An Adaptive Detection and Detection System for Low-Altitude Flight Inspection

Abstract

An embodiment of the present application provides an adaptive detection and detection system for low-altitude flight inspection, aiming to improve the efficiency and safety of UAV inspection operations. First, multiple UAV description information sets, UAV cluster description information, and flight inspection task description information are obtained. These information cover the performance parameters and task requirements of UAVs in different inspection scenarios. Subsequently, this information is loaded into the target adaptive inspection decision model of pre-knowledge learning. By analyzing this information, a target estimated confidence level is generated to evaluate the success probability of the UAV cluster when performing a specific inspection task. When the target estimated confidence level meets the set requirements, the corresponding UAV cluster is intelligently scheduled to perform the inspection task, thereby ensuring the efficient completion of the task. Thus, through intelligent evaluation and dynamic scheduling, the adaptability and flexibility of UAV inspection operations are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and in particular, to an adaptive detection and detection system for low-altitude flight inspection. Background Art

[0002] With the rapid development of unmanned aerial vehicle technology, low-altitude flight inspection has been widely used in many fields such as power line inspection, environmental monitoring, and disaster assessment. However, in practical applications, unmanned aerial vehicle inspection operations face many challenges, such as complex and changeable inspection environments, differences in unmanned aerial vehicle performance, and diversity of task requirements. Traditional unmanned aerial vehicle inspection scheduling methods often rely on manual experience and fixed rules, and it is difficult to dynamically adjust according to the specific requirements of inspection tasks and the real-time status of unmanned aerial vehicles, resulting in low inspection efficiency and increased safety risks.

[0003] To solve the above problems, related technologies have begun to explore data-driven unmanned aerial vehicle inspection scheduling methods. These methods collect and analyze unmanned aerial vehicle performance data, inspection environment information, task requirements, etc., and use advanced algorithms such as machine learning and deep learning to intelligently schedule and optimize unmanned aerial vehicle inspection tasks. However, most of the existing methods focus on the performance evaluation and task allocation of a single unmanned aerial vehicle, ignoring the synergistic effect between unmanned aerial vehicle clusters and the impact of different inspection scenarios on unmanned aerial vehicle performance, resulting in the scheduling results being difficult to achieve the optimal. Summary of the Invention

[0004] To at least overcome the above deficiencies in the prior art, an object of an embodiment of the present application is to provide an adaptive detection and detection system for low-altitude flight inspection.

[0005] According to one aspect of the present application, an adaptive detection method for low-altitude flight inspection is provided, and the method includes:

[0006] Obtain X sets of unmanned aerial vehicle description information, cluster description information, and flight inspection task description information, where the X sets of unmanned aerial vehicle description information include Y knowledge field description vectors of each unmanned aerial vehicle in X unmanned aerial vehicles, the X unmanned aerial vehicles are unmanned aerial vehicles scheduled to a flight inspection task in a target inspection execution scenario of a target low-altitude inspection operation, the Y knowledge field description vectors of each unmanned aerial vehicle include description vectors of each unmanned aerial vehicle in multiple inspection execution scenarios, the multiple inspection execution scenarios are fully authorized inspection execution scenarios in the target low-altitude inspection operation, the cluster description information includes description vectors of each unmanned aerial vehicle cluster in Z unmanned aerial vehicle clusters in the multiple inspection execution scenarios, the flight inspection task description information includes a description vector of the target inspection execution scenario, the Z unmanned aerial vehicle clusters are multiple unmanned aerial vehicle clusters generated by clustering and grouping the X unmanned aerial vehicles according to the current cluster division strategy, and X, Y, and Z are all not less than 2;

[0007] Load the X sets of UAV description information, the cluster description information, and the flight inspection task description information into the target adaptive inspection decision model learned in advance to generate the target estimated confidence generated by the target adaptive inspection decision model. The target estimated confidence characterizes the confidence that the Z UAV clusters can successfully complete the low-altitude flight inspection task in the one-time flight inspection task under the condition that the Z UAV clusters are scheduled as the UAV clusters in the one-time flight inspection task;

[0008] Under the condition that the target estimated confidence meets the set requirements, use the Z UAV clusters as the UAV clusters scheduled in the one-time flight inspection task to obtain the adaptive detection result before the low-altitude flight inspection.

[0009] According to another aspect of the present application, there is provided a detection system, which includes a processor and a machine-readable storage medium. Machine-executable instructions are stored in the machine-readable storage medium, and the machine-executable instructions are loaded and executed by the processor to implement the above-mentioned adaptive detection method for low-altitude flight inspection.

[0010] In the technical solutions provided by some embodiments of the present application, by integrating and analyzing the multi-dimensional description information of UAV individuals, UAV clusters, and flight inspection tasks, the present application embodiment realizes the accurate evaluation of the ability of UAV clusters to perform inspection tasks, and uses the target adaptive inspection decision model learned in advance to generate the target estimated confidence to quantify the possibility that the UAV cluster can successfully complete the low-altitude flight inspection in a specific inspection task. Under the condition that the target estimated confidence meets the set requirements, it is possible to intelligently schedule the appropriate UAV cluster to the inspection task, thus ensuring the efficiency and safety of the inspection operation. That is, by intelligently evaluating the matching degree between the UAV cluster and the inspection task, it can be ensured that the scheduled UAV cluster has the ability to complete the task, avoiding task delays or failures caused by insufficient UAV performance or improper cluster configuration, thus significantly improving the efficiency of the inspection operation. Before scheduling the UAV cluster, the comprehensive evaluation of the cluster's ability by the target adaptive inspection decision model effectively avoids the safety risks that may be caused by the performance defects of UAVs or clusters, enhancing the safety and reliability of the inspection task. In addition, it is possible to intelligently select the most suitable UAV cluster for scheduling according to the specific requirements of the inspection task, realizing the optimal allocation of UAV resources and improving the resource utilization efficiency. By introducing the target adaptive inspection decision model learned in advance, the intelligent evaluation and decision-making of the inspection ability of UAV clusters are realized, the intelligent level of inspection task scheduling is improved, and the possibility of manual intervention and decision-making errors is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the accompanying drawings required to be enabled in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and thus should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be extracted in combination with these drawings.

[0012] Figure 1 It is a schematic flowchart of the adaptive detection method for low-altitude flight inspection provided by the embodiments of the present application;

[0013] Figure 2 It is a schematic block diagram of the architecture of the detection system provided by the embodiments of the present application. Specific embodiments

[0014] The following description is for enabling those of ordinary skill in the art to implement and combine the present application, and this description is provided in a specific application scenario and the environment required by it. For those of ordinary skill in the art, it is obvious that various changes can be made to the disclosed embodiments, and when not departing from the principles and scope of the present application, the general principles defined in the present application can be applied to other embodiments and application scenarios. Therefore, the present application is not limited to the described embodiments, but should be given the broadest scope consistent with the claims.

[0015] Figure 1 It is a schematic flowchart of the adaptive detection method for low-altitude flight inspection provided by an embodiment of the present application. The following will introduce this adaptive detection method for low-altitude flight inspection in detail.

[0016] Step S110, obtain X sets of UAV description information, cluster description information, and flight inspection task description information. The X sets of UAV description information include Y knowledge field description vectors of each UAV in the X UAVs. The X UAVs are the UAVs for a flight inspection task in the target inspection execution scenario of the target low-altitude inspection operation. The Y knowledge field description vectors of each UAV include the description vectors of each UAV in multiple inspection execution scenarios. The multiple inspection execution scenarios are the fully authorized inspection execution scenarios in the target low-altitude inspection operation. The cluster description information includes the description vectors of each UAV cluster in the Z UAV clusters in the multiple inspection execution scenarios. The flight inspection task description information includes the description vector of the target inspection execution scenario. The Z UAV clusters are multiple UAV clusters generated by clustering and grouping the X UAVs according to the current cluster division strategy. X, Y, and Z are all not less than 2.

[0017] In detail, the drone description information set contains detailed description information of X drones participating in a flight inspection mission. Each drone has Y knowledge field description vectors, which cover the performance, status, historical data and other information of the drone in multiple inspection execution scenarios. The inspection execution scenario refers to the specific environment or conditions in which the drone performs the inspection mission. Different inspection scenarios may have different performance requirements for drones. The cluster description information contains the overall performance description of Z drone clusters in multiple inspection execution scenarios. Each drone cluster is composed of multiple drones, which can cooperate with each other when performing tasks. The drone cluster refers to a team composed of multiple drones, which can cooperate with each other and complete the task together when performing tasks. Drone clusters can improve the efficiency and success rate of inspection tasks.

[0018] The inspection task description information includes a specific description of the target inspection execution scenario, such as geographical features, task objectives, time limits, etc., which is crucial for evaluating whether the drone cluster is suitable for performing the task.

[0019] For example, in this embodiment, for a low-altitude inspection operation of a power line, there are multiple inspection execution scenarios in this low-altitude inspection operation of the power line, and all of them have been fully authenticated, such as mountain inspection scenarios, urban suburban inspection scenarios, and cross-river area inspection scenarios.

[0020] Now there is a flight inspection mission about to be executed, and the server needs to obtain relevant description information. Assume that there are X = 5 drones dispatched to this flight inspection mission under the target inspection execution scenario (assuming that this target inspection execution scenario is a mountain inspection scenario) in this target low-altitude inspection operation. For each drone, there are Y = 3 knowledge field description vectors. These 3 knowledge field description vectors contain description vectors of each drone in multiple inspection execution scenarios (mountain inspection scenario, urban suburban inspection scenario, and cross-river area inspection scenario).

[0021] Take one of the drones as an example. In the mountain inspection scenario, one of its knowledge field description vectors may contain information about its flight altitude adaptability. For example, in previous mountain inspections, this drone was able to fly stably at an average flight altitude of 100-300 meters, and the data related to this altitude range will be recorded in this knowledge field description vector. In the suburban inspection scenario, another knowledge field description vector may record its ability to deal with interference from small buildings, such as the probability of accurately avoiding areas with many low buildings when flying over them. In the inspection scenario across river areas, the third knowledge field description vector may contain data on its imaging capabilities when facing interference from water surface reflections.

[0022] Meanwhile, according to the current cluster division strategy, these 5 drones are divided into Z = 3 drone clusters. The server needs to obtain cluster description information, which includes the description vectors of each drone cluster in the 3 drone clusters under multiple inspection execution scenarios. For example, for one of the drone clusters, in the mountain inspection scenario, the description vector of this drone cluster may include the flight speed stability data of the drones in the drone cluster as a whole. If in previous mountain inspections, the average flight speed of the drones in this cluster fluctuates little, then this data will be recorded. In the suburban inspection scenario, it may include information such as the collaborative ability of the drones in the drone cluster to cope with signal interference. In the inspection scenario across river areas, it may include the overall assessment of the adaptability of the drones in the drone cluster to humid environments, etc.

[0023] Finally, the server also needs to obtain the flight inspection task description information, which is mainly the description vector of the target inspection execution scenario (mountain inspection scenario). This description vector may include the geographical features of the task area, such as the altitude range of the mountain area, the mountain range trend, etc. It may also include features related to the task target, such as the location of the power towers that need to be inspected key points, the trend of the power lines, etc. It may also include features related to the task time limit, such as this inspection task needs to be completed between 8 am and 5 pm, etc.

[0024] Step S120, load the X drone description information sets, cluster description information, and flight inspection task description information into the target adaptive inspection decision model of pre-knowledge learning, and generate the target estimated confidence level generated by the target adaptive inspection decision model. The target estimated confidence level represents the confidence level of successfully completing the low-altitude flight inspection task in the one-time flight inspection task corresponding to the Z drone clusters under the condition that the Z drone clusters are used as the drone clusters scheduled to the one-time flight inspection task.

[0025] Specifically, the target adaptive inspection decision model is a pre-trained model that can generate a target estimated confidence level based on the input drone description information sets, cluster description information, and flight inspection task description information to evaluate the success probability of the drone cluster when performing the inspection task. This target adaptive inspection decision model may be constructed based on technologies such as deep learning and machine learning. After inputting the description information of 5 drones, the description information of 3 drone clusters, and the description information of the mountain inspection task, the target adaptive inspection decision model will output a target estimated confidence level, such as 0.85, indicating that the drone cluster has an 85% probability of successfully completing the mountain inspection task.

[0026] For example, in this embodiment, after obtaining all the above information, this information needs to be loaded into the target adaptive inspection decision model for prior knowledge learning. First, encode and represent these 5 sets of UAV description information to generate corresponding 5 UAV description encoding vectors. For example, encode the data in the 3 knowledge field description vectors of each UAV according to certain rules. For the information on flight altitude adaptability in the mountain inspection scenario mentioned earlier, if the altitude range is represented by numbers, it may be encoded into a specific encoding segment. At the same time, encode and represent the cluster description information to generate a corresponding cluster description encoding vector. For example, encode the various ability data of the cluster in different inspection scenarios into a cluster description encoding vector. And encode and represent the flight inspection task description information to generate a corresponding flight inspection task description encoding vector.

[0027] Then, perform a correlation analysis on these 5 UAV description encoding vectors and the flight inspection task description encoding vector respectively. Taking one of the UAVs as an example, assume it is the first UAV, and interact its description encoding vector with the flight inspection task description encoding vector. First, decompose the first UAV description encoding vector into multiple different first feature groups, such as a flight ability related feature group (including the flight altitude adaptability mentioned earlier), a device status related feature group (such as the encoding corresponding to the UAV battery power, device temperature, etc.), and a task history related feature group (the encoding of the performance data in previous similar inspection tasks). At the same time, decompose the flight inspection task description encoding vector into multiple different second feature groups related to task scenario requirements, such as a task area geographical feature group (the encoding of the geographical data of the mountain area), a task objective related feature group (the encoding related to power towers and lines), and a task time limit related feature group (the encoding of the time period from 8 am to 5 pm).

[0028] For each first feature group, based on semantic understanding and logical relationships, conduct preliminary association analysis with all second feature groups one by one. For example, conduct preliminary association analysis between the flight altitude adaptability in the flight ability related feature group and the mountain altitude range in the mission area geographical feature group. If the flight altitude adaptability of this drone can work well within the mountain altitude range, then the association degree will be judged as high, and an association relationship description document will be established to record the feature group pair composed of the first feature group and the second feature group in this association, as well as the basis for the association (such as the adaptability principle of the flight altitude within the mountain altitude range) and the preliminarily judged association degree. Conduct in-depth interaction operations on the feature group pairs with a high association degree to generate encoded segments after in-depth interaction. For example, after in-depth interaction between the flight altitude and the mountain altitude, a new encoded segment may be obtained to represent the altitude adaptation advantage of this drone in the mountain inspection scenario. Conduct partial interaction on the feature group pairs with a medium association degree to generate encoded segments after partial interaction. Record the identification of the feature group pairs with a low association degree to generate identification record data. Finally, recombine the encoded segments after in-depth interaction and partial interaction in a set logical order (such as considering geographical factors first and then target factors in the task execution process, etc.), and append the identification record data in a predefined encoding form after the recombined encoded segments to generate the first drone association encoding vector. Generate the other 4 drone association encoding vectors in the same way.

[0029] Next, based on these 5 drone association encoding vectors, the cluster description encoding vector, and the flight inspection task description encoding vector, determine the target estimation confidence. First, pool these 5 drone association encoding vectors, the cluster description encoding vector, and the flight inspection task description encoding vector to generate a pooled encoding vector. Then, conduct full connection mapping processing on the pooled encoding vector based on the fully connected mapping layer. In this process, the model calculates, according to the knowledge and algorithms learned previously, a confidence representing the successful completion of the low-altitude flight inspection task in this flight inspection task under the condition that these 3 drone clusters are scheduled as the respective drone clusters in this flight inspection task, that is, the target estimation confidence.

[0030] Step S130, under the condition that the target estimation confidence meets the set requirements, use the Z drone clusters as the respective drone clusters scheduled in the one flight inspection task to obtain the adaptive detection result before the low-altitude flight inspection.

[0031] Specifically, the set requirement refers to the standard or threshold used to determine whether the target estimation confidence meets the requirement. Only when the target estimation confidence meets the set requirement will the corresponding UAV cluster be scheduled for the inspection task. For example, assume the set requirement is that the target estimation confidence is not less than 0.8. If the target estimation confidence output by the model is 0.85, then it meets the set requirement, and the corresponding UAV cluster can be scheduled for the inspection task.

[0032] The adaptive detection result refers to the evaluation result of the matching degree between the UAV cluster and the task by the target adaptive inspection decision model before scheduling the UAV cluster for the inspection task. If the evaluation result meets the set requirement, it is considered that the adaptive detection result is good and the inspection task can be executed. For example, assume the target estimation confidence output by the model is 0.85, which meets the set requirement (not less than 0.8), then it can be considered that the adaptive detection result is good, and the corresponding UAV cluster can be scheduled to perform the inspection operation in the mountain inspection task.

[0033] For another example, in this embodiment, assume the set requirement is that the deviation degree between the target estimation confidence and the threshold confidence is less than 0.1. If the target estimation confidence calculated by the server meets this requirement, for example, the target estimation confidence is 0.85 and the threshold confidence is 0.8, and their deviation degree is 0.05 which is less than 0.1, then the server will use these 3 UAV clusters as the UAV clusters scheduled for this flight inspection task. This completes the acquisition of the adaptive detection result before the low-altitude flight inspection. This adaptive detection result indicates that the current cluster division and task allocation are relatively reasonable, and these 3 UAV clusters have a high probability of successfully completing this low-altitude flight inspection task in the mountains. If the set requirement is not met, it may be necessary to re-adjust the cluster division strategy or further optimize the relevant parameters of the UAVs, etc., to ensure that the task requirements can be met and the confidence of task success can be improved.

[0034] In another case, assume the server obtains 3 estimation confidences, which are the estimation confidences determined by the target adaptive inspection decision model based on the UAV description information, cluster description information, and flight inspection task description information of each UAV cluster division data in 3 different UAV cluster division data. These 3 different cluster division strategies include the current cluster division strategy, and the number of UAV clusters included in each UAV cluster division data is the same, which is 3. If among these 3 estimation confidences, the deviation degree between the current target estimation confidence and the threshold confidence is the smallest, then the server will also use these 3 UAV clusters as the UAV clusters scheduled for this flight inspection task.

[0035] Another situation is that among the three estimated confidence levels, the deviation between the target estimated confidence level and the threshold confidence level is the smallest and this deviation is less than 0.1. The server will also use these three UAV clusters as the UAV clusters scheduled for this flight inspection task to obtain the adaptive detection results before the low-altitude flight inspection. This adaptive detection result is of great significance for the efficient and accurate execution of the entire low-altitude inspection operation. It can evaluate and optimize the feasibility and success probability of the task before the flight inspection, thereby improving the quality and efficiency of the entire inspection operation.

[0036] Based on the above steps, the embodiment of the present application realizes the accurate evaluation of the ability of the UAV cluster to perform the inspection task by integrating and analyzing the multi-dimensional description information of the UAV individual, the UAV cluster, and the flight inspection task. Using the target adaptive inspection decision model learned by prior knowledge, the target estimated confidence level is generated to quantify the possibility of the UAV cluster successfully completing the low-altitude flight inspection in a specific inspection task. Under the condition that the target estimated confidence level meets the set requirements, it is possible to intelligently schedule the appropriate UAV cluster into the inspection task, thereby ensuring the efficiency and safety of the inspection operation. That is, by intelligently evaluating the matching degree between the UAV cluster and the inspection task, it is possible to ensure that the scheduled UAV cluster has the capabilities required to complete the task, avoiding task delays or failures caused by insufficient UAV performance or improper cluster configuration, thereby significantly improving the efficiency of the inspection operation. Before scheduling the UAV cluster, the capabilities of the cluster are comprehensively evaluated through the target adaptive inspection decision model, effectively avoiding potential safety risks that may be caused by UAV or cluster performance defects, and enhancing the safety and reliability of the inspection task. In addition, it is possible to intelligently select the most suitable UAV cluster for scheduling according to the specific requirements of the inspection task, realizing the optimal allocation of UAV resources and improving the resource utilization efficiency. By introducing the target adaptive inspection decision model learned by prior knowledge, the intelligent evaluation and decision-making of the UAV cluster inspection ability are realized, the intelligent level of the inspection task scheduling is improved, and the possibility of manual intervention and decision-making errors is reduced.

[0037] In a possible implementation manner, step S120 includes:

[0038] Step S121, encoding and representing the X UAV description information sets to generate corresponding X UAV description encoding vectors, encoding and representing the cluster description information to generate a corresponding cluster description encoding vector, and encoding and representing the flight inspection task description information to generate a corresponding flight inspection task description encoding vector, where the X UAV description encoding vectors correspond to the X UAVs respectively.

[0039] Step S122: Perform a relevance analysis on the X drone description encoding vectors and the flight inspection task description encoding vector respectively to generate X drone relevance encoding vectors, where the X drone relevance encoding vectors correspond to the X drones respectively.

[0040] Step S123: Determine the target estimated confidence level based on the X drone relevance encoding vectors, the cluster description encoding vector, and the flight inspection task description encoding vector.

[0041] Based on the previous example of low-altitude power line inspection operations, after obtaining the X drone description information sets, the cluster description information, and the flight inspection task description information, these information are loaded into the target adaptive inspection decision model of the pre-knowledge learning to generate the target estimated confidence level. First, the server encodes the X drone description information sets to generate the corresponding X drone description encoding vectors, encodes the cluster description information to obtain the cluster description encoding vector, and encodes the flight inspection task description information to generate the flight inspection task description encoding vector. For example, in the scenario where 5 drones (X = 5) are involved in the power line inspection task mentioned before, for each drone's description information set, which contains the knowledge field description vectors in inspection scenarios such as mountainous areas, suburban areas, and cross-river areas, the server converts this information into drone description encoding vectors according to the predetermined encoding rules. In terms of the cluster description information, the description vectors of the capabilities of 3 drone clusters (Z = 3) in each inspection scenario are encoded to form the cluster description encoding vector, and the information such as the geographical features, task objectives, and time limits related to the mountain inspection scenario in the flight inspection task description information are also encoded as the flight inspection task description encoding vector, and these 5 drone description encoding vectors correspond to the 5 drones one by one.

[0042] Next, perform a correlation analysis on each of these X (5) UAV description encoding vectors with the flight inspection task description encoding vector to generate X (5) UAV correlation encoding vectors. Taking one of the UAVs as an example, such as the UAV with good flight altitude adaptability in the mountain inspection scenario mentioned before, its description encoding vector is decomposed into different first feature groups such as flight ability related feature groups (like flight altitude adaptability), device status related feature groups (battery power, device temperature, etc.), and task history related feature groups (performance in previous similar tasks), etc. The flight inspection task description encoding vector is decomposed into second feature groups such as task area geographical feature groups (mountain altitude, mountain range trend, etc.), task objective related feature groups (location and trend of power towers and lines), and task time limit related feature groups (8 am - 5 pm), etc. Then, based on semantic understanding and logical relationships, perform a preliminary correlation analysis between the first feature groups and the second feature groups one by one. For example, analyze the flight altitude adaptability in the flight ability related feature group with the mountain altitude range in the task area geographical feature group. If the flight altitude of this UAV adapts to the mountain altitude range and the correlation degree is high, establish a correlation relationship description document to record the feature group pair, correlation basis, and correlation degree. For those with a high correlation degree, perform in-depth interaction operations. For example, after the in-depth interaction between the flight altitude and the mountain altitude, obtain an encoding segment representing the altitude adaptation advantage. For those with a medium correlation degree, perform partial interaction to obtain the corresponding encoding segment. For those with a low correlation degree, perform identification recording to obtain identification record data. Finally, recombine the encoding segments after in-depth interaction and partial interaction in the order of the task execution process and append the identification record data to form the UAV correlation encoding vector of this UAV. Operate on the other 4 UAVs in the same way to obtain 5 UAV correlation encoding vectors, which correspond to the 5 UAVs respectively.

[0043] Finally, the server determines the target estimated confidence based on these X (5) drone association coding vectors, the cluster description coding vector, and the flight inspection task description coding vector. The server pools together these 5 drone association coding vectors, the cluster description coding vector, and the flight inspection task description coding vector to generate a pooled coding vector. Then, it performs a fully connected mapping process on the pooled coding vector using the fully connected mapping layer in the target adaptive inspection decision model. During this process, the model comprehensively analyzes and calculates various information in the pooled coding vector based on the previously learned knowledge and algorithms. For example, information such as the flight capabilities and equipment status of the drones in the drone association coding vector, the overall performance information of the cluster in the cluster description coding vector, and the task requirement information in the flight inspection task description coding vector are all taken into consideration. Through the complex calculation process within the model, a numerical value is finally obtained, and this numerical value is the target estimated confidence, which represents the confidence that these 3 (Z = 3) drone clusters can successfully complete the low-altitude flight inspection task in this flight inspection task under the condition that these 3 drone clusters are scheduled as the individual drone clusters for this flight inspection task.

[0044] In a possible implementation manner, step S122 includes:

[0045] Performing a relevance analysis on the a-th drone description coding vector among the X drone description coding vectors and the flight inspection task description coding vector based on the following steps to generate the a-th drone association coding vector among the X drone association coding vectors, where a is a positive integer not less than 1 and not greater than X:

[0046] Step S1221: Interact the a-th drone description coding vector with the flight inspection task description coding vector to generate the a-th blended coding vector.

[0047] Step S1222: Use the aggregation result of the a-th drone description coding vector and the a-th blended coding vector as the a-th drone association coding vector. Alternatively, perform a fusion calculation on the a-th blended coding vector and the pre-defined a-th influencing factor to generate the a-th fusion coding vector, and use the aggregation result of the a-th drone description coding vector and the a-th fusion coding vector as the a-th drone association coding vector.

[0048] Among them, step S1221 includes:

[0049] Step S1221-1: Decompose the a-th drone description coding vector into multiple different first feature groups, which are divided based on different attributes of the drone, specifically including a flight ability-related feature group, an equipment status-related feature group, and a task history-related feature group.

[0050] And, step S1221-2, decompose the described flight inspection task description coding vector into multiple second feature groups related to different task scenario requirements, specifically including a task area geographical feature group, a task objective related feature group, and a task time limit related feature group.

[0051] Step S1221-3, for each first feature group, based on semantic understanding and logical relationships, conduct preliminary association analysis with all second feature groups one by one, and establish an association relationship description document for each preliminary association analysis result. The association relationship description document is used to record the feature group pairs composed of the associated first feature group and second feature group, as well as the basis for the association and the initially judged degree of association.

[0052] Step S1221-4, according to the association relationship description document, perform in-depth interaction operations on the feature group pairs with a high degree of association to generate coded segments after in-depth interaction, perform partial interaction on the feature group pairs with a medium degree of association to generate coded segments after partial interaction, and perform identification records on the feature group pairs with a low degree of association to generate identification record data.

[0053] Step S1221-5, recombine the coded segments after in-depth interaction and partial interaction in a set logical order. The logical order is determined based on the task execution process or the importance of the data. Also, append the identification record data in a predefined coding form after the recombined coded segments to generate the a-th blended coding vector.

[0054] In this embodiment, continue to elaborate on this part in detail based on the previous example of low-altitude inspection operations of power lines.

[0055] In the previous steps, X drone description coding vectors and flight inspection task description coding vectors have been obtained. Now, perform correlation analysis on them to generate X drone correlation coding vectors. Take the a-th drone among the X drones (here assume a = 3, in the scenario where 5 drones participated in the inspection task before) as an example to illustrate this process.

[0056] First, the third UAV description encoding vector interacts with the flight inspection task description encoding vector to generate the third fusion encoding vector. First, the third UAV description encoding vector is decomposed into multiple different first feature groups, including flight ability-related feature groups, device status-related feature groups, and task history-related feature groups. For the flight ability-related feature groups, in the power line inspection scenario, it may include encoding information such as the maximum flight speed of the UAV during mountain inspections, the minimum turning radius, and the flight stability when dealing with different altitudes; the device status-related feature groups may cover encoding information such as the current battery power percentage of the UAV, the operating temperatures of various key devices (such as motors, sensors, etc.), and the remaining service life of the device; the task history-related feature groups may record encoding information such as the number of successful and failed times of this UAV in previous similar inspection tasks, and its performance when encountering special situations (such as strong wind interference, signal interruption, etc.).

[0057] At the same time, the flight inspection task description encoding vector is decomposed into multiple different second feature groups related to task scenario requirements, specifically the task area geographical feature group, the task objective-related feature group, and the task time limit-related feature group. In the mountain power line inspection scenario, the task area geographical feature group includes encoding information such as the average altitude of the mountains, the undulation of the mountains, the distribution of valleys and peaks, the distribution of rivers and roads and their potential impact on flight; the task objective-related feature group includes encoding information such as the specific coordinate positions of the power towers to be inspected, the orientation of the power lines (whether along valleys or ridges, whether crossing rivers, etc.), the height and type of the power towers; the task time limit-related feature group includes encoding information such as the specified start time (such as 8 am) and end time (such as 5 pm) of this inspection task, and possible special requirements for inspection speed or accuracy during different time periods (for example, reducing the flight speed to improve inspection accuracy when strong sunlight at noon may affect imaging quality).

[0058] Next, for each first feature group, based on semantic understanding and logical relationships, conduct preliminary correlation analysis with all second feature groups one by one, and establish a correlation relationship description document for each preliminary correlation analysis result. For example, for the maximum flight speed in the flight ability related feature group, conduct preliminary correlation analysis with the mountain undulation degree in the task area geographical feature group. If the mountain undulation is large and the UAV needs to have a high flight speed to efficiently complete the inspection task, then it will be judged that their correlation degree is high, and record in the correlation relationship description document the feature group pair composed of the maximum flight speed in the flight ability related feature group and the mountain undulation degree in the task area geographical feature group, and at the same time record the basis for the correlation (that is, large mountain undulation requires high speed to ensure inspection efficiency) and the preliminarily judged correlation degree (high). Conduct preliminary correlation analysis on the battery power percentage in the device status related feature group and the task duration in the task time limit related feature group. If the task duration is long and the battery power percentage is low, the correlation degree may be judged as medium and recorded in the correlation relationship description document. Another example is the preliminary correlation between the performance in dealing with special situations in the task history related feature group and the type of power tower in the task objective related feature group. If a certain type of power tower has never had special situations in previous inspections, then the correlation degree may be low, and it is also recorded in the correlation relationship description document.

[0059] According to the correlation relationship description document, perform in-depth interaction operations on the feature group pairs with a high correlation degree to generate encoded segments after in-depth interaction. For example, for the feature group pair of the maximum flight speed and the mountain undulation degree with a high correlation degree mentioned above, the in-depth interaction operation may be to accurately calculate the optimal flight speed range that meets efficient inspection based on the mountain undulation degree, and fuse the encoding related to this speed range with the original maximum flight speed encoding to generate an encoded segment after in-depth interaction. For the feature group pairs with a medium correlation degree, perform partial interaction to generate encoded segments after partial interaction. For example, for the feature group pair of the battery power percentage and the task duration with a medium correlation degree, the partial interaction may be to estimate the battery power consumption based on the task duration, and fuse the encoding of this estimation result with the original battery power percentage encoding to obtain an encoded segment after partial interaction. For the feature group pairs with a low correlation degree, perform identification records to generate identification record data. For example, for the low-correlation feature group pair of the performance in dealing with special situations and the type of power tower, the identification record data may be simply to record this low-correlation situation and the corresponding feature group content.

[0060] Then, the coded segments after deep interaction and partial interaction are reassembled in a set logical order, which is determined based on the task execution process or the importance of the data. In this power line inspection scenario, according to the task execution process, the coded segments related to the geographical features of the task area should be considered first, because it will directly affect the flight path planning of the drone, so the coded segments after deep interaction or partial interaction with geographical features such as the undulation of the mountain range are placed first. Next is the coded segment related to the task goal, because the inspection goal is the core of the entire task, and then the coded segments related to the task goals such as the type and location of the power tower are placed. Finally, the coded segments related to the task time limit are placed, because it is more of a time constraint on the entire inspection process. After the coded segments are reassembled in this order, the previously generated identification record data is attached to the reassembled coded segments in a predefined coding form, so that the third fusion coding vector is generated.

[0061] After the third interleaved coding vector is generated, there are two ways to obtain the third drone-associated coding vector. One way is to directly aggregate the third drone description coding vector with the third interleaved coding vector, merge or calculate the corresponding codes in the two vectors according to certain rules (such as simple addition or weighted addition), and the result is the third drone-associated coding vector. Another way is to fuse the third interleaved coding vector with the third previously defined influencing factor, which is a numerical value or coding form predetermined based on previous experience or a specific algorithm. For example, this influencing factor may be a quantitative value of the impact of external factors such as the current season and weather conditions on drone inspection. The third interleaved coding vector is fused and calculated with this influencing factor, such as multiplication or calculation according to a certain functional relationship, to generate the third fused coding vector, and then the third drone description coding vector is aggregated with the third fused coding vector (also according to the aforementioned merging or calculation rules), and finally the third drone-associated coding vector is obtained. In the same way, for other drones (from the 1st to the Xth), the corresponding drone association coding vectors can be generated respectively, thus completing the entire correlation analysis and generating the X drone association coding vectors. This process fully considers the various characteristics of the drone itself and the various requirements of the flight inspection mission. Through the above correlation analysis and interactive operations, it lays the foundation for the subsequent accurate determination of the target estimation confidence.

[0062] In a possible implementation, step S123 includes:

[0063] Step S1231: Pool the X drone-associated coding vectors, the cluster description coding vector, and the flight inspection task description coding vector to generate a pooled coding vector.

[0064] Step S1232: Perform a fully connected mapping process on the pooled coding vector based on a fully connected mapping layer to generate the target estimated confidence.

[0065] In this embodiment, in previous operations, X drone-associated coding vectors, a cluster description coding vector, and a flight inspection task description coding vector have been obtained respectively. Now, based on these coding vectors, the target estimated confidence is to be determined.

[0066] First, the server pools the X drone-associated coding vectors, the cluster description coding vector, and the flight inspection task description coding vector to generate a pooled coding vector. In the previous example, assume that 5 (X = 5) drones are involved in the inspection task. Then these 5 drone-associated coding vectors contain information after the correlation analysis of each drone with the flight inspection task in multiple aspects. For example, each drone-associated coding vector contains coding information after the interaction of the drone's flight ability with the geographical features of the task area, task objectives, and task time limits, coding information after the interaction of the device status with these task requirements, and coding information after the interaction of the task history with the task requirements, etc. The cluster description coding vector contains coding information of the overall performance and other description vectors of 3 (Z = 3) drone clusters in multiple inspection scenarios, such as the coding of the collaborative ability of drones within the cluster to cope with different geographical environments and the overall stability in different task scenarios. The flight inspection task description coding vector contains the coding of information such as the regional geographical features (mountain ranges, altitudes, etc.), task objectives (positions of power towers and lines, etc.), and task time limits (from 8 am to 5 pm) of this mountain power line inspection task.

[0067] When pooling these coding vectors, they can be combined into a new pooled coding vector according to certain rules. This pooled coding vector is like a large set that combines all relevant information, and each part of the coding in it represents factors related to the success or failure of the inspection task. For example, the coding related to the flight stability of the drone in the drone-associated coding vector, the coding related to the overall flight stability of the cluster in the cluster description coding vector, and the coding related to the regional geographical features of the task area (such as the mountain undulations affecting flight stability) in the flight inspection task description coding vector are correlated with each other in the pooled coding vector, jointly reflecting the overall situation of the entire inspection task in terms of flight stability.

[0068] Next, the server performs a fully connected mapping process on the aggregated encoded vector based on the fully connected mapping layer, thereby generating the target estimated confidence. The fully connected mapping layer is an important part of the target adaptive inspection decision model, which has preset connection weights and algorithms. When performing the fully connected mapping process, each element in the aggregated encoded vector is connected to the neurons in the fully connected mapping layer and calculated according to the set weights.

[0069] Taking the encoding related to the flight ability of the unmanned aerial vehicle (UAV) in the aggregated encoded vector as an example, in the fully connected mapping layer, the neurons related to flight ability will process this part of the encoding according to the pre-learned weights. If it is found in the previous learning process that the high flight speed of the UAV is an important success factor in the mountain inspection task, then during the fully connected mapping process, the encoding elements related to flight speed will be assigned higher weights for calculation. Similarly, for the encoding related to the cluster cooperation ability in the cluster description encoded vector, the corresponding neurons in the fully connected mapping layer will determine the weights and perform calculations according to the importance of the cluster cooperation ability for the success of the inspection task. For the encoding related to the task time limit in the flight inspection task description encoded vector, the fully connected mapping layer will also determine the weights for calculation according to the impact degree of the time limit on the task.

[0070] In this complex calculation process, the fully connected mapping layer will comprehensively consider all the information in the aggregated encoded vector, not only the individual flight ability, cluster cooperation ability, or task time limit, but also the relationships and interactive effects between them. For example, the relationship between the flight ability of the UAV and its equipment status, and the relationship between the cluster cooperation ability and the task objective will be included in the calculation scope.

[0071] After a series of calculations by the fully connected mapping layer, a value will finally be obtained, and this value is the target estimated confidence. This target estimated confidence represents the confidence that the three UAV clusters corresponding to the three previously divided UAV clusters will successfully complete the low-altitude flight inspection task in this flight inspection task under the condition of being scheduled to this flight inspection task. For example, if the obtained target estimated confidence is 0.8, it means that based on the existing UAV individual information, cluster information, and flight inspection task information, the server believes that there is an 80% probability that these three UAV clusters will successfully complete the low-altitude flight inspection task of the mountain power line in this inspection. This target estimated confidence provides an important basis for subsequent decisions, such as whether to execute the inspection task according to the current cluster division, or whether it is necessary to adjust certain parameters to improve the confidence, etc.

[0072] In a possible implementation manner, step S110 includes:

[0073] Step S111, obtain the flight trajectory description information, immediate status description information, and long-term status description information of each of the X drones, and generate the X-drone description information set. The flight trajectory description information of each drone includes the flight trajectory information of each drone in the previous b flight inspection tasks. The long-term status description information of each drone includes the record information of each drone in each inspection execution scenario among the multiple inspection execution scenarios within the previous set time period. The immediate status description information of each drone includes the current record information of each drone in each inspection execution scenario among the multiple inspection execution scenarios.

[0074] Step S112, obtain the average performance statistical data of each of the Z drone clusters in each inspection execution scenario among the multiple inspection execution scenarios.

[0075] Step S113, obtain the scenario label index of the target inspection execution scenario, or obtain the scenario label index of the target inspection execution scenario and the time sequence node scheduled for the one-time flight inspection task.

[0076] In this embodiment, in the scenario of power line inspection, the server needs to obtain the flight trajectory description information, immediate status description information, and long-term status description information of each of the X drones to generate the X-drone description information set.

[0077] Regarding the flight trajectory description information of each drone, assume X = 5. These 5 drones have all participated in b = 3 flight inspection tasks before. Taking one of the drones as an example, during these 3 flight inspection tasks, its flight trajectory information contains a lot of content. In the first inspection task, this drone took off from the base and flew along the set route towards the power line inspection area in the mountainous area. Its flight trajectory might be to fly along a valley first and then cross a small hill. During this process, the server recorded information such as its flight altitude, speed, turning points, etc. For example, when it was flying in the valley, in order to avoid some tall trees, the flight altitude was maintained at about 80 meters and the speed was 30 meters per second. When crossing the hill, the flight altitude increased to 120 meters and the speed slightly decreased to 25 meters per second. The second inspection task might be for power line inspection in the suburban area of the city, and the flight trajectory was different. It needed to avoid high-rise buildings and densely populated areas. The flight altitude might be continuously adjusted between 100 - 150 meters, and the speed also fluctuated between 20 - 30 meters per second according to the complexity of the surrounding environment. The server detailedly recorded information such as the change points and change times of these trajectories. If the third inspection task was for inspection across a river area, when the drone approached the river, due to possible influence of water surface reflection and airflow, the flight trajectory would be adjusted accordingly, such as maintaining a relatively high safe distance from the river surface. The flight altitude might be maintained at about 130 meters, and the speed would also be adjusted according to the wind direction and wind force. All these flight trajectory information were completely recorded.

[0078] The immediate status description information of each drone includes the current recorded information of each drone under multiple inspection execution scenarios. Still taking this drone as an example, in the current mountain inspection scenario, its immediate status description information includes the current battery power, assumed to be 80%, the operating temperatures of each device, such as the motor temperature being 40 degrees Celsius and the sensor temperature being 35 degrees Celsius. It also includes the current flight altitude of 90 meters, the flight speed of 28 meters per second, and the current good communication signal strength with the control center, etc. In the suburban area inspection scenario of the city, the immediate status might be a battery power of 70% (because signal interference in the city might increase power consumption), the device temperature slightly increased, the flight altitude of 120 meters, the flight speed of 25 meters per second, and the communication signal was interfered by some buildings but still remained stable, etc. In the inspection scenario across the river area, the immediate status might be a battery power of 75%, the device temperature normal, the flight altitude of 135 meters, the flight speed of 27 meters per second, and the communication signal good, etc.

[0079] The long-term status description information of each drone includes the recorded information of this drone in each inspection execution scenario within multiple inspection execution scenarios during the previous set period. Assuming the set period is the past month, within this month, the long-term status record of this drone in the mountain inspection scenario may show that it participated in a total of 5 inspection tasks. Among them, once the flight trajectory deviated from the predetermined route due to strong wind weather, but it was corrected in time through the automatic adjustment system. In the suburban inspection scenario, it had 3 inspection tasks within this month. Once there was a brief communication interruption due to signal blockage by buildings, but it recovered quickly. In the cross-river area inspection scenario, there were 2 inspection tasks, and the overall performance was good. However, during one inspection, it was found that the imaging quality of the imaging device decreased slightly when the water surface had strong reflection. All these long-term status information were collected by the server. By obtaining these three aspects of information (flight trajectory description information, instant status description information, and long-term status description information), the server generated corresponding drone description information sets for these 5 drones respectively.

[0080] Next is to obtain the cluster description information. In this power line inspection operation, according to the current cluster division strategy, 5 drones are divided into Z = 3 drone clusters. The server needs to obtain the average performance statistical data of each drone cluster in each inspection execution scenario within these 3 drone clusters. Taking one of the drone clusters as an example, in the mountain inspection scenario, its average performance statistical data includes a lot of content. The average flight speed of the drones within this cluster is 28 m / s, the average flight altitude is 95 m, the average battery power consumption during a complete mountain inspection task is 60%, and the average communication signal strength between the drones within the cluster is at a good level. When dealing with special mountain weather (such as strong wind), the overall stability of the drones within the cluster is good, and the average distance deviated from the predetermined route is within an acceptable range, such as not exceeding 5 m. In the suburban inspection scenario, the average performance statistical data of this cluster shows that the average flight speed is 25 m / s, the average flight altitude is 110 m, and the average battery power consumption is 65% (because of more signal interference). The communication and cooperation ability between the drones within the cluster is shown as when encountering signal blockage by buildings, it can maintain an information transmission success rate of more than 90% through the communication adjustment mechanism within the cluster. In the cross-river area inspection scenario, the average flight speed of this cluster is 27 m / s, the average flight altitude is 130 m, and the average battery power consumption is 62%. In terms of dealing with the impact of water surface reflection on imaging quality, the overall response ability of the drones within the cluster is to adjust the flight angle and imaging parameters, so that more than 80% of the power lines and power towers can be imaged clearly. The server obtains such average performance statistical data for each cluster in different inspection scenarios, thereby obtaining the cluster description information.

[0081] Finally, obtain the flight inspection task description information. In this example, the server needs to obtain the scenario tag index of the target inspection execution scenario, or obtain the scenario tag index of the target inspection execution scenario and the time sequence nodes scheduled for a flight inspection task. Suppose the target inspection execution scenario this time is the mountain inspection scenario, and its scenario tag index may be a specific code. For example, "001" represents the mountain inspection scenario. This scenario tag index contains some basic feature information of the mountain inspection scenario, such as the index of geographical environment features (mountain range orientation, altitude range, etc.), possible weather types (windy, foggy, etc.), and the distribution density of power lines and power towers. If it is also necessary to obtain the time sequence nodes scheduled for a flight inspection task, assume that this time sequence node is to start this flight inspection task at 9 am. This time sequence node information is of great significance for the planning and execution of the entire inspection task. For example, the weather conditions (such as light intensity, wind direction and wind force, etc.) at 9 am will affect the flight safety and inspection effect of the drone. At the same time, this time point is also related to factors such as the peak usage period of the power line. It may be necessary to focus on the status of certain key power equipment during this time period. By obtaining this flight inspection task description information, the server provides an important basis for subsequent decision-making and analysis.

[0082] In a possible implementation manner, step S130 includes:

[0083] Step S131, under the condition that the deviation degree between the target estimated confidence level and the threshold confidence level is less than the set value, use the Z drone clusters as the respective drone clusters scheduled for the flight inspection task. Or

[0084] Step S132, under the condition that the deviation degree between the target estimated confidence level and the threshold confidence level is the smallest among the A obtained estimated confidence levels, use the Z drone clusters as the respective drone clusters scheduled for the flight inspection task. The A estimated confidence levels include the estimated confidence levels determined by the target adaptive inspection decision model based on the drone description information, cluster description information, and flight inspection task description information of each drone cluster division data in the A drone cluster division data. Each drone cluster division data in the A drone cluster division data includes multiple drone clusters generated by grouping the X drones according to a corresponding one of the A different cluster division strategies. The A different cluster division strategies include the current cluster division strategy. The number of drone clusters included in each drone cluster division data is the same, and A is not less than 2. Or

[0085] Step S133, under the condition that the deviation degree between the target estimated confidence level and the threshold confidence level among the A estimated confidence levels is the smallest and the deviation degree is less than the set value, the Z UAV clusters are used as the UAV clusters scheduled for the one-time flight inspection task.

[0086] In this embodiment, in the previous steps, the target estimated confidence level has been obtained through the target adaptive inspection decision model. This target estimated confidence level represents the possibility that these clusters can successfully complete the low-altitude flight inspection task (such as the mountain power line inspection task mentioned above) when the currently set Z UAV clusters (assuming Z = 3) are used as the UAV clusters scheduled for this flight inspection task. Now, it is necessary to decide whether to schedule these Z UAV clusters for this flight inspection task according to whether the target estimated confidence level meets the set requirements.

[0087] First, consider the situation where the Z UAV clusters are used as the UAV clusters scheduled for the one-time flight inspection task under the condition that the deviation degree between the target estimated confidence level and the threshold confidence level is less than the set value. Assume that the threshold confidence level is set to 0.7 and the set value is 0.1. If the target estimated confidence level calculated by the server is 0.75, then the deviation degree between the target estimated confidence level and the threshold confidence level is 0.05, and this deviation degree is less than the set value of 0.1. This means that under the evaluation of the model, the current 3 UAV clusters are very close to the ideal confidence level of successfully completing the task (threshold confidence level).

[0088] From the perspective of the actual inspection task, the comprehensive capabilities shown by these 3 UAV clusters in the previous various information analyses (including the flight trajectories, immediate states, long-term states of each UAV, and the average performance statistical data of the clusters, etc.) are evaluated and matched with the requirements of this flight inspection task (such as the geographical characteristics of the task area, task objectives, task time limits, etc.). The result of being close to the threshold confidence level indicates that they have sufficient ability to perform the task. For example, the adaptability of the flight trajectories of the UAVs in these 3 clusters, the stability of the equipment states, and the ability to handle various emergencies (such as weather changes, geographical environment interference, etc.) in the previous mountain inspection tasks, combined with the specific requirements of this task (such as specific inspection routes, inspection times, etc.), are comprehensively evaluated by the model and considered to have a high possibility of success and a small deviation. Therefore, the server can use these 3 UAV clusters as the UAV clusters scheduled for this flight inspection task.

[0089] Next, consider the case where, under the condition that the deviation degree between the target estimated confidence and the threshold confidence among the A obtained estimated confidences is the smallest, Z UAV clusters are used as the respective UAV clusters scheduled for a single flight inspection task. Assume A = 3, that is, the server determines 3 estimated confidences based on the UAV description information, cluster description information, and flight inspection task description information of each UAV cluster division data in 3 different UAV cluster division data through the target adaptive inspection decision model. These 3 different cluster division strategies all group X UAVs (assuming X = 5) into clusters to generate multiple UAV clusters, and the number of clusters generated in each cluster division data is the same (both are 3).

[0090] Under the first cluster division strategy, the estimated confidence calculated by the model is 0.65; the target estimated confidence obtained under the second cluster division strategy (i.e., the current cluster division strategy) is 0.75; the estimated confidence obtained under the third cluster division strategy is 0.6. It can be seen that among these 3 estimated confidences, the deviation degree between the target estimated confidence of 0.75 obtained under the current cluster division strategy and the threshold confidence (assuming it is still 0.7) is the smallest.

[0091] Analyzing from the actual situation, different cluster division strategies will lead to different cluster combination methods. The UAVs in each cluster are matched differently, and their overall performance and ability to handle tasks will also vary. Under the first cluster division strategy, the estimated confidence may be low due to factors such as poor coordination between some UAVs or unreasonable resource allocation within the cluster. Under the current cluster division strategy, the performance of the 3 UAV clusters is more balanced and excellent in all aspects. For example, in a certain cluster, factors such as the flight capabilities, equipment status, and task history of the UAVs complement each other, making this cluster more adaptable to the complex geographical environment, power line distribution, and task time limits in mountainous areas. Therefore, in the model evaluation, the target estimated confidence with the smallest deviation degree from the threshold confidence is obtained, so the server will select these 3 UAV clusters (under the current cluster division strategy) as the respective UAV clusters scheduled for this flight inspection task.

[0092] Finally, consider the case where, among the A estimated confidence levels, the deviation between the target estimated confidence level and the threshold confidence level is minimized and the deviation is less than the set value, and Z UAV clusters are used as the UAV clusters scheduled for a single flight inspection task. Still assume that A = 3, the threshold confidence level is 0.7, and the set value is 0.1. The estimated confidence level obtained under the first cluster division strategy is 0.6, deviating from the threshold confidence level by 0.1; the target estimated confidence level obtained under the second cluster division strategy (the current cluster division strategy) is 0.75, deviating from the threshold confidence level by 0.05; the estimated confidence level obtained under the third cluster division strategy is 0.68, deviating from the threshold confidence level by 0.02. It can be seen that the deviation of 0.05 between the target estimated confidence level of 0.75 and the threshold confidence level under the current cluster division strategy is the smallest, and this deviation of 0.05 is less than the set value of 0.1.

[0093] From the perspective of the actual execution of the inspection task, this indicates that the current 3 UAV clusters not only have the smallest deviation from the threshold confidence level, but also this smallest deviation is within the acceptable set range. This means that this cluster division scheme is the most likely to successfully complete this flight inspection task after considering various factors such as the individual capabilities of UAVs, the overall performance of the cluster, and task requirements. For example, among the current 3 UAV clusters, the experience and capabilities accumulated by each cluster in different previous inspection scenarios (such as strong winds in mountainous areas, signal interference in suburban areas, and water surface reflection in cross-river areas) are combined with the specific requirements of this mountainous area power line inspection task (such as specific power tower locations, inspection time limits, etc.). Under the evaluation of the model, they are combined in an optimal way, which not only meets the requirement of closeness to the threshold confidence level but also is within the acceptable deviation range. Therefore, the server will use these 3 UAV clusters as the UAV clusters scheduled for this flight inspection task, thus providing a reliable cluster scheduling scheme for the low-altitude flight inspection task and helping to improve the success rate and efficiency of the inspection task.

[0094] In a possible implementation manner, the method further includes:

[0095] Step A110: Obtain a sample learning data sequence and a sample confidence sequence. The sample learning data sequence includes B sample learning data. Each sample learning data includes X example UAV cluster description information, example cluster description information, and example flight inspection task description information. The X example UAV cluster description information includes Y sample knowledge field description vectors of each of the X example UAVs in the example UAV cluster. The X example UAVs are the UAVs in a sample flight inspection task in the example inspection execution scenario scheduled for the target low-altitude inspection operation. The Y sample knowledge field description vectors of each of the X example UAVs include the description vectors of each of the X example UAVs in the multiple inspection execution scenarios. The example cluster description information includes the description vectors of each of the Z example UAV clusters in the multiple inspection execution scenarios. The example flight inspection task description information includes the description vector of the example inspection execution scenario. The Z example UAV clusters are multiple example UAV clusters generated by clustering and grouping the X example UAVs according to an example cluster division strategy. The sample confidence sequence includes B sample confidences corresponding to the B sample learning data respectively, and B is not less than 2.

[0096] Step A120: Perform model parameter learning on the initialized adaptive inspection decision model according to the sample learning data sequence until the value of the target training cost function corresponding to the adaptive inspection decision model meets the pre-defined training termination requirement, then terminate the training, and use the adaptive inspection decision model at the end of the training as the target adaptive inspection decision model. Under the condition that the value of the target training cost function corresponding to the adaptive inspection decision model does not meet the pre-defined training termination requirement, the neuron weight coefficients in the adaptive inspection decision model are updated. The value of the target training cost function is a loss value determined based on the example estimated confidence generated by the adaptive inspection decision model and the corresponding sample confidence among the B sample confidences. The example estimated confidence is the estimated confidence generated by the adaptive inspection decision model based on the X example UAV cluster description information, the example cluster description information, and the example flight inspection task description information.

[0097] In this embodiment, within the framework of low-altitude inspection operation of power lines, the server needs to train a model to obtain a target adaptive inspection decision model that can accurately make inspection decisions.

[0098] First, the server needs to obtain a sample learning data sequence and a sample confidence sequence. Assume B = 3, that is, the sample learning data sequence contains 3 sample learning data.

[0099] For the first sample learning data, it contains X pieces of description information of example UAV clusters. Assume X = 5, and these 5 example UAVs are the UAVs for a sample flight inspection task in an example inspection execution scenario where they are scheduled for target low-altitude inspection operations. Each example UAV has Y sample knowledge field description vectors. Assume Y = 3. These 3 sample knowledge field description vectors cover the description vectors of each example UAV in multiple inspection execution scenarios (such as mountain inspection scenarios, suburban inspection scenarios, cross-river area inspection scenarios).

[0100] Taking one of the example UAVs as an example, in a mountain inspection scenario, one of its sample knowledge field description vectors may contain data related to its flight altitude adaptability. For example, the optimal flight altitude range when it flies at different altitudes in the mountains, and the altitude adjustment strategy when flying near slopes with different gradients. In a suburban inspection scenario, another sample knowledge field description vector may record its flight path planning ability when dealing with densely built-up areas, such as how to bypass numerous buildings with the shortest path and ensure effective inspection of power lines. In a cross-river area inspection scenario, the third sample knowledge field description vector may contain data related to its anti-airflow interference ability when flying above the water surface and the adaptability of the imaging device to the water surface reflection.

[0101] The example cluster description information includes the description vectors of each example UAV cluster in the Z example UAV clusters in multiple inspection execution scenarios. Assume Z = 3. For one of the example UAV clusters, in a mountain inspection scenario, its description vector may contain data on the overall flight speed coordination of the UAVs within the cluster. For example, in the complex terrain of the mountains, how the UAVs within the cluster adjust their respective speeds to maintain a relatively stable spacing and flight order, so as to efficiently complete the inspection task. In a suburban inspection scenario, it may contain information such as the collaborative communication ability of the UAVs within the cluster when dealing with frequent communication signal switching (due to building blockage). In a cross-river area inspection scenario, it may contain relevant data such as the overall coping strategy of the UAVs within the cluster for the impact of a humid environment on equipment performance.

[0102] The sample flight inspection task description information includes the description vector of the sample inspection execution scenario. If the sample inspection execution scenario is a mountain inspection scenario, this description vector may contain detailed geographical features of the mission area, such as the specific altitude difference range of the mountains, the distribution density of valleys and peaks, the intersection of rivers and power lines, etc. It also includes mission target-related features, such as the specific type of power tower that needs to be inspected (high-voltage tower or ordinary voltage tower), the special direction of the power line (whether it is along the ridge and close to dangerous areas such as cliffs), etc. It also includes mission time limit-related features, such as whether the sample inspection task needs to complete a specific part of the inspection work in a specific time period (such as the early morning when the light is dim but the power load is low).

[0103] At the same time, the sample confidence sequence includes three sample confidences corresponding to the three sample learning data. These sample confidences are obtained by other methods (such as manual evaluation or statistical results of some verified data) before, indicating the confidence that the sample drone cluster successfully completes the sample flight inspection task under the corresponding sample learning data.

[0104] Next, the server learns the model parameters of the initialized adaptive inspection decision model based on the sample learning data sequence.

[0105] The initialized adaptive inspection decision model is like an intelligent evaluation system in its initial state, and its internal neuron weight coefficients are set randomly or according to some default rules. The server loads the first sample learning data into this initialized model. In the model, a series of processing is performed based on the X example drone cluster description information, the example cluster description information, and the example flight inspection task description information. For example, the model will analyze the description vectors of each example knowledge field of the example drone to identify the factors that have an important impact on the success of the inspection task. It will also analyze the overall cluster performance factors in the example cluster description information and the task requirement factors in the example flight inspection task description information.

[0106] Then, the model determines the confidence level of the generated example estimate based on this information. This example estimate confidence level is the model's assessment of the likelihood that the example drone cluster will successfully complete the task in the sample flight inspection task based on its current parameter state (initial neuron weight coefficients, etc.).

[0107] Based on this example estimation confidence and the corresponding sample confidence, determine the cost value of the target training cost function. This cost value is actually a loss value, which reflects the gap between the model's example estimation confidence and the known sample confidence. If this cost value is large, it means that the current parameter setting of the model makes its evaluation result differ greatly from the actual sample confidence, and the model needs to be adjusted.

[0108] If the cost value of this target training cost function does not meet the previously defined training termination requirement (for example, the training termination requirement may be that the cost value is less than a set threshold, and assume this threshold is 0.1), then the server will update the neuron weight coefficients in the adaptive patrol decision model. The update method can be based on some optimization algorithms, such as the gradient descent algorithm. By adjusting the neuron weight coefficients, the model will reprocess the example learning data, generate the example estimation confidence again, and recalculate the cost value of the target training cost function.

[0109] The server will process the second example learning data and the third example learning data in the same way. For each processed example learning data, the cost value of the target training cost function will be calculated based on the generated example estimation confidence and the corresponding example confidence, and it will be determined whether to update the neuron weight coefficients of the model according to whether this cost value meets the training termination requirement.

[0110] This process will continue until the cost value of the target training cost function corresponding to the adaptive patrol decision model meets the previously defined training termination requirement. For example, after multiple processes of the example learning data and adjustments of the model parameters, the cost value calculated from the gap between the example estimation confidence and the corresponding example confidence obtained after the model processes the third example learning data is 0.08, which is less than the set threshold of 0.1. At this time, the server will terminate the training and use the adaptive patrol decision model at the time of terminating the training as the target adaptive patrol decision model. This target adaptive patrol decision model has been adjusted to a relatively accurate state through the training of the example learning data, and can generate the target estimation confidence relatively accurately according to the input UAV description information, cluster description information, and flight patrol task description information, so as to provide a reliable basis for the actual UAV patrol task scheduling decision.

[0111] In a possible implementation manner, step A120 includes:

[0112] Perform model optimization in the a-th model optimization phase based on the following steps, where a is not less than 2:

[0113] Step A121, load the a-th example learning data in the B example learning data into the adaptive patrol decision model generated in the (a - 1)-th model optimization phase. The a-th example learning data corresponds to the a-th example confidence in the B example confidences. The a-th example learning data includes Xa pieces of example UAV cluster description information, the a-th example cluster description information, and the a-th example flight patrol task description information, where Xa = X.

[0114] Step A122: In the adaptive inspection decision model generated in the (a - 1)-th model optimization stage, encode the Xa sets of example UAV description information to generate corresponding Xa example UAV description encoding vectors, encode the a-th example cluster description information to generate a corresponding a-th example cluster description encoding vector, and encode the a-th example flight inspection task description information to generate a corresponding a-th example flight inspection task description encoding vector. The Xa example UAV description encoding vectors respectively correspond to the Xa example UAVs in the a-th sample learning data.

[0115] Step A123: In the adaptive inspection decision model generated in the (a - 1)-th model optimization stage, perform a correlation analysis on the Xa example UAV description encoding vectors and the a-th example cluster description encoding vector respectively to generate Xa example UAV correlation encoding vectors. The Xa example UAV correlation encoding vectors respectively correspond to the Xa example UAVs in the a-th sample learning data.

[0116] Step A124: In the adaptive inspection decision model generated in the (a - 1)-th model optimization stage, determine the a-th estimated confidence level based on the Xa example UAV correlation encoding vectors, the a-th example cluster description encoding vector, and the a-th example flight inspection task description encoding vector.

[0117] Step A125: Determine the a-th cost value of the target training cost function based on the a-th estimated confidence level and the a-th sample confidence level.

[0118] Step A126: Under the condition that the a-th cost value of the target training cost function meets the previously defined training termination requirement, terminate the training, and use the adaptive inspection decision model at the time of terminating the training as the target adaptive inspection decision model. Under the condition that the a-th cost value of the target training cost function does not meet the previously defined training termination requirement, update the neuron weight coefficients of the adaptive inspection decision model generated in the (a - 1)-th model optimization stage to obtain the adaptive inspection decision model generated in the a-th model optimization stage.

[0119] In this embodiment, during the entire model training process, in order to enable the adaptive inspection decision model to accurately evaluate and make decisions on UAV inspection tasks, it is necessary to gradually optimize the model according to the sample learning data sequence. Here, it is assumed that B = 3, and the model optimization of the a-th model optimization stage is started, where a is not less than 2.

[0120] First, perform the model optimization phase with a = 2. The server loads the second sample learning data among the B sample learning data into the adaptive inspection decision model generated in the first model optimization phase. This second sample learning data corresponds to the second sample confidence among the B sample confidences. The second sample learning data includes Xa pieces of exemplary UAV cluster description information, the second exemplary cluster description information, and the second exemplary flight inspection task description information, where Xa = X, and assume X = 5.

[0121] In the adaptive inspection decision model generated in the first model optimization phase, start processing the data. First, encode and represent these 5 sets of exemplary UAV description information to generate the corresponding 5 exemplary UAV description encoding vectors. Take one of the exemplary UAVs as an example. This UAV has various information in different previous inspection scenarios, such as the flight altitude adaptability in the mountain inspection scenario, the ability to handle complex terrains, etc., the signal interference coping ability in the suburban area inspection scenario, and the water surface reflection processing ability in the cross-river area inspection scenario. Encode this information according to the encoding rules set by the model to form an exemplary UAV description encoding vector. Perform the same operation on the other 4 exemplary UAVs. Finally, obtain 5 exemplary UAV description encoding vectors, and these encoding vectors correspond to the 5 exemplary UAVs in the second sample learning data respectively.

[0122] Next, encode and represent the second exemplary cluster description information to generate the corresponding second exemplary cluster description encoding vector. This exemplary cluster description information contains various information related to the performance of the exemplary UAV cluster in multiple inspection scenarios. For example, in the mountain inspection scenario, information such as the collaborative flight altitude control ability of the UAVs in the cluster and the overall stability when dealing with strong winds, in the suburban area inspection scenario, the communication coordination ability of the UAVs in the cluster to cope with a large number of building interferences, and in the cross-river area inspection scenario, the collaborative management ability of the UAVs in the cluster for equipment protection in a humid environment. Encode this information into an exemplary cluster description encoding vector according to the encoding rules.

[0123] At the same time, encode and represent the second exemplary flight inspection task description information to generate the corresponding second exemplary flight inspection task description encoding vector. If this exemplary flight inspection task is in the mountain inspection scenario, then the task description information may include the detailed terrain structure of the mountains in the task area, such as valley depth, peak height distribution, etc., the specific layout of the power lines to be inspected, such as line direction, power tower spacing, etc., and information related to the time limit of the task, such as whether it is necessary to complete the inspection of a specific area within a specific time period. Encode this information into an exemplary flight inspection task description encoding vector.

[0124] Then, in the adaptive inspection decision model generated in the first model optimization stage, perform correlation analysis on these 5 example UAV description encoding vectors and the second example cluster description encoding vector respectively to generate 5 example UAV correlation encoding vectors. Taking one of the example UAVs as an example, perform correlation analysis on its example UAV description encoding vector and the example cluster description encoding vector. First, decompose the example UAV description encoding vector into different feature groups, such as the flight ability related feature group (including flight altitude adaptability, speed control ability, etc.), the device status related feature group (including device health status, battery power, etc.), and the task history related feature group (including the number of successful times in previous inspection tasks, the handling results of encountering special situations, etc.). At the same time, the example cluster description encoding vector can also be regarded as containing features related to the flight ability of the cluster (such as the overall flight altitude coordination of the cluster, speed adjustment ability, etc.), features related to the device status of the cluster (such as the average health status of the devices in the cluster), and features related to the task history of the cluster (such as the overall performance of the cluster in previous inspection tasks).

[0125] Based on semantic understanding and logical relationships, perform preliminary correlation analysis on the flight ability related feature group of the example UAV and the flight ability related features of the example cluster. For example, if the flight altitude adaptability of the example UAV is good, and the example cluster has a specific flight altitude coordination strategy in the mountain inspection scenario, judge their degree of association. If the degree of association is high, record the basis for the association (such as the altitude adaptability of the UAV helps to meet the flight altitude coordination strategy of the cluster) and the initially judged degree of association, and perform in-depth interaction operations to generate the encoded segment after in-depth interaction. If the degree of association is medium, perform partial interaction to obtain the encoded segment after partial interaction. If the degree of association is low, perform identification recording to obtain identification record data. Perform correlation analysis on the device status related feature group and the task history related feature group in the same way. Finally, recombine the encoded segments after in-depth interaction and partial interaction in a set logical order (such as in the order of importance for the success of the inspection task), and append the identification record data in a predefined encoding form after the recombined encoded segments to obtain the example UAV correlation encoding vector of this example UAV. Operate on the other 4 example UAVs in the same way, and finally generate 5 example UAV correlation encoding vectors, which respectively correspond to the 5 example UAVs in the second sample learning data.

[0126] Next, in the adaptive inspection decision model generated in the first model optimization phase, based on these 5 example drone association coding vectors, the second example cluster description coding vector, and the second example flight inspection task description coding vector, determine the second estimated confidence level. This process is like the model comprehensively considering the relationship between the example drone individuals and the cluster as well as the task requirements, and then evaluating the possibility of the example drone cluster successfully completing the task in this example flight inspection task. For example, the model will consider factors such as how the capabilities of the individual drones in the example drone association coding vector are exerted in the example cluster, and the matching degree between the overall performance of the example cluster description coding vector and the task requirements in the example flight inspection task description coding vector. Through complex internal calculations, the second estimated confidence level is obtained.

[0127] Based on the second estimated confidence level and the second sample confidence level, determine the second cost value of the target training cost function. This cost value reflects the gap between the estimated confidence level obtained by the model and the known sample confidence level. If the second sample confidence level represents the true confidence level of the example drone cluster successfully completing the task in this example flight inspection task obtained through actual experience or accurate evaluation, then this cost value measures the deviation degree of the current evaluation result of the model from the actual situation.

[0128] If the second cost value of the target training cost function meets the pre-defined training termination requirement (for example, the training termination requirement is that the cost value is less than 0.1), then the server will terminate the training and use the adaptive inspection decision model at this time as the target adaptive inspection decision model. This target adaptive inspection decision model can be used for actual drone inspection task decision-making. However, if the second cost value does not meet the training termination requirement, the server will update the neuron weight coefficients of the adaptive inspection decision model generated in the first model optimization phase to obtain the adaptive inspection decision model generated in the second model optimization phase. The process of updating the neuron weight coefficients is like adjusting the internal evaluation logic of the model to enable the model to more accurately evaluate the success probability of the example drone cluster in the example flight inspection task.

[0129] Next, enter the model optimization stage with a = 3. The process is similar to that with a = 2. The server loads the 3rd sample learning data into the adaptive inspection decision model generated in the 2nd model optimization stage, and then encodes the example UAV description information set, the example cluster description information, and the example flight inspection task description information. Then, it conducts the correlation analysis between the example UAV description encoding vector and the example cluster description encoding vector to determine the 3rd estimated confidence level. Based on the 3rd estimated confidence level and the 3rd sample confidence level, it determines the 3rd cost value of the target training cost function. Finally, it decides whether to terminate the training or continue to update the neuron weight coefficients of the model according to whether this cost value meets the training termination requirement, so as to obtain a more accurate adaptive inspection decision model. This gradually optimized process continuously adjusts the parameters of the model to enable it to better adapt to different example learning data, and finally obtains a target adaptive inspection decision model that can accurately evaluate the UAV inspection task.

[0130] In a possible implementation manner, step S120 includes:

[0131] Loading the X UAV description information sets, the cluster description information, and the flight inspection task description information into the target adaptive inspection decision model of the pre-knowledge learning to generate the first estimated confidence level generated by the target adaptive inspection decision model. The target estimated confidence level includes the first estimated confidence level, and the first estimated confidence level represents the confidence level that a predefined UAV cluster among the Z UAV clusters successfully completes the low-altitude flight inspection task in the one-time flight inspection task under the condition that the Z UAV clusters are used as the UAV clusters scheduled to the one-time flight inspection task. Or

[0132] Loading the X UAV description information sets, the cluster description information, and the flight inspection task description information into the target adaptive inspection decision model of the pre-knowledge learning to generate the second estimated confidence level generated by the target adaptive inspection decision model. The target estimated confidence level includes the second estimated confidence level, and the second estimated confidence level represents the confidence level that each UAV cluster among the Z UAV clusters successfully completes the low-altitude flight inspection task in the one-time flight inspection task under the condition that the Z UAV clusters are used as the UAV clusters scheduled to the one-time flight inspection task.

[0133] In this embodiment, in the low-altitude inspection operation scenario of the power line, after the server obtains the X UAV description information sets, the cluster description information, and the flight inspection task description information, it needs to load these information into the target adaptive inspection decision model of the pre-knowledge learning to generate the target estimated confidence level. Here, the target estimated confidence level includes two cases, namely the first estimated confidence level and the second estimated confidence level.

[0134] First, consider the case of generating the first estimated confidence level. Suppose X = 5 drones are dispatched to a flight inspection task in the target inspection execution scenario (such as a mountain inspection scenario) of this low-altitude power line inspection operation, and these 5 drones are divided into Z = 3 drone clusters according to the current cluster division strategy. The server loads the X drone description information sets, cluster description information, and flight inspection task description information into the target adaptive inspection decision model.

[0135] For each drone's description information set, it contains information about each drone in multiple inspection execution scenarios (such as mountain inspection scenarios, suburban inspection scenarios, cross-river area inspection scenarios, etc.). Taking one of the drones as an example, in the mountain inspection scenario, its information may include the flight trajectory in the complex terrain of the mountains during previous inspections, the equipment status during flight (such as battery power consumption, equipment temperature, etc.), and the historical records of dealing with special situations in the mountains (such as strong wind interference). In the suburban inspection scenario, there may be information such as route adjustment during flight in densely built-up areas and the degree of signal interference. In the cross-river area inspection scenario, there are relevant information such as dealing with the problem of water surface reflection imaging and the impact of a humid environment on the equipment. The cluster description information contains the description vectors of these 3 drone clusters in multiple inspection scenarios, such as the collaborative flight ability between drones within a cluster in the mountain inspection scenario, the overall stability in dealing with strong winds, etc.; the communication and cooperation ability of drones within the cluster in dealing with signal interference in the suburban inspection scenario; the overall performance of drones within the cluster in adapting to a humid environment in the cross-river area inspection scenario. The flight inspection task description information contains the relevant information of the target inspection execution scenario (mountain inspection scenario), such as the topographical and geomorphic features of the mountains (mountain range orientation, altitude change, etc.), the distribution of power lines and power towers to be inspected, and the time limit of this inspection task (such as to be completed between 8 am and 5 pm).

[0136] After these information are loaded into the target adaptive inspection decision-making model, the model will comprehensively process this information to generate the first estimated confidence level. This first estimated confidence level represents the confidence level that, under the condition of scheduling these 3 UAV clusters as the respective UAV clusters for this flight inspection task, a predefined UAV cluster among these 3 UAV clusters successfully completes the low-altitude flight inspection task in this flight inspection task. For example, assume that a predefined UAV cluster is a cluster containing 2 UAVs. When the model analyzes, it will focus on the matching degree between the ability information of the UAV individuals in this cluster and the requirements of the flight inspection task, as well as the collaborative relationship between this cluster and other clusters in the overall inspection task (although here we focus on this specific cluster). For the UAV individuals in this cluster, the model will consider whether their flight capabilities (such as flight altitude adaptability, speed stability, etc.) in the mountain inspection scenario can meet the current mountain terrain and task requirements (such as whether they can accurately inspect the power line within the specified flight altitude range, whether they can maintain a stable flight speed under complex terrain, etc.). At the same time, the equipment status will also be considered, for example, whether the battery power is sufficient to support the entire inspection task process, whether the equipment temperature is within the normal range to avoid affecting the equipment performance, etc. From the perspective of the cluster, the model will evaluate the collaborative ability of this cluster in the mountain inspection scenario, such as whether the flight spacing between these 2 UAVs is maintained reasonably, and whether the task division and cooperation (such as one is responsible for the main inspection, and the other is responsible for assistance and data backup, etc.) are reasonable and effective. By comprehensively considering these factors and the adaptability of the cluster to the entire task scenario (including the collaborative relationship with other clusters, etc.), the model obtains the first estimated confidence level that this specific cluster successfully completes the task in this flight inspection task through internal complex calculations and analyses.

[0137] Next, look at the situation of generating the second estimated confidence level. Similarly, the X UAV description information sets, cluster description information, and flight inspection task description information are loaded into the target adaptive inspection decision-making model with prior knowledge learning. The model will still conduct a comprehensive and in-depth analysis of this information. At this time, the generated second estimated confidence level represents the confidence level that, under the condition of scheduling these 3 UAV clusters as the respective UAV clusters for this flight inspection task, each UAV cluster among these 3 UAV clusters successfully completes the low-altitude flight inspection task in this flight inspection task.

[0138] In this process, the model will conduct a similar but more comprehensive assessment of each UAV cluster. Taking three UAV clusters as an example, for the first UAV cluster, the model will consider the matching degree between the flight capabilities, equipment status, mission history, etc. of individual UAVs within the cluster and the requirements of the flight inspection mission, just as it analyzed a specific cluster before. For example, in the mountain inspection scenario, whether the parameters such as the flight altitude and speed of the UAVs within the cluster can adapt to the mountain terrain, whether the equipment can operate stably, and whether their previous performance in mountain inspections was good. At the same time, it will also focus on analyzing the coordination relationship between this cluster and the other two clusters. For example, in the entire inspection mission, whether their flight routes cooperate with each other and whether the task division is reasonable (for example, one cluster is responsible for inspecting the eastern section of the mountain, another cluster is responsible for inspecting the western section, and the third cluster is responsible for inspecting the middle area or key areas, etc.). For the second and third UAV clusters, the model will also conduct the same comprehensive analysis, including the situation of individual UAVs within each of them and their coordination relationship with other clusters. By comprehensively considering all these factors, the model evaluates the possibility of each of these three UAV clusters successfully completing the mission in this flight inspection mission, and finally obtains the second estimated confidence level. This second estimated confidence level can comprehensively reflect the confidence level of these three UAV clusters as a whole in successfully completing the mission in this flight inspection mission, providing an important basis for subsequent decisions (such as whether to schedule the inspection mission according to the current cluster division, etc.).

[0139] Figure 2 The detection system 100 shown includes: a processor 1001 and a memory 1003. Among them, the processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the detection system 100 may further include a transceiver 1004, and the transceiver 1004 can be used for data interaction between this server and other servers, such as data sending and / or data receiving, etc. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of this detection system 100 does not constitute a limitation to the embodiments of the present application.

[0140] The processor 1001 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute various exemplary logical blocks, modules, and circuits described in connection with the disclosure of this application. The processor 1001 can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0141] The bus 1002 can include a path for transmitting information between the above components. The bus 1002 can be a PCI (Peripheral Component Interconnect) bus, an EISA (Extended Industry Standard Architecture) bus, or the like. The bus 1002 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 2 only a thick line is used to represent it in the figure, but it does not mean that there is only one bus or one type of bus.

[0142] The memory 1003 can be a ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, a RAM (Random Access Memory) or other type of dynamic storage device that can store information and instructions, or it can also be an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium that can be used to carry or store program code and can be read by a computer, which is not limited here.

[0143] The memory 1003 is used to store the program code for implementing the embodiments of the present application, and is controlled by the processor 1001 to execute. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0144] In addition, an embodiment of the present application further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the adaptive detection method for low-altitude flight inspection as described above is implemented.

[0145] Similarly, it should be noted that, in order to simplify the description of the present application disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing or description thereof. Similarly, it should be noted that, in order to simplify the description of the present application disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing or description thereof.

Claims

1. An adaptive detection method for low-altitude flight inspection, characterized in that: The method comprises: Obtain X drone description information sets, cluster description information, and flight inspection task description information, wherein the X drone description information sets include Y knowledge field description vectors of each drone in the X drones, the X drones are drones that are dispatched to a flight inspection task in a target inspection execution scenario in a target low-altitude inspection operation, the Y knowledge field description vectors of each drone include description vectors of each drone in multiple inspection execution scenarios, the multiple inspection execution scenarios are inspection execution scenarios with full authority authentication in the target low-altitude inspection operation, the cluster description information includes description vectors of each drone cluster in the Z drone clusters in the multiple inspection execution scenarios, the flight inspection task description information includes description vectors of the target inspection execution scenario, the Z drone clusters are multiple drone clusters generated by clustering the X drones according to the current cluster division strategy, and X, Y, and Z are all not less than 2; The X drone description information sets, cluster description information and flight inspection task description information are loaded into a target adaptive inspection decision model learned in advance, and a target estimation confidence generated by the target adaptive inspection decision model is generated, wherein the target estimation confidence represents the confidence that the Z drone clusters corresponding to the Z drone clusters successfully complete the low-altitude flight inspection task in the flight inspection task under the condition that the Z drone clusters are used as the respective drone clusters scheduled to the flight inspection task; Under the condition that the target estimation confidence meets the set requirements, the Z drone clusters are used as the drone clusters scheduled to the one flight inspection task to obtain the adaptive detection results before the low-altitude flight inspection; The step of loading the X drone description information sets, cluster description information, and flight inspection task description information into a target adaptive inspection decision model learned in advance, and generating a target estimation confidence generated by the target adaptive inspection decision model, comprises: Encode and represent the X drone description information sets to generate corresponding X drone description encoding vectors, encode and represent the cluster description information to generate corresponding cluster description encoding vectors, and encode and represent the flight inspection task description information to generate corresponding flight inspection task description encoding vectors, wherein the X drone description encoding vectors correspond to the X drones respectively; Performing correlation analysis on the X drone description coding vectors and the flight inspection task description coding vectors to generate X drone association coding vectors, where the X drone association coding vectors correspond to the X drones respectively; The target estimation confidence is determined based on the X drone association coding vectors, the cluster description coding vector and the flight inspection task description coding vector.

2. The adaptive detection method for low-altitude flight inspection according to claim 1 is characterized in that: The step of performing correlation analysis on the X drone description coding vectors and the flight inspection task description coding vectors to generate X drone association coding vectors includes: Based on the following steps, the ath drone description coding vector among the X drone description coding vectors is analyzed for correlation with the flight inspection task description coding vector to generate the ath drone association coding vector among the X drone association coding vectors, where a is a positive integer not less than 1 and not greater than X: Interact the a-th UAV description coding vector with the flight inspection task description coding vector to generate an a-th blending coding vector; The aggregation result of the a-th UAV description coding vector and the a-th blending coding vector is used as the a-th UAV associated coding vector; or, the a-th blending coding vector is fused with the a-th influencing factor defined previously to generate the a-th fused coding vector, and the aggregation result of the a-th UAV description coding vector and the a-th fused coding vector is used as the a-th UAV associated coding vector; The step of interacting the a-th UAV description coding vector with the flight inspection task description coding vector to generate the a-th blended coding vector includes: Decomposing the a-th drone description coding vector into a plurality of different first feature groups, wherein the first feature groups are divided based on different attributes of the drone, specifically including a flight capability-related feature group, a device status-related feature group, and a mission history-related feature group; And, decomposing the flight inspection task description coding vector into a plurality of second feature groups related to different task scenario requirements, specifically including a task area geographical feature group, a task target related feature group and a task time limit related feature group; For each first feature group, based on semantic understanding and logical relationship, preliminary association analysis is performed one by one with all second feature groups, and an association relationship description document is established for each preliminary association analysis result, wherein the association relationship description document is used to record the feature group pairs formed by the associated first feature group and the second feature group, as well as the basis for the association and the preliminary determined degree of association; According to the association relationship description document, a deep interaction operation is performed on a feature pair with a high association degree to generate a coded segment after the deep interaction, and a partial interaction is performed on a feature pair with a medium association degree to generate a coded segment after the partial interaction, and an identification record is performed on a feature pair with a low association degree to generate identification record data; The coding segments after deep interaction and partial interaction are recombined according to a set logical order, where the logical order is determined based on the task execution process or the importance of the data, and the identification record data is attached to the recombined coding segments in a predefined coding form to generate the ath interleaved coding vector.

3. The adaptive detection method for low-altitude flight inspection according to claim 1 is characterized in that: The determining the target estimation confidence based on the X drone association coding vectors, the cluster description coding vector and the flight inspection task description coding vector includes: Aggregate the X drone association coding vectors, the cluster description coding vector, and the flight inspection task description coding vector to generate an aggregated coding vector; The aggregated coding vector is subjected to full connection mapping processing based on a full connection mapping layer to generate the target estimation confidence.

4. The adaptive detection method for low-altitude flight inspection according to claim 1 is characterized in that: The step of obtaining X sets of drone description information includes: Acquire flight trajectory description information, immediate state description information, and long-term state description information of each drone among the X drones, and generate a description information set of the X drones, wherein the flight trajectory description information of each drone includes the flight trajectory information of each drone in the previous b flight inspection tasks, the long-term state description information of each drone includes the record information of each drone in each of the multiple inspection execution scenarios within a previous set time period, and the immediate state description information of each drone includes the current record information of each drone in each of the multiple inspection execution scenarios; And, get the cluster description information, including: Obtain average performance statistics of each drone cluster in the Z drone clusters in each inspection execution scenario in the multiple inspection execution scenarios; And obtain the flight inspection task description information, including: Obtain the scenario label index of the target inspection execution scenario, or obtain the scenario label index of the target inspection execution scenario and the timing node scheduled to the one-time flight inspection task.

5. The adaptive detection method for low-altitude flight inspection according to claim 1 is characterized in that: The method of dispatching the Z drone clusters as the respective drone clusters in the one flight inspection task under the condition that the target estimation confidence meets the set requirements includes: Under the condition that the deviation between the target estimation confidence and the threshold confidence is less than a set value, the Z drone clusters are used as the respective drone clusters scheduled to the one flight inspection mission; or Under the condition that the deviation between the target estimated confidence and the threshold confidence in the acquired A estimated confidences is minimal, the Z drone clusters are used as the drone clusters scheduled to the flight inspection task, the A estimated confidences include the estimated confidences determined by the target adaptive inspection decision model based on the drone description information, cluster description information and flight inspection task description information of each drone cluster division data in the A drone cluster division data, each drone cluster division data in the A drone cluster division data includes a plurality of drone clusters generated by clustering the X drones according to a corresponding cluster division strategy among A different cluster division strategies, the A different cluster division strategies include the current cluster division strategy, the number of clusters of the plurality of drone clusters included in each drone cluster division data is the same, and A is not less than 2; or Under the condition that the deviation between the target estimated confidence and the threshold confidence among the A estimated confidences is the smallest and the deviation is less than the set value, the Z drone clusters are dispatched as the respective drone clusters in the one flight inspection mission.

6. The adaptive detection method for low-altitude flight inspection according to any one of claims 1 to 5, characterized in that: The method further comprises: Obtain a sample learning data sequence and a sample confidence sequence, wherein the sample learning data sequence includes B sample learning data, each sample learning data includes X sample drone cluster description information, sample cluster description information, and sample flight inspection task description information, the X sample drone cluster description information includes Y sample knowledge field description vectors of each sample drone in the X sample drones, the X sample drones are drones that are dispatched to a sample flight inspection task in the sample inspection execution scenario in the target low-altitude inspection operation, and the Y sample knowledge field description vectors of each sample drone The quantity includes description vectors of each example drone in the multiple inspection execution scenarios, the example cluster description information includes description vectors of each example drone cluster in the Z example drone clusters in the multiple inspection execution scenarios, the example flight inspection task description information includes description vectors of the example inspection execution scenarios, the Z example drone clusters are multiple example drone clusters generated by clustering the X example drones according to the example cluster division strategy, the sample confidence sequence includes B sample confidences corresponding to the B sample learning data, and B is not less than 2; Model parameter learning is performed on the initialized adaptive patrol decision model according to the sample learning data sequence until the cost value of the target training cost function corresponding to the adaptive patrol decision model meets the previously defined training termination requirement, the training is terminated, and the adaptive patrol decision model at the time of the termination of the training is used as the target adaptive patrol decision model. Under the condition that the cost value of the target training cost function corresponding to the adaptive patrol decision model does not meet the previously defined training termination requirement, the neuron weight coefficient in the adaptive patrol decision model is updated; the cost value of the target training cost function is a loss value determined based on the example estimated confidence generated by the adaptive patrol decision model and the corresponding sample confidence of the B sample confidences, and the example estimated confidence is an estimated confidence determined by the adaptive patrol decision model based on the X example drone cluster description information, the example cluster description information and the example flight inspection task description information.

7. The adaptive detection method for low-altitude flight inspection according to claim 6 is characterized in that: The performing model parameter learning on the initialized adaptive inspection decision model according to the sample learning data sequence includes: Perform model optimization for the a-th model optimization stage based on the following steps, where a is not less than 2: Loading the ath sample learning data among the B sample learning data into the adaptive inspection decision model generated in the a-1th model optimization stage, wherein the ath sample learning data corresponds to the ath sample confidence among the B sample confidences, and the ath sample learning data includes Xa example drone cluster description information, the ath example cluster description information and the ath example flight inspection task description information, Xa=X; In the adaptive inspection decision model generated in the a-1th model optimization stage, the Xa example drone description information sets are encoded to generate corresponding Xa example drone description encoding vectors, the ath example cluster description information is encoded to generate the corresponding ath example cluster description encoding vector, and the ath example flight inspection task description information is encoded to generate the corresponding ath example flight inspection task description encoding vector, and the Xa example drone description encoding vectors correspond to the Xa example drones in the ath example learning data, respectively; In the adaptive inspection decision model generated in the a-1th model optimization stage, the Xa example drone description coding vectors are respectively analyzed for correlation with the ath example cluster description coding vector to generate Xa example drone association coding vectors, and the Xa example drone association coding vectors correspond to the Xa example drones in the ath sample learning data respectively; In the adaptive inspection decision model generated in the a-1th model optimization stage, determining an ath estimated confidence based on the Xa example drone association coding vectors, the ath example cluster description coding vector, and the ath example flight inspection task description coding vector; Determining an ath cost value of the target training cost function based on the ath estimated confidence and the ath sample confidence; Under the condition that the ath cost value of the target training cost function meets the previously defined training termination requirements, the training is terminated, and the adaptive patrol decision model at the time of termination of the training is used as the target adaptive patrol decision model; under the condition that the ath cost value of the target training cost function does not meet the previously defined training termination requirements, the neuron weight coefficients of the adaptive patrol decision model generated in the a-1th model optimization stage are updated to obtain the adaptive patrol decision model generated in the ath model optimization stage.

8. The adaptive detection method for low-altitude flight inspection according to any one of claims 1 to 5, characterized in that: The step of loading the X drone description information sets, cluster description information, and flight inspection task description information into a target adaptive inspection decision model learned in advance, and generating a target estimation confidence generated by the target adaptive inspection decision model, comprises: Loading the X drone description information sets, cluster description information, and flight inspection task description information into the target adaptive inspection decision model learned in advance, generating a first estimated confidence generated by the target adaptive inspection decision model, wherein the target estimated confidence includes the first estimated confidence, and the first estimated confidence represents the confidence that a previously defined drone cluster among the Z drone clusters successfully completes the low-altitude flight inspection task in the flight inspection task under the condition that the Z drone clusters are used as the drone clusters scheduled to the flight inspection task; or The X drone description information sets, cluster description information and flight inspection task description information are loaded into the target adaptive inspection decision model learned in advance, and a second estimated confidence generated by the target adaptive inspection decision model is generated, wherein the target estimated confidence includes the second estimated confidence, and the second estimated confidence represents the confidence that each drone cluster in the Z drone clusters successfully completes the low-altitude flight inspection task in the flight inspection task under the condition that the Z drone clusters are used as each drone cluster scheduled to the flight inspection task.

9. A detection system, characterized in that: The detection system includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the adaptive detection method for low-altitude flight inspection as described in any one of claims 1-8.

Images