An unmanned aerial vehicle task identification method based on map task elements and state feature fusion

Abstract

This invention provides a method for UAV task recognition by fusing map task elements and state features. It completes the mapping of task elements on the flight map, with the mapping result showing the proportion of five task categories in each grid cell; obtains UAV flight state feature data; for historical data with corresponding tasks, the task type is directly used as the sample label Y; for historical data without recorded task types, the feature samples are clustered, and the clustering results are mapped to task categories; the feature data and sample data are fed into a deep learning neural network for training, and the network structure and parameters are optimized; new feature data is input to predict the task, with the prediction result showing the probability of five task categories; the category proportion of the flight map task elements is fused with the output category probability to obtain the fused probability of the associated environmental space and flight state features. This invention improves the comprehensiveness, objectivity, and accuracy of UAV task recognition in airspace.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) technology, specifically relating to a UAV mission recognition method that fuses map mission elements and state features. Background Technology

[0002] As the drone industry matures, the integration of civilian drones into controlled airspace is an unstoppable trend. The burgeoning commercialization of drones has led to an explosive growth in both the types and numbers of civilian drones. Different types of drones possess numerous specific mission attributes; therefore, efficient and accurate identification of the missions currently being performed by drones within airspace can provide insights into drone behavior and intentions, offering a reference for avoiding potential conflicts. This can further enhance my country's civilian drone regulatory capabilities and provide technical support for the rapid integration of drones into the airspace regulatory system.

[0003] Currently, UAV flight monitoring technology is relatively mature. It mainly uses radar, ADS-B and other technologies to detect, receive and collect flight status characteristic data such as position coordinates, speed, attitude and timestamps of cooperative and non-cooperative UAVs, providing reliable data support for the task identification of UAVs in this solution.

[0004] Besides the drone's own flight characteristics, the surrounding environment is also a crucial factor. Although drones perform different tasks, they may exhibit similar flight characteristics at a given moment. Therefore, it is difficult to make an objective and accurate judgment about the drone's current mission based solely on flight data characteristics.

[0005] Among the existing publicly available technologies, CN114169367A, "A Method and System for Identifying Low-Speed, Small, and Slow Unmanned Aerial Vehicles Based on Deep Learning," identifies the type of unmanned aerial vehicle (UAV) by detecting and processing radio signals, extracting features based on an UAV classification and identification algorithm, and utilizing flight signal frequency band feature parameters and a deep learning convolutional neural network. CN110958200A, "A Method for Identifying UAVs Based on Radio Features," establishes a UAV signal feature database, extracts and identifies UAV signal features, and finally compares the UAV signals using the feature database to identify the UAV model.

[0006] Both of the aforementioned inventions classify and identify drone types and models by processing signal characteristics during drone communication. While these methods reliably classify drones based on the specific radio signal characteristics of each model, they still have certain limitations for the drone mission field identification scenario mentioned in this solution. First, drones of the same model may still perform different tasks depending on the user's needs; furthermore, with the development of the civilian drone industry, it is increasingly common for multiple types of drones and even other electronic devices to share the same signal frequency band, which will inevitably lead to severe co-channel interference, posing significant challenges to hardware design and anti-interference capabilities. Second, the need to establish a database matching signal characteristics and drone models means that models outside the database cannot be effectively identified, limiting the generalization performance of the identification results.

[0007] The paper "Research on UAV Behavioral Intent Classification Method Based on Appearance Features" (Science and Technology Innovation and Application, Vol. 11, No. 14, 2021, pp. 139-142) classifies civilian UAVs into five categories based on the appearance characteristics of their payloads: surveillance and reconnaissance, emergency rescue, transportation, agricultural and forestry plant protection, and aerial photography and mapping. A database for UAV behavior pattern discrimination is established based on these appearance features. The aforementioned methods often require the use of photoelectric detection to collect video or image data of the UAVs. The appearance features are then processed to determine the UAV's category. However, photoelectric detection is affected by objective factors such as the environment, obstacles, and equipment detection capabilities, resulting in significant errors in the identification results. This makes it difficult to conduct all-weather, indiscriminate surveillance of UAVs in the airspace, exhibiting clear limitations. Summary of the Invention

[0008] This invention addresses the problems of poor generalization and low accuracy in task recognition caused by the limited feature dimensions of civilian UAVs and the lack of multi-dimensional and three-dimensional task recognition capabilities. It proposes a multi-feature heterogeneity extraction method based on task division and a task element division method for UAV flight maps, which integrates two feature dimensions: UAV flight status and environmental space, to achieve effective identification and division of UAVs at the task level.

[0009] The specific technical solution of this invention is as follows:

[0010] A method for UAV mission recognition that fuses map mission elements and state features, characterized in that:

[0011] Firstly, the monitored area must be equipped with radar and ADS-B devices to effectively capture the position coordinates, speed, and timestamp flight status characteristics of UAVs within the airspace.

[0012] The drone missions are categorized into five types: last-mile logistics, aerial photography and mapping, emergency rescue, agricultural and forestry plant protection, and industrial inspection. Feature functions are constructed for each of the five mission types and their flight status characteristics. When the maximum recognition probability of the five mission types is less than 50%, the mission is declared unidentified.

[0013] The specific steps of the identification method are as follows:

[0014] Step 1: Based on the land use characteristics of the UAV's flight airspace, complete the task meta-mapping of the flight map. Each grid may simultaneously possess several task attributes. The mapping result is the proportion M of the five task types in each grid. k ={m k1 ,m k2 ,m k3 ,m k4 ,m k5}, M k This represents the distribution of the five task types in the k-th grid, {m k1 ,...,m k5 The percentages of five types of tasks—last-mile logistics, aerial surveying and mapping, emergency rescue, agricultural and forestry plant protection, and industrial inspection—are displayed in the task grid.

[0015] Step 2: Collect historical flight status data of the UAV, and obtain feature data X after preprocessing. i ={x i1 ,x i2 ,x i3 ,x i4 ,x i5}, X i This represents a feature vector constructed based on historical data of the drone, {x} i1 ,...,x i5} represent the five construction features respectively;

[0016] Step 3: For historical data with corresponding tasks, directly use the task type as the sample label Y. i For historical data that does not record task types, cluster the feature samples X, and map the clustering results to task categories to ensure that each feature data has a corresponding label value.

[0017] Step 4: Train the feature data and sample data into a deep learning neural network with 5 input and 5 output neurons, and optimize the network structure and parameters;

[0018] Step 5: Using the trained network parameters and new feature data, predict the task, with the prediction result being the probability of 0 for each of the 5 task classes. j ={o j1 ,o j2 ,o j3 ,o j4 ,oj5}, where O j Let {o} represent the neural network prediction result vector of UAV j. j1 ,...,o j5} represent the predicted probabilities of the five task categories, respectively;

[0019] Step 6: Fuse the category proportions of the flight map task elements with the output category probabilities to obtain the fusion probabilities of the associated environmental space and flight state features:

[0020] FP jk ={αo j1 +βm k1 ,αo j2 +βm k2 ,αo j3 +βm k3 ,αo j4 +βm k4 ,αo j5 +βm k5}

[0021] FP jk Let α represent the task fusion probability vector of UAV j operating in grid k, where α and β are fusion coefficients, representing the proportions of environmental space and flight state.

[0022] The five types of tasks are as follows:

[0023] A. Last-mile logistics:

[0024] In last-mile logistics scenarios, due to limitations in infrastructure, the current origin and destination of last-mile logistics drones are freight warehouses or transit stations, and they fly along the reported flight routes in between.

[0025] B. Aerial surveying and mapping:

[0026] For visual range flight;

[0027] C. Emergency Rescue:

[0028] In emergency rescue scenarios, the origin and destination points of drones are mostly located at emergency response departments, and they may hover in residential areas; D. Agricultural and forestry plant protection:

[0029] In the context of agricultural and forestry plant protection, the flight area is basically located in agricultural and forestry areas;

[0030] E. Industrial Inspection:

[0031] In industrial inspection scenarios, drones fly along pre-planned routes and speeds based on the layout of pipelines, with the flight area located in areas where pipelines are concentrated.

[0032] Furthermore, the flight characteristics of UAVs in the aforementioned five mission scenarios are constructed using the following feature functions:

[0033] (1) Yaw

[0034] Yaw refers to the degree of deviation between the actual flight path of the UAV and the reported planned flight path. It is calculated by taking the distance d between the two points based on the coordinates of the reported planned flight path nodes and the current actual coordinates of the UAV. i :

[0035] d i =1000·R·arccos(siny′) i sin y i +cos y′ i cos y i cos(x′ i -c i ))

[0036] R is the Earth's radius, with a mean of 6370 km. The actual coordinates of the UAV node i are (x... i ′,y i The coordinates (latitude and longitude) of the reported planned route node i are (x'). i ,y i Based on the degree of distance deviation, set the yaw counter C. i :

[0037]

[0038] The method for calculating the overall yaw of the UAV's flight range is as follows:

[0039]

[0040] (2) Horizontal cumulative range

[0041] Horizontal cumulative range refers to the cumulative flight distance of a UAV's flight path. It is an important reference for the UAV's operating area and reflects the coverage range of the current mission.

[0042]

[0043] (3) Average hovering time ratio

[0044] The average hovering time ratio refers to the ratio of the average hovering time per segment of a drone's flight to the total time. It represents the proportion of time the drone spends hovering during flight and is used to reflect the drone's hovering state.

[0045]

[0046] th iT represents the time the drone hovers during the i-th segment of its flight; i This represents the time taken for the i-th segment of the journey;

[0047] (4) Rate of change of horizontal distance

[0048] The rate of change of horizontal distance refers to how quickly the UAV's spatial position changes in the horizontal dimension, reflecting the UAV's velocity component in the horizontal direction. The calculation method for the rate of change of horizontal distance for the i-th segment of the flight is as follows:

[0049]

[0050] (5) Rate of change of vertical distance

[0051] The rate of change of vertical distance refers to how quickly the spatial position of the UAV changes in the vertical dimension. It reflects the velocity component of the UAV in the vertical direction. The method for calculating the rate of change of vertical distance for the i-th segment of the flight is as follows:

[0052]

[0053] h i+1 ′ and h i ′ represent the vertical height of the UAV at time i+1 and time i, respectively; T i This represents the time taken for the i-th segment of the journey.

[0054] This invention addresses the problems of incomplete features, complex identification methods, and low efficiency in task identification scenarios. It innovatively proposes to use the UAV flight status and environmental space as reference factors for task identification and calculate the fusion probability, thereby improving the comprehensiveness, objectivity, and accuracy of UAV task identification in the airspace. Attached Figure Description

[0055] Figure 1 This is the flight map grid task meta-mapping in the embodiment;

[0056] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0057] The specific technical solution of the present invention will be described in conjunction with the accompanying drawings.

[0058] This invention provides a method for UAV mission identification by fusing map mission elements and state features, comprising the following steps:

[0059] Firstly, the monitored area must be equipped with radar and ADS-B devices to effectively capture the flight status characteristics of UAVs within the airspace, such as their position coordinates, speed, and timestamps.

[0060] This invention categorizes UAV missions into five types: last-mile logistics, aerial photography and mapping, emergency rescue, agricultural and forestry plant protection, and industrial inspection. Feature functions are constructed for each of these five mission types and their flight status characteristics. When the maximum recognition probability for any of the five mission types is less than 50%, the mission is declared unidentified.

[0061] A. Last-mile logistics:

[0062] In last-mile logistics scenarios, the coverage range of drones is mostly within 5 kilometers, generally not exceeding 20 kilometers. Limited by infrastructure conditions, current last-mile logistics drones typically originate and terminate at freight warehouses or transit points, flying along pre-registered flight routes. B. Aerial surveying and mapping:

[0063] In aerial surveying scenarios, to ensure the shooting angle, clarity, and continuity, drones need to perform actions such as climbing, descending, moving, and even hovering with minimal attitude changes. They mainly fly in public areas such as parks, squares, and roads. Furthermore, because these drones have high requirements for real-time video transmission, they mostly fly within visual line of sight.

[0064] C. Emergency Rescue:

[0065] In emergency rescue scenarios, the origin and destination points of drones are mostly located in emergency response departments such as hospitals and fire stations, and they may also hover in residential areas.

[0066] D. Agricultural and forestry plant protection:

[0067] In agricultural and forestry plant protection scenarios, drones are often designed with pre-defined flight altitudes and routes to ensure even spraying of the materials, and the flight area is basically located in agricultural and forestry areas.

[0068] E. Industrial Inspection:

[0069] In industrial inspection scenarios, drones fly along pre-planned routes and speeds based on the layout of pipelines, with the flight area mostly located in areas where pipelines are concentrated.

[0070] Based on the flight characteristics of UAVs in the above five mission scenarios, the following feature functions are constructed:

[0071] (1) Yaw

[0072] Yaw refers to the degree of deviation between the actual flight path of the UAV and the reported planned flight path. The distance d between the two points is calculated based on the coordinates of the reported planned flight path nodes and the current actual coordinates of the UAV.

[0073] d i =1000·R·arccos(siny′) i siny i +cosy′ i cosyi cos(x′ i -x i ))

[0074] R is the Earth's radius, with a mean of 6370 km. The actual coordinates of the UAV node i are (x... i ′,y i The coordinates (latitude and longitude) of the reported planned route node i are (x'). i ,y i Based on the degree of distance deviation, set the yaw counter C. i :

[0075]

[0076] The method for calculating the overall yaw of the UAV's flight range is as follows:

[0077]

[0078] (2) Horizontal cumulative range

[0079] Horizontal cumulative range refers to the cumulative flight distance of a UAV's flight path. It is an important reference for the UAV's operating area and reflects the coverage range of the current mission.

[0080]

[0081] (3) Average hovering time ratio

[0082] The average hovering time ratio refers to the ratio of the average hovering time per segment of a drone's flight to the total time. It represents the proportion of time the drone spends hovering during flight and reflects the drone's hovering status.

[0083]

[0084] th i T represents the time during which the UAV remains in a hovering state (i.e., its altitude does not change) during the i-th segment of the flight. i This represents the time taken for the i-th segment of the journey.

[0085] (4) Rate of change of horizontal distance

[0086] The rate of change of horizontal distance refers to how quickly a UAV's spatial position changes in the horizontal dimension, reflecting the UAV's velocity component in the horizontal direction. The calculation method for the rate of change of horizontal distance for the i-th segment of the flight is as follows:

[0087]

[0088] (5) Rate of change of vertical distance

[0089] The rate of change of vertical distance refers to how quickly the spatial position of the UAV changes in the vertical dimension. It reflects the velocity component of the UAV in the vertical direction. The method for calculating the rate of change of vertical distance for the i-th segment of the flight is as follows:

[0090]

[0091] h i+1 ′ and h i ′ represent the vertical height of the UAV at time i+1 and time i, respectively; T i This represents the time taken for the i-th segment of the journey.

[0092] Figure 2 This is a flowchart of the entire technical solution. The specific steps of this technical solution are as follows:

[0093] Step 1: Based on the land use characteristics of the UAV's flight airspace, complete the task meta-mapping of the flight map. Each grid may simultaneously possess several task attributes. The mapping result is the proportion M of the five task types in each grid. k ={m k1 ,m k2 ,m k3 ,m k4 ,m k5}, M k This represents the distribution of the five task types in the k-th grid, {m k1 ,...,m k5 The percentages of five types of tasks—last-mile logistics, aerial surveying and mapping, emergency rescue, agricultural and forestry plant protection, and industrial inspection—are displayed in the task grid.

[0094] Step 2: Collect historical flight status data of the UAV, and obtain feature data X after preprocessing. i ={x i1 ,x i2 ,x i3 ,x i4 ,x i5}, X i This represents a feature vector constructed based on historical data of the drone, {x} i1 ,...,x i5} represent the five construction features respectively;

[0095] Step 3: For historical data with corresponding tasks, directly use the task type as the sample label Y. i For historical data that does not record task types, cluster the feature samples X, and map the clustering results to task categories to ensure that each feature data has a corresponding label value.

[0096] Step 4: Train the feature data and sample data into a deep learning neural network with 5 input and 5 output neurons, and optimize the network structure and parameters;

[0097] Step 5: Using the trained network parameters and new feature data, predict the task, with the prediction result being the probability of 0 for each of the 5 task classes. j ={o j1 ,o j2 ,o j3 ,o j4 ,o j5}, where O j Let {o} represent the neural network prediction result vector of UAV j. j1 ,...,o j5} represent the predicted probabilities of the five task categories, respectively;

[0098] Step 6: Fuse the category proportions of the flight map task elements with the output category probabilities to obtain the fusion probabilities of the associated environmental space and flight state features:

[0099] FP jk ={αo j1 +βm k1 ,αo j2 +βm k2 ,αo j3 +βm k3 ,αo j4 +βm k4 ,αo j5 +βm k5}

[0100] FP jk Let α represent the task fusion probability vector of UAV j operating in grid k, where α and β are fusion coefficients, representing the proportions of environmental space and flight state.

Claims

1. A method for UAV mission recognition by fusing map task elements and state features, characterized in that: First, the monitored area must be equipped with radar and ADS-B equipment to effectively capture the position coordinates, speed, and timestamp flight status characteristics of UAVs in the airspace. The tasks of drones are divided into five categories: last-mile logistics, aerial photography and mapping, emergency rescue, agricultural and forestry plant protection, and industrial inspection. Feature functions are constructed for the five types of tasks and flight status characteristics. When the maximum probability of recognizing the 5 types of tasks is less than 50%, the task is declared unrecognized. The specific steps of the identification method are as follows: Step 1: Based on the land use characteristics of the UAV's flight airspace, complete the task meta-mapping of the flight map. Each grid may have several task attributes simultaneously. The mapping result is the proportion of the five types of tasks in each grid. , Indicates the first The distribution of the five task types in each grid. The percentage of tasks in the task grid is divided into five categories: last-mile logistics, aerial surveying and mapping, emergency rescue, agricultural and forestry plant protection, and industrial inspection. Step 2: Collect historical flight status data of the UAV and obtain feature data after preprocessing. , This represents a feature vector constructed based on historical data of the drone. These represent five structural features respectively; Step 3: For historical data with corresponding tasks, directly use the task type as the sample label. For historical data that does not record task types, feature samples will be used. Perform clustering and map the clustering results to task categories to ensure that each feature data has a corresponding label value; Step 4: Train the feature data and sample data into a deep learning neural network with 5 input and 5 output neurons, and optimize the network structure and parameters; Step 5: Using the trained network parameters and new feature data, predict the task, and the prediction result is the probability of the 5 task classes. ,in Indicates drone The neural network prediction result vector, These represent the predicted probabilities for the five task categories, respectively. Step 6: Fuse the category proportions of the flight map task elements with the output category probabilities to obtain the fusion probabilities of the associated environmental space and flight state features: ， Indicates in grid Operating drones The fusion probability vector of each task, and The fusion coefficient represents the proportion of environmental space and flight status.

2. The UAV mission recognition method based on the fusion of map task elements and state features according to claim 1, characterized in that, The five types of tasks are as follows: A. Last-mile logistics: In last-mile logistics scenarios, due to limitations in infrastructure, the current origin and destination of last-mile logistics drones are freight warehouses or transit stations, and they fly along the reported flight routes in between. B. Aerial surveying and mapping: For visual range flight; C. Emergency Rescue: In emergency rescue scenarios, the origin and destination of drones are often located at emergency response departments, and they may hover in residential areas. D. Agricultural and forestry plant protection: In the context of agricultural and forestry plant protection, the flight area is basically located in agricultural and forestry areas; E. Industrial Inspection: In industrial inspection scenarios, drones fly along pre-planned routes and speeds based on the layout of pipelines, with the flight area located in areas where pipelines are concentrated.

3. The UAV mission recognition method based on the fusion of map task elements and state features according to claim 2, characterized in that, The construction of feature functions for the five types of tasks and flight status features is as follows: (1) Yaw Yaw refers to the degree of deviation between the actual flight path of a UAV and the reported planned flight path. It is calculated by taking the distance between the two points based on the coordinates of the reported planned flight path nodes and the current actual coordinates of the UAV. : ， The radius is the Earth's radius, with an average of 6370 km, and the actual node of the drone. The coordinates are latitude and longitude ( , ), reported planned route nodes The coordinates of latitude and longitude are ( , ); Set the yaw counter according to the degree of distance deviation. : ， The method for calculating the overall yaw of the UAV's flight range is as follows: ， (2) Horizontal cumulative range Horizontal cumulative range refers to the cumulative flight distance of a UAV's flight path. It is an important reference for the UAV's operating area and reflects the coverage range of the current mission. ， (3) Average hovering time ratio The average hovering time ratio refers to the ratio of the average hovering time per segment of a drone's flight to the total time. It represents the proportion of time the drone spends hovering during flight and is used to reflect the drone's hovering state. ， Indicates the first The flight range, the time the drone remains hovering; Indicates the first The time taken for a segment of the voyage; (4) Rate of change of horizontal distance The rate of change of horizontal distance refers to how quickly a drone's spatial position changes in the horizontal dimension, reflecting the drone's position in the surrounding environment. The horizontal velocity component; The method for calculating the rate of change of horizontal distance in a segment of the voyage is as follows: ， (5) Rate of change of vertical distance The rate of change of vertical distance refers to how quickly a drone's spatial position changes in the vertical dimension. It reflects the drone's velocity component in the vertical direction. The method for calculating the rate of change of vertical distance for a segment of the voyage is as follows: ， and These represent the nodes respectively. Time and Node At any given moment, the drone's vertical altitude; Indicates the first The time taken for a segment of the voyage.

Images