location are consistent with the outlier object pointed to by the representative result. The ratio of this second number to the number of other outlier candidate results is determined as the consistency ratio. The consistency ratio is used to measure the degree of agreement between other outlier candidate results and the representative result in terms of outlier identification. When other outlier candidate results provide new evidence but the outlier object they point to is inconsistent with the representative result, it indicates a conflict, and the credibility of this new evidence should be suppressed.

[0060] To more clearly describe the quantitative calculation method of the degree of evidence complementarity in the embodiments, a formula for calculating the degree of evidence complementarity is constructed. Since the degree of evidence complementarity needs to simultaneously reflect the richness of newly added evidence and the consistency of the anomaly indication, and the value of newly added evidence should be suppressed when the consistency is low, the formula for calculating the degree of evidence complementarity is as follows: In the formula, This indicates the degree of complementarity of evidence from collaborative inspection segments. The larger the value, the greater the scale of non-redundant evidence among the anomalous candidate results; This indicates the supplementary evidence gain, which reflects the richness of new evidence provided by other anomalous candidate results relative to the representative result; This represents the consistency ratio, used to reflect the degree of consistency between other anomalous candidate results and the representative result in terms of anomalous indication. The calculation means that when... A higher value indicates that although the supplementary results come from different robots or different sensors, they still consistently point to the same abnormal object. Mainly composed of Decision; when A lower value indicates a significant conflict between the results, specifically in the denominator. The item increases, It is suppressed, thereby inhibiting the complementary identification of conflicting information. Furthermore, when At that time, the degree of complementarity of evidence is set to zero and is not included in the calculation of the above formula.

[0061] Finally, after determining the coverage association status and the degree of evidence complementarity, the necessity of transmission is determined.

[0062] The transmission necessity metric is used to comprehensively measure the urgency of uploading candidate anomaly results from collaborative inspection segments to the cloud. Different anomaly categories have varying degrees of urgency. Personnel intrusion, liquid leaks, and equipment overheating often require faster triggering of cloud-based response tasks, while in some scenarios, ordinary garbage accumulation or foreign objects on equipment surfaces can be handled by uploading representative results first, with other results supplemented later. Therefore, the transmission necessity metric also needs to be considered in conjunction with the anomaly urgency.

[0063] In some embodiments, the segment anomaly category corresponding to a collaborative inspection segment can be determined based on the anomaly category corresponding to each anomaly candidate result in the collaborative inspection segment. Optionally, the anomaly category of the anomaly candidate result with the highest confidence among all anomaly candidate results in the collaborative inspection segment can be determined as the segment anomaly category of the collaborative inspection segment, or the anomaly category with the highest frequency among all anomaly candidate results in the collaborative inspection segment can be determined as the segment anomaly category of the collaborative inspection segment. Then, based on the segment anomaly category and combined with a preset correspondence between anomaly category and urgency, a dimensionless value is mapped to obtain a value, which is the anomaly urgency of the collaborative inspection segment. The anomaly urgency is used to characterize the urgency of handling the anomaly category in the target area, and the value of the anomaly urgency ranges from 0 to 1.

[0064] To more clearly describe the quantitative calculation method of transmission necessity in the embodiments, a formula for calculating transmission necessity is constructed. Transmission necessity needs to rapidly approach a high value range as the coverage correlation state, the degree of evidence complementarity, and the urgency of anomalies all increase together. Furthermore, transmission necessity should remain at a low level when any one of the three factors is low. For example, the formula for calculating transmission necessity is as follows: In the formula, Indicates the necessity of transmitting collaborative inspection segments. The larger the value, the more necessary it is to upload the collaborative inspection segment to the cloud. Indicates the overriding associated status; Indicate the degree of complementarity of evidence; Indicates an exceptionally high level of urgency; This represents an exponential function with the natural constant as its base.

[0065] in, This indicates the combined effect of structural collaboration value, added evidence value, and park risk requirements in the same collaborative inspection segment, outer layer This causes the transmission necessity to rapidly approach a high value range as all three factors increase together; conversely, if any one of the three factors is low, the product remains at a low level, and the transmission necessity is also limited to a low level. (This is similar to coverage-related states.) A low value indicates that multiple robots have not yet formed effective collaborative coverage around the same anomalous object. In this case, even if the degree of evidence complementarity and the urgency of the anomaly are high, the necessity of transmission is limited to a low level to avoid uploading multiple sets of anomaly candidate results before a spatial collaborative coverage structure has been formed. For example, the degree of evidence complementarity... A low value indicates that other anomalous candidate results provide limited new evidence compared to the representative result, and are mostly redundant information. The necessity of transmission is limited to a low level to avoid repeatedly uploading similar anomalous candidate results when the evidence does not provide effective complementarity; such as the urgency of the anomalous response. A lower value indicates that the timeliness requirement for handling this anomaly category is not high, and the necessity of transmission is limited to a low level, avoiding the uploading of all data content in the collaborative inspection segment when the business risk is low.

[0066] S104. Determine the data upload strategy based on the transmission necessity, and send the abnormal data corresponding to the collaborative inspection segment to the cloud side based on the data upload strategy.

[0067] The data upload strategy includes a combination of uploaded content.

[0068] It should be noted that the transmission necessity determined in step S103 above comprehensively reflects the overall necessity of uploading collaborative inspection segments across three dimensions: spatial collaborative coverage, complementary evidentiary value, and business urgency. In actual logistics park inspection scenarios, communication bandwidth and cloud computing resources are limited. Uploading all abnormal candidate results from all collaborative inspection segments indiscriminately would result in redundant load on the transmission link and ineffective consumption of cloud computing power. Uploading only a single representative result might lead to the loss of supplementary information obtained by other robots from different perspectives or sensors, resulting in incomplete information during cloud-side verification. Therefore, it is necessary to determine corresponding data upload strategies based on the level of transmission necessity, ensuring that the uploaded data content matches the upload necessity of the collaborative inspection segments, thereby reducing invalid data transmission while guaranteeing sufficient verification information for the cloud side.

[0069] One possible implementation involves comparing the transmission necessity with a preset first threshold and a second threshold, and determining the combination of upload content in the data upload strategy based on the comparison result. The first threshold is used to distinguish between high and medium upload necessity, and the second threshold is used to distinguish between medium and low upload necessity, with the second threshold being lower than the first threshold.

[0070] It should be noted that the theoretical range of the transmission necessity determined in step S103 above is 0 to... Between these parameters, the transmission necessity reaches its maximum value when the coverage association status, the degree of evidence complementarity, and the anomaly urgency all reach their theoretical maximum value of 1. The value is approximately 0.632. In practical applications, the specific values ​​of the first and second thresholds can be set according to the requirements of data transmission efficiency and processing resources in the actual application scenario. However, the first threshold should be less than the theoretical maximum value. For example, the first threshold can be taken as an empirical value of 0.5, and the second threshold can be taken as an empirical value of 0.2.

[0071] When the transmission necessity is greater than or equal to the first threshold, it indicates that the collaborative inspection segment is at a high level in terms of spatial collaborative coverage, evidence complementarity, and anomaly urgency. Multiple robots have formed effective collaborative coverage around the same anomaly object, and other anomaly candidate results provide significant new evidence relative to the representative result. Simultaneously, the anomaly category has a high urgency for handling. At this point, the data upload strategy is determined to include a combination of the representative result and anomaly candidate results with new evidence. Anomaly candidate results with new evidence refer to those that, when determining the degree of evidence complementarity, are judged to provide new visible surface categories, thermal imaging information, or local details relative to the representative result. This representative result and anomaly candidate results with new evidence are sent to the cloud as the anomaly data corresponding to the collaborative inspection segment, enabling the cloud to obtain complementary information from multiple perspectives and sensors for a complete review of the anomaly object. This allows the cloud to receive the representative result as early as possible and initiate a refined review process. Simultaneously, cross-validation is performed upon receiving supplementary perspective or thermal imaging data, ensuring the completeness of the review while shortening the overall response time.

[0072] When the transmission necessity is less than the first threshold but greater than the second threshold, it indicates that although the collaborative inspection segment has achieved a certain level of collaborative coverage, the amount of new evidence from other anomaly candidate results relative to the representative result has not yet reached a high level, or the anomaly urgency is moderate. In this case, the combination of uploaded content included in the data upload strategy is determined as the representative result, and this representative result is sent to the cloud as the anomaly data corresponding to the collaborative inspection segment. For other anomaly candidate results within the collaborative inspection segment, their data content is retained and not uploaded temporarily. The cloud side waits for the representative result to be reviewed. If it determines that supplementary viewpoint or thermal imaging information is needed, a request is sent to the endpoint. Upon receiving the request, the endpoint uploads the retained anomaly candidate results. This ensures that the cloud side can obtain basic anomaly identification information while reducing unnecessary supplementary data transmission and lowering the load on the 5G communication link and cloud computing power.

[0073] When the transmission necessity is less than or equal to the second threshold, it indicates that the collaborative inspection segment has low collaborative coverage, low evidence complementarity, and low anomaly urgency. Multiple robots may have only repeatedly observed the same visible surface, or the anomaly category does not have high timeliness requirements for handling. In this case, the uploaded content combination included in the data upload strategy is determined as the anomaly category and anomaly location representing the result, and sent to the cloud as the anomaly data corresponding to the collaborative inspection segment. This avoids the bandwidth and computing power consumption caused by complete anomaly candidate results for anomalies with insufficient collaborative coverage, insufficient evidence complementarity, and low urgency. The cloud can initially register the anomaly based on its category and location, and then make a comprehensive judgment after obtaining more observation data for that area in subsequent inspections.

[0074] Understandably, the above-mentioned hierarchical upload strategy achieves dynamic matching between the necessity of transmission and the combination of uploaded content. This ensures that data transmission from the device side to the cloud side avoids resource waste caused by repeated and redundant uploads, while also ensuring that abnormal data with complementary value can be delivered to the cloud side for review in a timely and complete manner.

[0075] S105. In response to the processing needs fed back from the cloud side, issue processing task instructions to the processing robot.

[0076] In some embodiments, after receiving abnormal data sent by the end side based on the data upload policy, the cloud side reviews the abnormal data and generates a handling requirement.

[0077] Optionally, the review process includes the cloud side first determining whether the received abnormal data contains image data items. If the abnormal data only contains abnormal category and abnormal location information but not image data items due to the threshold grading strategy, the cloud side skips the image integrity verification step and the large model fine recognition step, and directly registers and records the abnormal data based on the abnormal category and abnormal location information, generating a corresponding handling requirement. The handling action category in this handling requirement is "review," and the handling inspection unit identifier is the inspection unit identifier pointed to by the abnormal location information. If the abnormal data contains image data items, the image data integrity is first verified to confirm that the task identifier, robot identifier, inspection unit identifier, collection timestamp, abnormal category, and abnormal box fields in the image data are complete and valid. After the verification is passed, the cloud side uses the large model deployed on the cloud to perform fine recognition of the image data, and combines the thermal imaging results and abnormal location information to reconfirm and refine the classification of abnormal candidate results. For example, waste accumulation can be further subdivided into small amounts, medium amounts, or large amounts; liquid leaks can be further subdivided into water leaks, oil leaks, or chemical leaks; and equipment overheating can be further subdivided into localized overheating or overall overheating. The cloud-based system then generates structured disposal requirements based on these detailed classifications.

[0078] For example, a handling requirement includes a handling action category and a handling inspection unit identifier. The handling action category indicates the specific handling operation to be performed, including at least one of cleaning, handling, verification, and maintenance. The handling inspection unit identifier indicates the inspection unit that needs to perform the handling operation. For instance, when the cloud side identifies a liquid leak on the back side of the loading / unloading platform and further classifies it as an oil leak, the generated handling requirement would have handling action categories of cleaning and verification, and the handling inspection unit identifier would be the identifier of the inspection unit where the liquid leak occurred.

[0079] Furthermore, upon receiving a handling request from the cloud side after reviewing the uploaded data, the system matches a handling robot according to the request and issues a handling task instruction to the matched robot. This handling task instruction includes the type of handling action and the identification of the handling inspection unit.

[0080] In some embodiments, when matching disposal robots, candidate disposal robots are first determined based on the disposal action category and robot category in the disposal requirement. Robot categories include cleaning robots, handling robots, and verification robots, etc. Different disposal action categories correspond to different robot categories: cleaning actions correspond to cleaning robots, handling actions correspond to handling robots, and verification actions correspond to verification robots. Robots matching the required robot category are then retrieved from the robot status database as candidate disposal robots.

[0081] After identifying candidate robots, the final robot to perform the task needs to be selected. Since multiple candidate robots meeting the category requirements may exist simultaneously in a logistics park, and each robot has different current locations, remaining battery power, and current workloads, it is necessary to select the robot with the most suitable overall conditions. Current location affects the robot's travel time to the target inspection unit; remaining battery power affects whether the robot can complete the task and return; and current workload affects the robot's ability to take on new tasks. Therefore, a robot matching degree calculation formula can be constructed to comprehensively quantify these three factors, allowing the robot with the highest matching degree to be selected as the final robot to perform the task. For example, the robot matching degree calculation formula is as follows: In the formula, This indicates the matching degree of the candidate disposal robots. The larger the value, the more suitable the candidate robot is to undertake the current task. This indicates the current location of the candidate disposal robot and the travel distance between it and the inspection unit corresponding to the disposal inspection unit identifier; This indicates the maximum allowable travel distance for the candidate disposal robot, for example, an empirical value of 800 meters; This indicates the remaining battery percentage of the candidate disposal robot; This indicates the current task load of the candidate disposal robots; This indicates the maximum number of task loads allowed for the candidate disposal robot, for example, an empirical value of 3; , and For the preset weighting coefficients, and These are used to adjust the relative importance of three factors—driving distance, remaining battery power, and task load—in the matching degree calculation. For example, , , The calculation means that the candidate handling robot with the shorter the travel distance, the more remaining battery power, and the lighter the current task load, the higher its matching degree. The higher the score, the higher the priority for selection as a disposal robot to perform disposal tasks.

[0082] Then, after calculating the matching degree of each candidate disposal robot according to the above robot matching degree calculation formula, the candidate disposal robot with the highest matching degree is selected as the disposal robot, and a disposal task instruction is issued to the disposal robot. The disposal task instruction is sent to the disposal robot through a message queue protocol. The disposal task instruction is used to instruct the disposal robot to perform the corresponding disposal operation and return disposal completion information upon completion.

[0083] In some embodiments, while issuing the task instruction, the inspection unit corresponding to the task requirement and its adjacent inspection units with path connections are added to a key revisit list. The key revisit list represents the set of inspection units that require increased review density and attention level in subsequent inspection cycles. The attention level of inspection units belonging to the key revisit list in subsequent inspection paths is adjusted, causing subsequent inspection robots to increase sampling frequency or lower anomaly detection thresholds when passing these units. This is because if a current anomaly requires multiple robots to relay, it indicates that the inspection unit itself has occlusion, corner, or multi-faceted observation requirements in the geographic spatial structure of the target area. Subsequent inspections should increase the review density of these units, rather than only handling a single inspection unit once.

[0084] In some embodiments, after the handling robot executes the handling task instruction and returns handling completion information, the robot receives the handling post-handling image and compares it with the anomaly candidate results corresponding to the handling requirement to confirm the handling result. This handling result is used to characterize whether the anomaly has been effectively cleared or repaired. For anomaly cases where handling completion is confirmed, the corresponding image data (such as the image data with the highest confidence among the anomaly candidate results or the handling post-handling image) and annotation information are extracted and added to the edge-side incremental optimization dataset. The annotation information includes the inspection unit identifier, anomaly category, and sensor category. Then, the edge-side incremental optimization dataset is used to periodically perform incremental training and updates on the edge-side lightweight model deployed on the inspection robot, enabling the edge-side lightweight model to maintain continuous adaptability to common occlusion anomalies, thermal anomalies, and multi-faceted observation anomalies in the target area.

[0085] Understandably, in the multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making provided in this embodiment of the invention, by pre-dividing the target area into multiple inspection units and defining the path connection relationships between the inspection units, abnormal candidate results scattered across different robots can be grouped into collaborative inspection segments according to spatial and temporal conditions, rather than being processed independently as isolated events. This establishes a structural association between multi-robot observations at the edge stage. Based on this, the spatial collaborative coverage of the same abnormal object by multiple robots is measured by the coverage association state, and the scale of non-redundant evidence among each abnormal candidate result is measured by the degree of evidence complementarity. This distinguishes between redundant and necessary complementary observations in multi-robot collaboration, and determines the transmission necessity based on both the coverage association state and the degree of evidence complementarity. This ensures that the uploaded decision can simultaneously reflect the structural completeness of collaborative coverage and the complementary value of evidentiary information. Furthermore, a data upload strategy is determined based on the hierarchy of transmission necessity. Different combinations of uploaded content correspond to different transmission necessity intervals. This ensures that communication bandwidth and cloud computing resources prioritize anomalies with sufficient collaborative coverage, significant complementary evidence, and high urgency for handling. This avoids the occupation of link and computing resources by invalid and repetitive data transmission, while ensuring that anomaly candidate results with multi-view and multi-sensor complementary value can be completely delivered to the cloud for verification. Upon receiving the handling request from the cloud, a handling task instruction is issued to the handling robot, forming a complete closed-loop link between edge screening, cloud verification, and handling execution. This improves the overall data transmission efficiency and anomaly handling response capability of the multi-robot inspection system in the logistics park.

Claims

1. A multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making, characterized in that, The method includes: Based on the geospatial structure of the target area, multiple inspection units are pre-divided, and the path connection relationship between each inspection unit is defined. Obtain abnormal candidate results collected by multiple inspection robots performing inspection tasks along a preset path. Based on the inspection unit pointed to by each abnormal candidate result and the path connection relationship, construct the abnormal candidate results that meet the preset spatiotemporal merging conditions into a collaborative inspection segment. The transmission necessity is determined based on the coverage association status and the degree of evidence complementarity of the collaborative inspection segments; the coverage association status is used to characterize the degree of spatial collaborative coverage of the same abnormal object by multiple robots, and the degree of evidence complementarity is used to characterize the scale of non-redundant evidence among each abnormal candidate result. Based on the transmission necessity, a data upload strategy is determined, and based on the data upload strategy, abnormal data corresponding to the collaborative inspection segment is sent to the cloud side. The data upload strategy includes a combination of upload content. In response to the processing needs reported from the cloud side, processing task instructions are issued to the processing robot.

2. The multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making according to claim 1, characterized in that, Based on the geospatial structure of the target area, multiple inspection units are pre-divided, and the path connection relationships between each inspection unit are defined, including: Based on the spatial boundaries of each functional area in the target area, the target area is divided into multiple inspection units, each inspection unit covering a functional area or a part of a functional area. Adjacent inspection units that have a continuous observation relationship on the inspection path are defined as inspection units with the path succession relationship.

3. The multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making according to claim 1, characterized in that, Obtain anomaly candidate results collected by multiple inspection robots performing inspection tasks along a preset path. Based on the inspection unit pointed to by each anomaly candidate result and the path connection relationship, construct collaborative inspection segments from the anomaly candidate results that meet preset spatiotemporal merging conditions, including: The system obtains the anomaly candidate results output by each inspection robot after initial screening of the collected images through the edge lightweight model during the execution of the inspection task. The anomaly candidate results include anomaly category, confidence level, collection timestamp, robot identifier, and anomaly location. Based on the anomaly location, each anomaly candidate result is mapped to the corresponding inspection unit, and the inspection unit identifier corresponding to each anomaly candidate result is determined. The abnormal candidate results that are mapped to the target inspection unit or the inspection unit that has the path connection relationship with the target inspection unit and whose collection time interval is less than the preset time threshold are determined as abnormal candidate results that meet the preset spatiotemporal merging conditions and are included in the same collaborative inspection segment. The target inspection unit is the inspection unit pointed to by any abnormal candidate result.

4. The multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making according to claim 3, characterized in that, Determining the coverage association status of the collaborative inspection segment includes: From the collaborative inspection segment, the abnormal candidate result pairs that are determined to be the same abnormal object by comparing the abnormal object category and the abnormal location are taken as valid observation pairs, and the path acceptance ratio is determined according to the number of valid observation pairs. The visible surface categories of the abnormal objects covered by each abnormal candidate result are marked, the number of different visible surface categories is counted as the number of replacement surfaces, and the replacement coverage rate is determined based on the number of replacement surfaces and the maximum number of visible surfaces allowed under the inspection unit to which the same abnormal object belongs. The visible surface categories include front side, side surface, back replacement surface and thermal replacement surface. The coverage association status is determined based on the total number of abnormal candidate results contained in the collaborative inspection segment, the path acceptance ratio, and the supplementary coverage rate.

5. The multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making according to claim 4, characterized in that, Determining the degree of complementarity of evidence for the collaborative inspection segments includes: The abnormal candidate results with the highest confidence and image quality that meet the preset conditions in the collaborative inspection segment are taken as the representative results. For other anomalous candidate results in the collaborative inspection segment besides the representative result, a first number of anomalous candidate results with new evidence relative to the representative result is determined. The new evidence includes new visible surface categories, thermal imaging information, or local details. Determine a second number of abnormal candidate results among the other abnormal candidate results, where both the abnormal category and the abnormal location are consistent with the abnormal object pointed to by the representative result; The ratio of the first quantity to the quantity of other abnormal candidate results is determined as the supplementary evidence gain; The ratio of the second quantity to the quantity of the other abnormal candidate results is determined to be a consistent ratio; The degree of complementarity of the evidence is determined based on the supplementary evidence gain and the consistency ratio.

6. The multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making according to claim 3, characterized in that, Determining the necessity of transmission includes: Based on the anomaly category corresponding to each anomaly candidate result in the collaborative inspection segment, determine the segment anomaly category corresponding to the collaborative inspection segment; Based on the segment anomaly category and combined with the preset correspondence between anomaly category and urgency, the anomaly urgency of the collaborative inspection segment is determined. The transmission necessity is determined based on the coverage association status, the degree of evidence complementarity, and the degree of anomaly urgency.

7. The multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making according to claim 5, characterized in that, Determining a data upload strategy based on the aforementioned transmission necessity includes: If the transmission necessity is greater than or equal to the first threshold, the combination of uploaded content included in the data upload strategy is determined as the representative result and the abnormal candidate result with the new evidence. If the transmission necessity is less than the first threshold and greater than the second threshold, the combination of upload content included in the data upload strategy is determined as the representative result, and the other abnormal candidate results are reserved to be uploaded after cloud side request, where the second threshold is less than the first threshold. If the transmission necessity is less than or equal to the second threshold, the combination of uploaded content included in the data upload strategy is determined as the anomaly category and anomaly location of the representative result.

8. The multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making according to claim 1, characterized in that, In response to the handling needs reported from the cloud side, handling task instructions are issued to the handling robot, including: Receive processing requests from the cloud side after reviewing the uploaded data. The processing requests include the type of processing action and the identifier of the processing inspection unit. According to the disposal requirements, a disposal robot is matched, and a disposal task instruction is issued to the matched disposal robot. The disposal task instruction includes the disposal action category and the disposal inspection unit identifier.

9. The multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making according to claim 8, characterized in that, Matching a disposal robot according to the disposal requirements includes: Candidate disposal robots are determined based on the disposal action category and robot category in the disposal requirements; The disposal robot is determined from the candidate disposal robots based on their current location, remaining battery power, and current task load.

10. The multi-robot intelligent inspection method combining edge-side perception and cloud-side decision-making according to claim 8, characterized in that, The method further includes: After the disposal robot executes the disposal task instruction and returns disposal completion information, the disposal robot collects the post-disposal image; The processed image is compared with the abnormal candidate results corresponding to the processing requirements to confirm the processing result; The image data and annotation information corresponding to the confirmed and handled abnormal cases are added to the edge-side incremental optimization dataset, and the edge-side lightweight model deployed on the inspection robot is periodically trained and updated using the edge-side incremental optimization dataset.

Images