Low-altitude operation multi-aircraft cooperative scheduling method and system

Abstract

The application discloses a low-altitude operation multi-aircraft cooperative scheduling method and system, and the method comprises the following steps: dividing a to-be-inspected object into multiple inspection sub-tasks, setting an initial priority and distributing the initial priority to multiple aircrafts; establishing a space correlation graph according to the physical structure relationship and engineering attributes among the inspection targets. After the aircrafts complete the inspection sub-tasks, the inspection data is subjected to abnormality detection, and the priority increment of the unfinished tasks is calculated in combination with the abnormality score, the correlation strength among the targets and the spatial distance; when the same unfinished task receives multiple priority increments, the maximum value is taken, and the maximum value is combined with the initial priority value to synthesize an updated priority value. The unfinished tasks are reordered according to the updated priority value, and when the sorting change exceeds a rescheduling trigger threshold, the task is redistributed in combination with the current state of each aircraft. The application can dynamically adjust the subsequent task sequence according to the inspection abnormality, and improve the priority coverage capability of high-risk targets and the cooperative operation efficiency of multiple aircrafts.

Description

Technical Field

[0001] This invention relates to the field of low-altitude aircraft collaborative operation and mission scheduling technology, and in particular to a method and system for low-altitude multi-aircraft collaborative scheduling. Background Technology

[0002] With the development of low-altitude aircraft and airborne sensing technology, using multiple aircraft to collaboratively perform inspections of objects such as power transmission lines, bridges, and building facades has become an important way to improve the efficiency of large-scale inspections. Existing multi-aircraft scheduling methods typically generate a task allocation scheme and execution queue before the operation begins, based on the location of the inspection target, preset task priorities, aircraft endurance, and sensor configuration, and plan the path with the goal of minimizing total flight distance, task completion time, or energy consumption.

[0003] However, low-altitude inspection missions are inherently dynamic. During operation, aircraft may discover anomalies such as cracks, deterioration, and corrosion. Other inspection targets with physical connections, spatial proximity, or from the same installation batch often pose a high risk of association. Existing scheduling methods typically use inspection results for post-operation defect analysis, failing to promptly adjust the priority of unfinished tasks based on discovered anomalies. This can easily lead to high-risk targets being inspected as originally planned, and when operational windows are limited by weather, airspace control, or flight time, potential hazards requiring priority inspection may be overlooked.

[0004] Furthermore, some dynamic scheduling methods primarily rely on the degree of anomaly or flight distance of individual tasks to directly switch tasks, lacking comprehensive consideration of the structural relationships, engineering attributes, and spatial distances between inspected targets. This makes it difficult to accurately determine the impact range of abnormal information on other targets. When abnormal results occur continuously, task reallocation may be frequently triggered, causing aircraft to repeatedly change course, increasing flight energy consumption and scheduling overhead. Therefore, how to utilize the abnormal information obtained during inspections, combined with the correlation between targets and the real-time status of aircraft, to implement reasonable and stable dynamic scheduling for unfinished tasks remains a technical problem that needs to be solved in low-altitude multi-aircraft collaborative operations. Summary of the Invention

[0005] This invention provides a method and system for coordinated scheduling of multiple aircraft for low-altitude operations, in order to at least solve some of the technical problems existing in the background art as described above.

[0006] According to a first aspect of the present invention, a method for coordinated scheduling of multiple aircraft for low-altitude operations is provided, comprising the following steps: The objects to be inspected within the work area are divided into multiple inspection sub-tasks, each inspection sub-task is assigned an initial priority value, and each inspection sub-task is assigned to multiple aircraft. Based on the physical structural relationships and engineering attributes between each inspection target, a spatial association graph of the inspection targets is established. Each node in the spatial association graph corresponds to an inspection sub-task, and the edge weight between nodes is the association strength. After the aircraft completes the inspection sub-task, it performs anomaly detection on the collected inspection data and outputs the anomaly score of the completed task; based on the anomaly score, the correlation strength, and the spatial distance between the completed and incomplete tasks, it calculates the priority increment of each incomplete task. When multiple completed tasks propagate priority increments to the same incomplete task, the maximum value among the priority increments is taken; based on the maximum value and the initial priority value of the incomplete task, the updated priority value is obtained. All incomplete tasks are reordered according to the updated priority value. When the change in order exceeds the rescheduling trigger threshold, the incomplete tasks are reallocated according to the current state of each aircraft and the updated priority value.

[0007] Preferably, the correlation strength is calculated based on the structural correlation coefficient and the process batch similarity coefficient; the structural correlation coefficient is determined according to the physical positional relationship and structural connection relationship between the two inspection targets, and the process batch similarity coefficient is determined according to the installation batch information of the two inspection targets; the values ​​of the structural correlation coefficient and the process batch similarity coefficient are both between 0 and 1.

[0008] Preferably, the method for determining the structural correlation coefficient includes: statistically analyzing the conditional probability of two inspection targets having defects simultaneously from historical inspection records, and using the conditional probability as the value of the structural correlation coefficient; the method for determining the process batch similarity coefficient includes: querying the installation batch information of two inspection targets from the equipment ledger, statistically analyzing the conditional probability of common defects appearing in the same batch of equipment from historical records, and using the conditional probability as the value of the process batch similarity coefficient.

[0009] Preferably, when calculating the priority increment, a distance attenuation coefficient is applied based on the spatial distance between completed and uncompleted tasks; the distance attenuation coefficient decreases as the spatial distance increases, and adopts a Gaussian attenuation form; the attenuation radius parameter is determined based on the spatially relevant scale of the inspection object, and the spatially relevant scale is calibrated according to the distance corresponding to the conditional probability of defect co-occurrence attenuating to a preset lower limit value.

[0010] Preferably, the reallocation of unfinished tasks includes: constructing a weighted bipartite graph with each aircraft and each unfinished task, wherein the allocation cost from each aircraft to each unfinished task is the ratio of the flight time required for the aircraft to reach the task to the updated priority value of the task; using the Hungarian algorithm to solve the minimum weight matching of the weighted bipartite graph to obtain a reallocation scheme; and after the aircraft completes the currently executing inspection subtask, switching to the next inspection subtask specified by the reallocation scheme.

[0011] Preferably, the anomaly detection employs a lightweight convolutional neural network, which uses MobileNetV2 as its backbone network. The original classification layer at the end of MobileNetV2 is replaced with a fully connected layer containing a single neuron. The activation function of this fully connected layer is the sigmoid function, which directly outputs a continuous anomaly score. The image collected by the inspection sensor is used as input, and the output is a continuous scalar value between 0 and 1 as the anomaly score.

[0012] Preferably, the training data for the lightweight convolutional neural network comes from historical inspection images, and each image is labeled with a severity score according to the defect grading standard; during training, the SmoothL1 loss function is used, and the loss value is calculated based on the deviation between the model's predicted score and the labeled score; the optimizer is the Adam optimizer, and the learning rate adjustment strategy is cosine annealing.

[0013] Preferably, after receiving the redistribution scheme, if the time interval since the last task switch is less than the preset minimum execution interval, the switch is temporarily suspended, and the next inspection sub-task specified in the redistribution scheme is executed after the current en route inspection sub-task is completed; the minimum execution interval is determined based on the average execution time of a single inspection sub-task.

[0014] Preferably, when solving the minimum weight matching of the weighted bipartite graph, an energy feasibility constraint is imposed on each aircraft: the total expected energy consumption of all unfinished tasks allocated to the same aircraft does not exceed a preset safe proportion of the aircraft's current remaining energy.

[0015] According to a second aspect of the present invention, a low-altitude multi-aircraft cooperative scheduling system is provided, comprising: The task initialization module is used to divide the objects to be inspected in the work area into multiple inspection sub-tasks, assign an initial priority value to each inspection sub-task, and assign each inspection sub-task to multiple aircraft. The spatial association graph establishment module is used to establish a spatial association graph of inspection targets based on the physical structure relationship and engineering attributes between each inspection target. Each node in the spatial association graph corresponds to an inspection sub-task, and the edge weight between nodes is the association strength. The anomaly detection and priority increment calculation module is used to detect anomalies in the collected inspection data after the aircraft completes the inspection sub-task, and output the anomaly score of the completed task; and calculate the priority increment of each uncompleted task based on the anomaly score, the correlation strength, and the spatial distance between the completed and uncompleted tasks. The priority update module is used to take the maximum value among the priority increments when multiple completed tasks propagate priority increments to the same unfinished task; and to obtain the updated priority value based on the maximum value and the initial priority value of the unfinished task. The task reassignment module is used to reorder all incomplete tasks according to the updated priority value. When the change in order exceeds the rescheduling trigger threshold, the incomplete tasks are reassigned according to the current state of each aircraft and the updated priority value.

[0016] This invention establishes a spatial correlation diagram reflecting the physical structural relationships and engineering attributes of inspection targets. It incorporates anomaly detection results from completed missions into the online scheduling process of incomplete missions. Priority increments are calculated by combining the correlation strength between targets and the spatial distance attenuation relationship. This allows local inspection anomalies to influence the subsequent inspection order of related targets according to their actual correlation, avoiding delays in high-risk target investigation due to relying solely on preset priorities. For multiple priority increments received for the same incomplete mission, the maximum value is used for updating, suppressing artificially high priorities caused by the superposition of multiple weakly correlated anomalies and improving the rationality of risk propagation results. Furthermore, reassignment is triggered only when mission order changes meet set conditions, and subsequent missions are determined by comprehensively considering aircraft arrival time, mission priority, remaining battery power, and mission switching intervals. This enables timely response to anomaly risks while reducing ineffective rescheduling and frequent heading adjustments. Therefore, it improves the priority coverage of high-risk inspection targets within a limited operational window, reduces the possibility of overlooking critical hidden dangers, and balances the endurance safety of multiple aircraft with overall operational efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of a low-altitude multi-aircraft collaborative scheduling method provided in an embodiment of the present invention.

[0019] Figure 2 A flowchart illustrating the reassignment of unfinished tasks provided in this embodiment of the disclosure.

[0020] Figure 3 This is a structural block diagram of a low-altitude multi-aircraft collaborative scheduling system provided in an embodiment of the present disclosure.

[0021] Figure 4 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] According to embodiments of this disclosure, a multi-aircraft collaborative scheduling method for low-altitude operations is provided. This method is applicable to scenarios where multiple aircraft collaboratively perform inspection-type operations in low-altitude airspace. Typical applications include power transmission line inspection, bridge structure inspection, and building facade inspection. The operation area contains a large number of target points to be inspected. A single aircraft cannot complete all tasks within a single operation window, requiring multiple aircraft to collaborate. Each aircraft is equipped with inspection sensors and an onboard edge computing unit, enabling preliminary analysis of collected data during flight. Simultaneously, it maintains communication with the ground dispatch station via a wireless data link, allowing for real-time reporting of flight status and task execution results. The operation window is constrained by weather conditions or airspace control, and its duration is limited, for example, it can be configured to be 2 to 6 hours. Within this limited operation window, how to dynamically adjust the priority of subsequent tasks based on the abnormal information gradually acquired during the inspection process, ensuring that high-risk targets are prioritized for inspection, is the key problem addressed by this method.

[0024] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0025] This invention provides a method for coordinated scheduling of multiple aircraft for low-altitude operations, such as... Figure 1 As shown, the method includes the following steps: S1 divides the objects to be inspected in the work area into multiple inspection sub-tasks, assigns an initial priority value to each inspection sub-task, and assigns each inspection sub-task to multiple aircraft. S2. Based on the physical structure relationship and engineering attributes between each inspection target, establish an inspection target spatial association graph. Each node in the spatial association graph corresponds to an inspection sub-task, and the edge weight between nodes is the association strength. S3. After the aircraft completes the inspection sub-task, it performs anomaly detection on the collected inspection data and outputs the anomaly score of the completed task. Based on the anomaly score, the correlation strength, and the spatial distance between the completed and uncompleted tasks, it calculates the priority increment of each uncompleted task. S4, when multiple completed tasks propagate priority increments to the same unfinished task, take the maximum value among the priority increments; based on the maximum value and the initial priority value of the unfinished task, obtain the updated priority value; S5, reorder all unfinished tasks according to the updated priority value. When the change in order exceeds the rescheduling trigger threshold, the unfinished tasks are reallocated according to the current state of each aircraft and the updated priority value.

[0026] In some embodiments, for step S1, before the operation begins, the scheduling system divides the objects to be inspected within the entire operation area into multiple inspection sub-tasks. Each inspection sub-task corresponds to an inspection target point or a continuous inspection section, and has clear geographical coordinates, estimated operation time, and required sensor types. The scheduling system assigns an initial priority value to each inspection sub-task, which reflects the prior importance of the task. Specifically, the initial priority value is a weighted sum of the following three items: equipment level coefficient (valued according to the equipment importance classification in industry standards, for example, 1.0 for level 1 equipment, 0.7 for level 2, and 0.4 for level 3 equipment), inspection interval coefficient (number of days since the last inspection divided by the standard inspection cycle days, with an upper limit of 1.0), and regional historical failure rate (number of failures in the region over the past few years divided by the number of inspections). The weights of the three items are 0.4, 0.3, and 0.3, respectively. This combination is determined based on operation and maintenance management experience and can be adjusted according to the operation and maintenance procedures of different operation objects.

[0027] After initial priority values ​​are assigned, the scheduling system allocates each inspection subtask to multiple aircraft based on their takeoff and landing locations, maximum endurance, and sensor configurations, using either a greedy allocation strategy or an integer linear programming method. This generates an ordered execution queue containing several inspection subtasks for each aircraft. The greedy allocation strategy assigns tasks to the nearest aircraft with sufficient remaining endurance, in descending order of priority. The integer linear programming method, on the other hand, aims to minimize the total flight distance or total completion time, solving for the globally optimal allocation scheme under aircraft endurance constraints.

[0028] In some embodiments, for step S2, before task initiation, the scheduling system establishes a spatial association graph of the inspection targets based on the physical structural relationships and engineering attributes between each inspection target. Each node in the spatial association graph corresponds to an inspection subtask, and the edges between nodes indicate that two inspection targets are related at the physical or engineering level; the weight of the edge is called the association strength, denoted as . ,in and These are the numbers of the two inspection targets. The range of values ​​is Dimensionless. The meaning of correlation strength is that when an anomaly is detected in one target, the greater the likelihood that another target with a higher correlation strength will exhibit the same anomaly. The correlation strength between two inspection targets is lower than a preset boundary threshold. When the value is 0.05 (e.g., 0.05), no edge connections are established to reduce the number of redundant edges in the graph.

[0029] For example, constructing edge thresholds Based on the defect co-occurrence statistics of historical inspection records, the empirical correlation strength between all target pairs is calculated, and the cumulative distribution function (CDF) curve of the correlation strength is plotted. The correlation strength quantile corresponding to a cumulative probability of 90% in the CDF curve is taken as... The value of is determined by retaining the 10% of target pairs with the strongest correlation as edge construction objects, and disconnecting the remaining weakly correlated edges. If historical data is lacking, The default value is 0.10. In small-scale scenarios where the number of targets in the work area is less than 50, it can be... Lowered to 0.05 to retain more related information; in large-scale scenarios with more than 500 targets, it can be... Increased to 0.15 to control graph density. The range of values ​​is .

[0030] Among them, correlation strength Based on structural correlation coefficient Similarity coefficient with process batch The product is determined as follows: in Reflecting the target With the goal The range of values ​​for physical structural connections or adjacencies. ; This reflects the degree to which both parties share the same batch of materials or the same construction team during the construction and installation process; the range of values ​​is [not specified]. Both are dimensionless quantities, and their product is... Both are dimensionless quantities; the reason for using the product form is that when two targets are strongly correlated in both structural location and process batch, the risk of common defects is the highest; if one of the correlations is very weak (close to 0), even if the other correlation is strong, the overall correlation strength will be reduced accordingly.

[0031] Among them, the structural correlation coefficient The determination method is as follows: Calculate the conditional probability of two inspection targets simultaneously having defects from historical inspection records, and use this conditional probability as... The value of is taken as follows. Taking transmission line inspection as an example, the inspection records of the past few years (e.g., 5 years) are statistically analyzed. Cases in which insulator deterioration is found in one phase conductor on the same tower, and the same problem also exists in another phase on the same tower are summarized, and the conditional probability is calculated: this probability is equal to the number of inspections in which defects exist in both phases at the same time divided by the number of inspections in which defects exist in either phase. If this value is approximately 0.65, then the probability of defects between different phases on the same tower is... Take 0.65; if the conditional probability of defect co-occurrence in the same circuit conductor between two adjacent towers is approximately 0.25, then Take 0.25; between targets with a spacing of two or more base towers For other scenarios such as bridge inspection, a value of 0.05 or lower is recommended. The calibration logic is the same, and statistics are performed according to the historical detection records of the corresponding scenario.

[0032] Among them, the process batch similarity coefficient The method of determination and Similarly, specifically, the installation batch information of two inspection targets is retrieved from the equipment ledger, and the conditional probability of common defects appearing in the same batch of equipment is calculated from historical records. This conditional probability is then used as... The value of is determined by the shared material sources and construction processes used in the same batch of equipment, which increases the likelihood of common defects compared to different batches. For example, if the conditional probability of all insulators in the same batch deteriorating after one insulator deteriorates is approximately 0.55, then the probability of common defects in the same batch is higher. Use 0.55; use 0.20 for different batches but from the same supplier; use 0.10 for different suppliers. When no historical statistical data is available, and Initial values ​​can be preset by domain engineers based on experience, and recalibrated after a certain number of inspection records have been accumulated. Optionally, when the inspection object ledger information is incomplete or batch data cannot be obtained, A neutral default value of 0.5 is uniformly adopted, relying solely on the structural correlation coefficient. It plays a key role in differentiating between different entities.

[0033] Once the spatial association graph is constructed, it is stored in the scheduling system and remains unchanged throughout the job. In subsequent steps, the scheduling system propagates priority increments along the edges of the graph to the associated targets.

[0034] In some embodiments, for step S3, after the aircraft completes the inspection sub-task, the onboard edge computing unit performs anomaly detection on the collected inspection data and outputs the anomaly score of the completed task. , The number of the completed task; The range of values ​​is , dimensionless, the larger the value, the more serious the abnormality, and close to 0, the less abnormality is found.

[0035] In one embodiment, anomaly detection employs a lightweight convolutional neural network (CNN), using images acquired by inspection sensors as input and outputting a continuous scalar value between 0 and 1 as the anomaly score. The lightweight network is chosen to account for the computing power and power consumption constraints of airborne edge computing platforms. Typical airborne embedded platforms have limited computing power and need to process one frame of image within tens of milliseconds of a single inference iteration.

[0036] This lightweight convolutional neural network uses MobileNetV2 as its backbone. MobileNetV2 consists of 17 inverse residual modules stacked sequentially. Each module is composed of a 1×1 pointwise convolutional layer, a 3×3 depthwise separable convolutional layer, and a 1×1 pointwise convolutional layer, with a channel expansion ratio of 6 for each module. When the number of input and output channels of a module is the same, the input is directly added to the output through residual skip connections. MobileNetV2 uses depthwise separable convolution to decompose standard convolution into two steps: depthwise convolution and pointwise convolution, significantly reducing computational cost and making it suitable for deployment on edge devices with limited computing power.

[0037] In this embodiment, the original classification layer at the end of MobileNetV2 is replaced with a fully connected layer containing a single neuron. This fully connected layer uses the sigmoid function for activation, directly outputting continuous anomaly scores. The original MobileNetV2 classification layer at the end contains a fully connected layer and a softmax activation function, outputting a discrete class probability distribution, only able to determine a binary category of "defective" or "non-defective." After the replacement, the model outputs continuous severity scores, distinguishing the degree of anomaly severity. Severe defects produce larger anomaly scores, resulting in a greater increase in the priority of adjacent targets after propagation; minor defects produce smaller anomaly scores, with a correspondingly weaker propagation impact. If binary classification is still used, the impact of severe defects and minor defects on adjacent targets cannot be distinguished, and the discriminative power of the propagation results will decrease.

[0038] The model input consists of RGB images collected by inspection sensors, which are cropped and normalized to a uniform size of 224×224 pixels. The data format is a three-dimensional tensor with dimensions of batch size × 3 × 224 × 224, where 3 corresponds to the three color channels of RGB. During image cropping, the region containing the main body of the target is extracted centered on the inspection target. Normalization linearly scales the pixel values ​​from 0 to 255 to the range of 0 to 1.

[0039] For example, a lightweight object detector (such as SSD-MobileNet) is used to locate the bounding box of the inspection target in the original image. A cropping window is then created by extending 20% ​​of the shorter side of this bounding box outwards, capturing the region containing the target and appropriate surrounding background. If the target is a linear target (such as a power transmission line), a strip-shaped region containing complete defect candidate areas is cropped along the direction of the transmission line. The cropped image is then scaled to 224×224 pixels using bicubic interpolation. The pixel values ​​are then... linear scaling to Then subtract the mean and standard deviation of the ImageNet dataset by channel (mean = Standard deviation = This ensures that the input data distribution is compatible with the pre-trained weights. The image data is organized into the NCHW format under the PyTorch framework (i.e., batch size × 3 × 224 × 224, where 3 corresponds to the three RGB color channels, and the first dimension is the channel dimension), or the NHWC format under the TensorFlow framework (batch size × 224 × 224 × 3). The specific format is determined according to the inference framework used during deployment.

[0040] The above preprocessing is performed in real time on the airborne edge computing unit, with a single frame image preprocessing time of no more than 5 milliseconds. The model output is a one-dimensional scalar. Inspection images contain information about the surface condition of equipment. The higher the degree of anomaly, the more significant the texture breaks and color deviations appear in the image. Convolutional neural networks can learn the correspondence between image features and the severity of anomalies by extracting image features layer by layer.

[0041] Optionally, the training data comes from a historical inspection image database. Domain engineers label each image with a severity score according to industry defect grading standards, with values ​​ranging from 0 to 1. The labeling follows industry-standard equipment defect grading criteria. For example, in the power industry, transmission line equipment defects are classified into three levels: general, severe, and urgent. The corresponding labeling scores are mapped to the intervals of 0 to 0.3, 0.3 to 0.7, and 0.7 to 1.0, respectively. Engineers then assign a precise score within each interval based on the specific defect severity. During training, the SmoothL1 loss function (also known as Huber loss) is used. The loss value is calculated based on the deviation between the model's predicted score and the labeled score. The specific expression is as follows: in For the model to the first The predicted score for each sample. These are the corresponding labeled scores, and both are dimensionless quantities. This represents the number of samples in the batch. For Huber functions, when hour ,when hour Compared to mean squared error (MSE), SmoothL1 exhibits a constant and non-linear gradient growth even with large biases, is less sensitive to individual biases in the labeled scores, and has a more stable training process. The optimizer used is the Adam optimizer, with an initial learning rate set to... The learning rate adjustment strategy uses cosine annealing, with training epochs ranging from 50 to 100, and batch sizes set from 16 to 32. To prevent overfitting, a Dropout layer can be added before the final fully connected layer, with a dropout rate set to 0.3, meaning that 30% of the neuron outputs are randomly reset to zero during training. After training, the model parameters are fixed and deployed to an onboard edge computing unit.

[0042] After each inspection subtask is completed, the aircraft inputs the images collected by that task into the aforementioned model for forward inference to obtain anomaly scores. Then along with the mission number The report was sent to the ground dispatch station.

[0043] In some embodiments, for step S4, the ground dispatch station receives the completed task. Abnormal scores Then, based on the abnormal scores Determined correlation strength And completed tasks Unfinished tasks The spatial distance between tasks is used to calculate the priority increment of each unfinished task. This allows the inspection results of completed tasks to be incorporated into the online scheduling decision-making process from post-event analysis, enabling the scheduling system to infer risks in adjacent areas based on discovered local anomalies.

[0044] For spatial association diagrams and tasks Each incomplete task in the series Priority increment The calculation method is as follows: in For completed tasks The abnormal scores are output from the aforementioned steps; For the task With the task The strength of the correlation between them is determined by the aforementioned steps; This is the distance attenuation coefficient; For the task With the task The Euclidean distance between the two tasks, in meters, is calculated based on their geographic coordinates.

[0045] Among them, the distance attenuation coefficient It decreases with increasing spatial distance, and adopts a Gaussian decay form: in This refers to spatial distance, expressed in meters. The attenuation radius parameter, in meters, represents the effective spatial range of the associated influence; the exponential part... The numerator and denominator are both in meters squared. A characteristic of Gaussian decay is that it decays slowly at close ranges, and decays further at distances exceeding... After the decay accelerates, the distance reaches to The attenuation coefficient is close to 0, which is consistent with the actual law that the mutual influence between adjacent devices decreases rapidly with distance.

[0046] Attenuation radius parameter Based on the spatial relevance of the inspected objects, specifically, from historical defect records, the conditional probability of defect co-occurrence between two inspected targets is statistically analyzed according to different distance intervals, resulting in a statistical curve showing the conditional probability decaying with distance; the distance value corresponding to the conditional probability decaying to a preset lower limit (e.g., 0.1) is taken as... The reference value, i.e. Take half of that distance value. For example, in transmission line inspections, the environmental impacts (such as icing, strong winds, and salt spray) within the same section typically cover 3 to 5 spans. If the average span is 350 meters, the probability of defect co-occurrence decreases to below 0.1 at a distance of approximately 1200 meters. It can be configured to 600 meters. In bridge inspection scenarios, the spatial correlation scale between components within the same span is typically 0.3 to 0.5 times the span. The smaller value should be selected accordingly.

[0047] In one embodiment, during task execution, multiple completed tasks may simultaneously send requests to the same incomplete task. Priority increment propagation. Simply summing the increments from various sources can easily lead to an artificially inflated priority due to the superposition of multiple weakly correlated signals. A target with weak correlations to many completed tasks might have an increment exceeding that of a target with strong correlations but only receiving a single-source signal, even if the anomalous signals from each source are very minor. This contradicts the true meaning of risk propagation; the risk level of a target should depend on the strongest single correlated signal, not the accumulation of numerous weak signals. Therefore, when multiple completed tasks propagate priority increments to the same incomplete task, the maximum value among the priority increments should be used, rather than summing them.

[0048] Let the set of completed tasks be... The maximum value is added to the initial priority value of the unfinished task to obtain the updated priority value. : in This is the initial priority value, determined during the task allocation phase. The operation retrieves all completed task pairs. The maximum value in the priority increments propagated; each time a new task is completed and an anomaly score is reported, the scheduling system re-executes the above calculation to update the updated priority values ​​of all incomplete tasks. .

[0049] For example, suppose there are 200 sub-tasks in a transmission line inspection operation, and 40 have been completed so far. Abnormal scores (Corresponding to a severe insulator crack), and the unfinished task. correlation strength (Both are located in the same base but different phases), spatial distance rice, rice: Meanwhile, another task has been completed. Abnormal scores (Corresponding to a slight surface oxidation), with the same target correlation strength (Both are located in adjacent bases), spatial distance rice: The priority increments for the two sources are 0.39 and 0.048 respectively, with the maximum value of 0.39 being taken. If... initial priority value Then the updated priority value This objective has been upgraded from medium priority to high priority and will be placed at the top in subsequent reordering.

[0050] In some embodiments, for step S5, please refer to Figure 2 , Figure 2 A flowchart illustrating the reassignment of incomplete tasks according to an embodiment of this disclosure is shown. At step 201, all incomplete tasks are reordered based on the updated priority values. Each time the scheduling system receives a new exception score and completes the priority update according to the aforementioned steps, it reorders the tasks based on the updated priority values. All incomplete tasks are reordered in descending order and compared to the task order within each aircraft queue in the current allocation scheme. The change in order is defined as the number of inversion pairs between the updated and current task orders; that is, the number of task pairs whose relative order is reversed in the two orders. A larger number of inversion pairs indicates a more drastic change in priority distribution and a greater deviation of the current allocation scheme from the latest risk assessment.

[0051] In S202, when the change in priority exceeds the rescheduling trigger threshold, the scheduling system determines that the current allocation scheme is no longer suitable for the latest priority distribution and triggers the rescheduling process. The rescheduling trigger threshold is based on the total number of currently incomplete tasks. Multiply by the preset scaling factor And round up to the nearest integer to determine, that is . The value ranges from 0.1 to 0.3, meaning that rescheduling is triggered when more than 10% to 30% of the task ordering relationships change. A smaller value indicates a more sensitive response to priority changes, suitable for scenarios with short job windows requiring rapid adjustments; a larger value indicates a more conservative response, suitable for scenarios with a large number of aircraft and high costs associated with frequent switching. In one example, The value is 0.15. If there are currently 160 unfinished tasks remaining, the threshold will be triggered. This means that a redistribution is triggered when the number of inversions exceeds 24.

[0052] The calculation of the number of inversions is a well-known method in this field, and can be obtained through the merging process of merge sort. That is, the current sort and the new sort are represented as task number sequences. When merging two ordered subsequences, whenever an element in the right subsequence is placed into the result sequence before an element in the left subsequence, the number of remaining elements in the left subsequence is accumulated to obtain the number of new inversions. The total computational complexity is O(n log n). .

[0053] In step 203, after triggering a reallocation, the scheduling system constructs a weighted bipartite graph with each aircraft and each incomplete task. One side of the weighted bipartite graph represents the available aircraft, and the other side represents the incomplete tasks. The allocation cost from each aircraft to each incomplete task is the ratio of the flight time required for the aircraft to reach the task to the updated priority value of that task. in For aircraft Unfinished tasks The cost of allocation The unit is seconds. It is a dimensionless quantity. The meaning of cost value is: after the update, the higher the priority value of the task, the larger the denominator, the smaller the cost, and the easier it is to be prioritized; the closer the spacecraft is, the shorter the flight time, the smaller the numerator, and the smaller the cost. According to aircraft Current location and task The straight-line distance between the locations is estimated by dividing the cruising speed of the aircraft by the cruising speed of the aircraft. The cruising speed is obtained from the performance parameters of the aircraft (for example, the cruising speed of a certain multi-rotor aircraft is 10 meters per second).

[0054] The scheduling system employs the Hungarian algorithm to find the minimum-weighted matching in this weighted bipartite graph, thus obtaining a redistribution scheme. The Hungarian algorithm is a classic method for finding optimal matching in a bipartite graph. Through alternating path search and potential function adjustment, it iteratively seeks a one-to-one correspondence scheme that minimizes the total matching weight. The global optimal solution is found in time, where This refers to the number of nodes on the side with more nodes. In real-world scenarios, since each aircraft is typically assigned multiple tasks, the scheduling system transforms the problem into a multi-round matching process: each round of matching assigns a highest-priority task to each aircraft. After the assignment is completed, the assigned task is removed from the set of tasks to be assigned, the aircraft's position is updated to the position of the assigned task, and then the next round of matching begins, until all tasks are assigned or the aircraft's remaining battery power is insufficient.

[0055] Specifically, in real-world scenarios, since each aircraft is typically assigned multiple tasks, the scheduling system transforms the reallocation problem into a multi-round sequential matching process, where each round assigns one unfinished task to each available aircraft. The specific process is as follows: Before the first round of matching begins, the real-time 3D coordinates of each aircraft (obtained from status reports) are used as the starting position of each aircraft, and all incomplete tasks (tasks that have not yet been completed by any aircraft) are set as tasks to be assigned. The scheduling system constructs a weighted bipartite graph for this round: the left-hand nodes represent all available aircraft (denoted as the set). The quantity is The right-hand node contains all tasks to be assigned (the number is...). Cost matrix elements in (No. (Wheel) according to the formula in S4.2 Calculation, where For aircraft Current location to task The estimated flight time. When At that time, a rectangular extended version of the standard Hungarian algorithm (i.e., zero-padding for dummy nodes) is used: if ,Add to A virtual spacecraft, the cost of which is used for all missions is set to a maximum constant, such as... If the virtual spacecraft is matched, the corresponding task will not be assigned in this round; if ,Add to Each virtual task, the value of which is allocated to all aircraft is set as follows: (The spacecraft assigned a virtual task after matching will idle in this round). The Hungarian algorithm outputs a solution. Binary matching matrix ,in Indicates aircraft Tasks assigned in this round , This indicates that the allocation is unassigned.

[0056] After a round of matching is completed, the scheduling system performs the following update operations: 1) Remove the tasks already assigned in this round from the set of tasks to be assigned. 1) Remove from the list; 2) Update the current position of each aircraft to the position coordinates of the task assigned to it in this round (if an aircraft is not assigned a task in this round, its position remains unchanged and it will continue to participate in matching in the next round); 3) Check the set of tasks to be assigned. If the value is empty, the reallocation ends; otherwise, proceed to the next round of matching.

[0057] The multi-round matching process terminates when any of the following conditions are met: (a) the set of tasks to be assigned. (a) Empty, all unfinished tasks have been assigned; (b) All aircraft have reached the upper limit of the power feasibility constraint and can no longer receive any new tasks; (c) No tasks are assigned in two consecutive rounds of matching (i.e. all remaining tasks do not meet the power constraints for all aircraft), at which point the remaining tasks are marked as "to be executed after subsequent return" and re-planned by the scheduling system in the next job window.

[0058] In each round of matching, the computational complexity of the Hungarian algorithm is O(n log n). For typical scenarios (5-20 aircraft, 50-500 tasks), the solution time for a single round of matching is in the millisecond to second range, and the total time for multiple rounds of matching does not exceed 5 seconds, meeting the requirements for online scheduling.

[0059] During the solution process, a power feasibility constraint is imposed on each aircraft, meaning that the total estimated energy consumption of all unfinished tasks assigned to the same aircraft must not exceed a preset safety percentage of the aircraft's current remaining power. This safety percentage is typically set at 80% to 85%, reserving 15% to 20% of power for return and emergency use. The estimated total energy consumption is estimated by multiplying the total flight distance (including inter-task transfer distance and flight distance within the operational area) of the aircraft sequentially performing its assigned tasks by the aircraft's average energy consumption rate per unit distance, which is obtained from the aircraft's factory performance parameters. Allocation schemes that do not meet the power feasibility constraint are eliminated during the matching process, and the aircraft will no longer receive new tasks.

[0060] In step 204, after the reallocation scheme is determined, the scheduling system issues new task queues to each aircraft via data link. Once an aircraft completes its currently executing inspection subtask, it switches to the next inspection subtask specified by the reallocation scheme. If an aircraft is currently executing an inspection subtask and has not yet completed it, the task is not interrupted; it is switched to after completion to avoid generating incomplete inspection data.

[0061] Optionally, when the priority value of an inspection subtask suddenly rises to the highest value among all incomplete tasks after an update, and the task is closest to an aircraft currently performing a task, the scheduling system can issue an emergency task insertion command to the aircraft, inserting the high-priority task into the next position of the aircraft's current queue.

[0062] In one embodiment, a reassignment might require an aircraft to abandon a low-priority task in its original queue and instead execute a newly elevated high-priority task. The abandoned task is then reassigned by the scheduling system to other aircraft with available capacity. To avoid repeated adjustments to the allocation scheme and frequent aircraft rerouting due to multiple abnormal results within a short period, a minimum execution interval constraint is introduced. Specifically, after receiving a reassignment scheme, if the time interval since the last task switch is less than the preset minimum execution interval, the switch is postponed until the currently en route inspection subtask is completed before executing the next inspection subtask specified in the reassignment scheme. The minimum execution interval is determined based on the average execution time of a single inspection subtask, for example, it can be configured to be 0.5 to 1.0 times the average execution time; in one example, if the average time for a single inspection subtask is 5 minutes, the minimum execution interval is set to 3 minutes. The scheduling system maintains the last switch timestamp in the status record of each aircraft. Before issuing each reassignment command, it first queries this timestamp and compares it with the current time. A switch command is only issued if the interval exceeds the minimum execution interval.

[0063] In some embodiments, when an aircraft needs to return to base early due to insufficient remaining power, its unfinished inspection subtasks, along with the current updated priority values ​​of each task, are transferred to the scheduling system. The scheduling system then redistributes these inspection subtasks to the remaining aircraft according to the aforementioned redistribution method.

[0064] In some embodiments, each aircraft reports its current position coordinates, flight altitude, flight speed, remaining battery percentage, and current mission number to the ground control station at a fixed interval. The reporting interval can be configured, for example, every 5 seconds. Based on this information, the control system updates the status information of each aircraft for flight time management. The system estimates and assesses the feasibility of power constraints. When communication is interrupted for more than a preset duration (e.g., 30 seconds), the scheduling system marks the aircraft as having an unknown status, suspends the assignment of new tasks to it, and reinstates it for assignment once communication is restored.

[0065] Among them, the abnormal score Uploaded via a separate data channel, and transmitted separately from flight status data, the raw image data collected during inspections is large in volume. After inference is completed by the onboard edge computing unit, only the scalar result of the anomaly score (e.g., a 4-byte floating-point number) is uploaded, resulting in a very small data volume. The impact of transmission delay on scheduling response speed is negligible. The raw images can be downloaded uniformly after the aircraft returns to base for subsequent detailed defect analysis and report generation.

[0066] The construction of the spatial association graph is completed offline before the mission begins, without occupying flight operation time. The priority incremental calculation involves multiplication and maximum value operations, and the computational complexity is proportional to the number of edges in the spatial association graph. For a typical inspection scenario with hundreds of target points, conventional ground station computing equipment can complete the calculation in milliseconds. The computational complexity of weighted bipartite graph matching is on the polynomial order of the number of aircraft and the number of missions. For a common scale, such as 5 to 20 aircraft and 50 to 500 inspection sub-missions, the solution time is within seconds.

[0067] Therefore, this method describes the physical and engineering relationships between targets by establishing a spatial correlation diagram of inspection targets before the mission starts. During the mission execution, the anomaly scores of completed tasks are used to propagate priority increments to uncompleted tasks along the correlation diagram. When the priority ranking changes significantly, task reallocation is triggered. This allows the scheduling system to infer the risks of adjacent areas and adjust the focus of subsequent operations based on the discovered local anomalies within a limited operation window. The proportion of high-risk areas being covered by priority is increased, reducing the possibility of missing key hidden dangers due to insufficient window. At the same time, low-risk areas do not occupy aircraft resources prematurely, thus improving the overall scheduling efficiency.

[0068] According to another aspect of the invention, see [reference needed]. Figure 3, Figure 3 This is a structural block diagram of a low-altitude multi-aircraft cooperative scheduling system provided in an embodiment of this application. Figure 3 As shown, the system includes: The task initialization module 301 is used to divide the objects to be inspected in the work area into multiple inspection sub-tasks, assign an initial priority value to each inspection sub-task, and assign each inspection sub-task to multiple aircraft. The spatial association graph establishment module 302 is used to establish a spatial association graph of inspection targets based on the physical structure relationship and engineering attributes between each inspection target. Each node in the spatial association graph corresponds to an inspection sub-task, and the edge weight between nodes is the association strength. The anomaly detection and priority increment calculation module 303 is used to perform anomaly detection on the collected inspection data after the aircraft completes the inspection sub-task, and output the anomaly score of the completed task; and calculate the priority increment of each uncompleted task based on the anomaly score, the correlation strength and the spatial distance between the completed task and the uncompleted task. The priority update module 304 is used to take the maximum value among the priority increments when multiple completed tasks propagate priority increments to the same unfinished task; and to obtain the updated priority value based on the maximum value and the initial priority value of the unfinished task. The task reassignment module 305 is used to reorder all unfinished tasks according to the updated priority value. When the order change exceeds the rescheduling trigger threshold, the unfinished tasks are reassigned according to the current status of each aircraft and the updated priority value.

[0069] Based on the same inventive concept, this application also provides an electronic device, the method corresponding to which can be the method in the foregoing embodiments, and its problem-solving principle is similar to that method. For example... Figure 4 As shown, Figure 4 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods and / or technical solutions of the foregoing embodiments of the present application.

[0070] Embodiments of the present invention also provide a non-transitory computer-readable storage medium that can be disposed in an electronic device to store at least one instruction or at least one program related to implementing a method in the method embodiments, wherein the at least one instruction or the at least one program is loaded and executed by the processor to implement the method provided in the above embodiments.

[0071] While specific embodiments of the invention have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of the invention. It should also be understood that various modifications can be made to the embodiments without departing from the scope and spirit of the invention. The scope of the invention is defined by the appended claims.

Claims

1. A method for coordinated scheduling of multiple aircraft for low-altitude operations, characterized in that, include: The objects to be inspected within the work area are divided into multiple inspection sub-tasks, each inspection sub-task is assigned an initial priority value, and each inspection sub-task is assigned to multiple aircraft. Based on the physical structural relationships and engineering attributes between each inspection target, a spatial association graph of the inspection targets is established. Each node in the spatial association graph corresponds to an inspection sub-task, and the edge weight between nodes is the association strength. After the aircraft completes the inspection sub-task, it performs anomaly detection on the collected inspection data and outputs the anomaly score of the completed task. Based on the anomaly score, the correlation strength, and the spatial distance between completed and incomplete tasks, calculate the priority increment of each incomplete task; When multiple completed tasks propagate priority increments to the same incomplete task, the maximum value among the priority increments is taken; based on the maximum value and the initial priority value of the incomplete task, the updated priority value is obtained. All incomplete tasks are reordered according to the updated priority value. When the change in order exceeds the rescheduling trigger threshold, the incomplete tasks are reallocated according to the current state of each aircraft and the updated priority value.

2. The method according to claim 1, characterized in that, The correlation strength is calculated based on the structural correlation coefficient and the process batch similarity coefficient; the structural correlation coefficient is determined according to the physical positional relationship and structural connection relationship between the two inspection targets, and the process batch similarity coefficient is determined according to the installation batch information of the two inspection targets; the values ​​of the structural correlation coefficient and the process batch similarity coefficient are both between 0 and 1.

3. The method according to claim 2, characterized in that, The method for determining the structural correlation coefficient includes: statistically analyzing the conditional probability of two inspection targets having defects simultaneously from historical inspection records, and using the conditional probability as the value of the structural correlation coefficient; the method for determining the process batch similarity coefficient includes: querying the installation batch information of two inspection targets from the equipment ledger, statistically analyzing the conditional probability of common defects appearing in the same batch of equipment from historical records, and using the conditional probability as the value of the process batch similarity coefficient.

4. The method according to claim 1, characterized in that, When calculating the priority increment, a distance attenuation coefficient is applied based on the spatial distance between completed and uncompleted tasks; the distance attenuation coefficient decreases as the spatial distance increases, and adopts a Gaussian attenuation form; The attenuation radius parameter is determined based on the spatial correlation scale of the inspection object. The spatial correlation scale is calibrated according to the distance corresponding to the conditional probability of defect co-occurrence attenuating to a preset lower limit value with distance.

5. The method according to claim 1, characterized in that, The reallocation of unfinished tasks includes: constructing a weighted bipartite graph with each aircraft and each unfinished task, where the allocation cost from each aircraft to each unfinished task is the ratio of the flight time required for the aircraft to reach the task to the updated priority value of the task; using the Hungarian algorithm to solve the minimum weight matching of the weighted bipartite graph to obtain a reallocation scheme; and after the aircraft completes the currently executing inspection subtask, switching to the next inspection subtask specified by the reallocation scheme.

6. The method according to claim 1, characterized in that, The anomaly detection employs a lightweight convolutional neural network, which uses MobileNetV2 as its backbone network. The original classification layer at the end of MobileNetV2 is replaced with a fully connected layer containing a single neuron. The activation function of this fully connected layer is the sigmoid function, which directly outputs a continuous anomaly score. The image acquired by the inspection sensor is used as input, and the output is a continuous scalar value between 0 and 1 as the anomaly score.

7. The method according to claim 6, characterized in that, The training data for the lightweight convolutional neural network comes from historical inspection images, with each image labeled with a severity score according to the defect grading standard. During training, the SmoothL1 loss function is used, and the loss value is calculated based on the deviation between the model's predicted score and the labeled score. The optimizer uses the Adam optimizer, and the learning rate adjustment strategy is cosine annealing.

8. The method according to claim 1, characterized in that, After receiving the redistribution plan, if the time interval since the last task switch is less than the preset minimum execution interval, the switch will be postponed until the current en route inspection sub-task is completed before the next inspection sub-task specified in the redistribution plan is executed. The minimum execution interval is determined based on the average execution time of a single inspection subtask.

9. The method according to claim 5, characterized in that, When solving the minimum weight matching of the weighted bipartite graph, an energy feasibility constraint is imposed on each aircraft: the total expected energy consumption of all unfinished tasks allocated to the same aircraft does not exceed a preset safety ratio of the aircraft's current remaining energy.

10. A low-altitude multi-aircraft collaborative scheduling system, characterized in that, include: The task initialization module is used to divide the objects to be inspected in the work area into multiple inspection sub-tasks, assign an initial priority value to each inspection sub-task, and assign each inspection sub-task to multiple aircraft. The spatial association graph establishment module is used to establish a spatial association graph of inspection targets based on the physical structure relationship and engineering attributes between each inspection target. Each node in the spatial association graph corresponds to an inspection sub-task, and the edge weight between nodes is the association strength. The anomaly detection and priority incremental calculation module is used to detect anomalies in the collected inspection data after the aircraft completes the inspection sub-task and output the anomaly score of the completed task. Based on the anomaly score, the correlation strength, and the spatial distance between completed and incomplete tasks, calculate the priority increment of each incomplete task; The priority update module is used to take the maximum value among the priority increments when multiple completed tasks propagate priority increments to the same unfinished task; and to obtain the updated priority value based on the maximum value and the initial priority value of the unfinished task. The task reassignment module is used to reorder all incomplete tasks according to the updated priority value. When the change in order exceeds the rescheduling trigger threshold, the incomplete tasks are reassigned according to the current state of each aircraft and the updated priority value.

Images