[Top Talk/Top Talk] Sponsored by Meituan Institute of Technology and Research Cooperation Department, for all technical students, we regularly invite Meituan senior technical experts, industry experts, university scholars and best-selling authors to share best practices and Internet Contents such as hot topics and cutting-edge technology advancements in academia help Meituan students broaden their horizons and enhance their cognition.
On October 20, 2020, Top Talk invited a full-time teacher from Beijing University of Aeronautics and Astronautics, and asked him to bring a sharing titled "Sky Highway Based on Big Data in Space and Time: A Distributed Control Framework for Urban Air Mobility Management". This article is a transcript shared by Quanquan Teacher, I hope it can be helpful or enlightening to everyone.
/Report Guests/
Full Power｜Beijing University of Aeronautics and Astronautics
Quanquan, associate professor at Beijing University of Aeronautics and Astronautics, visiting professor at the University of Toronto (E-mail: [email protected]). Long-term reliable flight control. As the first author, he has completed three English books and published nearly 100 articles. Won the gold medal (ranked first) in the 2nd National University Automation Professional Young Teacher Experimental Equipment Design "Maker Competition".
/Report Summary/
UAV traffic and the recent rise of aerial mobility management have received widespread attention. To this end, traditional aviation giants such as Boeing, Airbus, Honeywell and Bell, as well as emerging world-class influential companies such as Uber have joined. This report proposes an air highway program and simulation test. The solution is based on big time and space data, considering transportation network, route planning and distributed control design. On the air highway, each drone has its own planned route and can fly autonomously and freely to support dense and three-dimensional flight traffic.
1. Background
Today I bring you some work done by our laboratory ( Reliable Flight Control Research Group of Beihang University ). The main content is to talk about a distributed control framework for urban air mobility management.
I believe that the students of Meituan must also very much look forward to the realization of drone delivery. The future drone delivery service will greatly change our current lifestyle. Although in densely populated areas, we still need people to complete the delivery service, but in sparsely populated areas, such as suburbs, it is better to use drones for delivery. The report shows that networked drones will bring 7-10 times the industry opportunities for the industry, which is why we started to do related work in this area about three years ago.
Can the transportation network and traffic management of drones use existing transportation methods? Through research, we found that traditional civil aviation networks are actually not suitable. Civil aviation aircraft are actually very sparse. In a three-dimensional space, the frequency of changes in the entire network is relatively low. When there are applications for network access, they basically go through centralized planning. . Although the highway network is very dense, it is a two-dimensional space, so the traffic network management is also autonomous. The railway is also two-dimensional. The dynamic change of the network means that just like we are on the Internet, we need to access the network, and the road cannot be built for us immediately. UAV transportation has something in common with roads and railways. The difference is that UAVs are in a three-dimensional space, and the dynamic changes of the network are relatively high. Therefore, when designing the UAV flight control framework, we hope to design a framework that can adapt to changes including the increase of UAVs and the expansion of the network. Regarding path planning, it can actually be done in a centralized and autonomous way. Regarding the content of this part, we published a review article " Overview of Low-altitude UAV Traffic Management and Recommendations " this year, and interested students can refer to it.
We hope to provide a low-altitude intelligent brain for the continuous growth of low-altitude drones and applications. For technical research, the difference between low altitude and high altitude is not that big, but at present we mainly consider more low altitude. For example, open the low altitude below 120 meters. Its characteristic is that there will be a lot of buildings in these places. If the drone falls, it will cause some harm to the personal and property safety in the area below, which is harmful to our urban traffic. A very big challenge. Therefore, we mainly consider the following three requirements:
-
Plan the route and take-off time of the drone to ensure that the drone takes off under the premise of avoiding conflicts: the take-off time of the drone route is the same as our current flight experience, sometimes the plane will stop somewhere and delay the take-off , The purpose is to avoid collisions and conflicts with other aircraft on the route. Of course, sometimes due to weather and other reasons, the aircraft needs to be delayed because the aircraft is waiting on the ground and the cost is minimal.
-
Successfully complete the flight mission without collision
-
Responding to the influence of uncertain factors such as weather and no-fly zones: We sometimes need to cut off certain routes during the flight of the drone. In this case, we hope that the aircraft can still fly to the destination, at least it can be safe Back to the nearby airport.
2. Air highway foundation: network and space-time big data
To carry out research on air highways, we need some research foundation of network and space-time big data.
First of all, we need infrastructure such as communication, navigation, and surveillance functions. These functions act as the eyes, ears, and nervous system of the entire UAV traffic management system, and are responsible for situational awareness and information transmission. Among them, communication is 4G/5G, satellite communication, etc. This is our network. Navigation is the navigation of our aircraft, such as navigation through base station positioning, radar, satellite, inertial navigation, and vision. The difference between surveillance and navigation is that navigation is for aircraft, and surveillance is what we, as officials, need to understand the dynamics of aircraft in the air. Some aircraft may navigate by themselves and tell underground ground stations through navigation or communication, so that we can monitor them. But some aircrafts are model airplanes and have no communication function and can only be seen passively. Then we can monitor them through some visible light, sound waves, etc., and aircraft may broadcast their own information through some ADS-Bs. Therefore, through the above functions, we can realize our understanding of the surrounding environment of the aircraft in flight and the understanding of the air environment on the ground. This is a basis for our research on air highways.
On the other hand, we need the support of space-time big data. First of all, we need to understand all no-fly zones, which will also change dynamically. Second, we need to understand big weather data so that we can plan aircraft to avoid extreme weather. At the same time, we also need to obtain geographic big data information. For example, through geographic big data, we can understand where there are obstacles and which areas are under grass. Based on this information, we can further extract some information to plan the aircraft’s route and route network. Etc. and plan the route of the aircraft. In addition, we can also know which places are densely populated through the mobile Internet, so that we can avoid these densely populated places when planning airway networks or routes. The above are all the spatiotemporal big data foundations that we need for research.
As mentioned earlier, the Internet is the basis for our research on air highways. At present, in air traffic, information is mainly shared through the Internet. But the network will have network quality, so what is the relationship between network quality and flight safety? Network quality is usually determined by three factors: noise, delay, and packet loss. Due to the network quality, the location of the obstacle obtained by the drone may be different from the actual location of the obstacle. Take the following picture as an example, the UAV has a deviation in its estimated position, and the estimated obstacle position has also been deviation. Therefore, we need to design a safe flight distance to deal with this uncertainty caused by network quality. This is the same reason that we need to keep the distance between cars when driving on the highway. Our concept of distance between cars uses drone air traffic, hoping to deal with these packet loss delays. Of course, some people use some control methods to solve these problems. Our method should be more suitable for traffic scenarios.
The distance of traditional air traffic is not so complicated. The distance between aircrafts is very far apart, so how should the distance between drones be controlled? Through research, we believe that the safety radius of the drone should meet the relationship in the above figure. rm is the radius of the aircraft itself, ro is the radius of the obstacle itself, rv is related to the speed and maneuverability of the aircraft, and re represents the influence of the network. . How to understand network influence? The delayed packet loss rate is Θ. If the packet loss rate is 1, it means that the aircraft is completely lost. From a conservative perspective, the aircraft may be in any position. Therefore, the closer Θ is to 1, the greater the safety radius, and τd represents a delay, and there will be a delay in network transmission. Some students may think that the delay of our phone calls is already very small, but in the air, we have verified through experiments that there is a certain delay in the network. On the other hand, the packet loss rate will increase as the distance increases. Therefore, we need to evaluate the network impact and design the aircraft safety radius based on the evaluation results. Under the safe radius, it can be considered that the aircraft has no network noise and is completely accurate. As long as the safety radii of the two aircraft do not intersect, the aircraft will definitely not collide. This is the design of our safety radius.
On the other hand, we need to use some data for risk assessment. As shown in the figure below, there are at least two factors: accident probability assessment (fGIA) and accident casualty assessment. Accident probability assessment refers to the possibility that the aircraft will fall without continuous flight, which can be estimated by statistical methods. One advantage of drones over airplanes is that even if they fall, they may not hurt people. There is an exposure model for aircraft falling. For example, falling on a tree or roof will have less impact on people. Therefore, we need geographic information support. Once the aircraft needs to make an emergency landing, we can use geographic information to find areas suitable for emergency landing such as lawns. At the same time, the exposure model is also related to population density. The damage model is related to the design of the aircraft and the momentum of the aircraft's fall. These factors are combined to obtain a risk assessment, which can be used for route network planning, route planning, emergency landing, etc. In formulating relevant standards, the relevant laws and regulations will be more concerned about the risk assessment of aircraft. Some risk assessment companies will also prove how risky the aircraft is based on the results of the risk assessment.
3. Sky Highway
The air highway is divided into three parts: model establishment, algorithm design, and experimental verification. Among them, the model establishment is divided into two types: route network model and drone model. We need to model the drone to manage the drone, and the ground also needs to send instructions to the drone, which is equivalent to a standard model. ; Algorithms can be divided into centralized air traffic control and distributed air traffic control. The centralized type can be considered that all instructions are issued by the ground station to the aircraft, and the aircraft do not communicate with each other, and the communication is completely coordinated through the ground. Distributed is relatively more flexible. The last is our simulation and experiment.
3.1 Model establishment
3.1.1 Airway network model
The airway network model can be considered as a network composed of nodes and edges. Our design goal is to maintain a safe distance between drones on different routes without interfering with each other. If the aircraft is on different waterways, for example, there are two directions on a road, but the aircraft on different waterways needs to maintain a corresponding safe distance, similar to two roads, the angle between them is very small, and you have to go to a node , If the angle is small to a certain extent, then the distance between two aircraft on different channels is very close, which may be dangerous. In the actual process, we cannot know the specific location of the drones, but can only know an approximate location that is not particularly accurate. Therefore, we need to maintain a safe distance between the drones.
The construction of the airway network requires an understanding of the airspace, such as the geographic information and population information we mentioned earlier. As shown in the figure below, the black area can be considered as a no-fly zone. There are two types, one is relatively sparse and the other is relatively dense. Either way, they can be connected by nodes. These nodes may be the take-off and landing points of the aircraft, and some nodes may be the intersections of air routes, just like the intersections of highways.
There are two optimization goals for the airway network construction. One is to hope that the total length of the airway network is as short as possible. Because the construction of the airway network is equivalent to an infrastructure project, it is necessary to ensure the communication, navigation, and monitoring on the airway network. Cost is required; second, we hope that the risk of the airway network is the lowest. Taking into account factors such as population density, we hope to draw the airway network as shown in the figure below, but this is a multi-objective optimization problem.
We have done some work on airway network modeling, using several methods as shown in the figure below.
-
The first is the morphological skeleton method , which is similar to the skeleton of image processing. Given a picture, its skeleton needs to be generated. The principle is very simple. Black is the dangerous boundary, and the resulting skeleton is these routes. The distance between the route and the two sides should be as far as possible.
-
The second is the triangulation method . The shortest path connecting the three points is not connecting the three sides into a triangle, but may be connecting them through the Fermat point.
-
Finally, there is the comprehensive method . The morphological skeleton method is suitable for dense maps, while the triangulation method is more suitable for sparse maps. The comprehensive method takes into account both dense and sparse situations, and the airway network is built in a semi-automatic way.
The implementation process of the morphological skeleton method is shown in the figure below. The first is skeleton extraction. Some students may ask why these variables are generated after skeleton extraction. This is because we need to ensure that the extracted skeleton is separated from the black danger zone on both sides. If it is greater than a certain threshold, if it is not met, it will be disconnected and removed, and then a straight line fitting will be performed. Of course, some target points must be added to it to connect with the entire network. Finally, the structure of the graph needs to be extracted, and the relationship between nodes and edges is extracted according to the modeling method of graph theory.
The triangulation method is to connect these nodes together through Fermat points. Some edges will pass through obstacles, and we will avoid them through optimization methods, and finally form a network. In addition, we can add factors such as population density to the map as a black no-fly zone. Usually in the map, "1" means there is an obstacle, and "0" means there is no obstacle. We have done a further work in the airway network modeling, using a probability between 0 and 1 to indicate that some no-fly zones such as walls can never fly in, but some areas are sparsely populated and are not suitable for use. "1" means that in this case, the probability of 0.4 and 0.5 can be used. We hope to build a route network in this way, which is not shown in the figure below.
The synthesis method is in the case of the following figure, we use the morphological skeleton method in the dense obstacle area, and use the Fermat point in the outer sparse area, and finally combine them into a network. Sometimes my own students will ask me, what is dense and what is sparse? I don’t think we should consider this issue and judge by ourselves, because the airway network modeling does not have to be a fully automated process, and once it is built, it does not need to be changed. Therefore, we need to manually determine each area during the modeling process. What is the shape of the airway network? In this way, the above two methods can be well taken into consideration, and finally a different airway network can be formed.
The figure below is a schematic diagram of the abstract structure of the channel. Inside the channel, we refer to the current highway. There is an isolation zone in the middle. This isolation zone is the safe distance between the two aircraft we mentioned earlier. It can be bidirectional.
On the other hand, the nodes of the airway network also have a structure, as shown in the figure below, we generally have a cylindrical structure. The node has multiple channels connected, and it is necessary to consider that the aircraft on different channels should not be too close. Therefore, it is necessary to increase the radius of the node to ensure that the aircraft on different channels are sufficiently far apart.
The airway network has an abstract structure and a specific structure inside. Through some constraints, it is ensured that different aircraft can be greater than the safe distance under any circumstances, which allows us to know how to design the entire airway network.
3.1.2 UAV model
We need to issue instructions to the drone, which requires a certain standard interface. For the interface, we have some modes such as power-off mode, waiting for authorization mode, pre-position mode, flight mode, obstacle avoidance mode, The emergency landing mode sends instructions to the UAV. In this case, the UAV is equivalent to being controlled by our traffic control system. This interface is not that standard yet. I hope that eventually we can have a standard for air traffic systems.
3.2 Algorithm design
Low-altitude traffic control algorithms include centralized air traffic control and distributed air traffic control.
3.2.1 Centralized Air Traffic Control
Centralized control can be divided into two parts: offline planning and online control. Offline planning means that you need to declare your flight plan before taking off, and then accept it for review. If the current airway network is very congested, the review will not pass, and you need to wait or re-declare the flight plan. If the review is passed, a waiting flight plan containing information such as take-off time and location will be generated, and the waiting flight plan will be written into the database of the air traffic management system to predict the air traffic situation.
However, there will be many uncertainties during the flight of the aircraft due to factors such as weather and its own state, and the changes in the flying speed of the aircraft will cause conflicts. Therefore, during the flight of the aircraft, we need to carry out some quantitative control of the aircraft. This is the online control of the aircraft. We can control the altitude and speed of the aircraft so that it can avoid conflicts. If conflicts cannot be avoided in the entire airway network, then we certainly do not want to have the same effect as the dominoes, because a local factor makes the entire airway network Changes. The easiest way to avoid conflicts is to avoid obstacles. One aircraft flies upwards and the other flies downwards. This is the advantage of air traffic. There is no way for cars to do it. It is probably the logic.
The following is the algorithmic process based on several key modules in the centralized low-altitude traffic control system:
-
Plan review algorithm
Step 1 : Obtain the obstacle avoidance distance of the newly added UAV?????a and planning information collection????????
Step 2 : Update the route network information ???? and the current time ????, obtain the waypoint ℎ???? of the drone through the Dijkstra algorithm, and calculate the time required to complete the flight?????? ??;
Step 3 : Solve the optimization problem for drones;
Step 4 : If there is a solution, go directly to step 5; if there is no solution and the cause is a conflict, judge the priority of the drones that conflict with it????????????????? ???????????????????, ???? ∈ ????????, whether the collision is smaller than yourself. If yes, reject the application for drone testing and go to step 5; otherwise, reject the application for drone testing and wait for ???????????????????? After a long period of time, perform step two; if there is no solution and the reason is the capacity problem, temporarily shill the fixed capacity of the route corresponding to the overcapacity to 0, and then perform step two; other situations are recommended????? ????????? step 2 after the duration;
Step 5 : If the timeout constraints are met, feedback the UAV waypoint and take-off time; otherwise, it is recommended to perform step 2 after the duration of????????????????????
-
Conflict detection and flow control algorithm
Step 1 : Obtain the estimated time ????max and obstacle avoidance distance ????a;
Step 2 : Update the airway network information ????, the current time ????, the information of all drones that have passed the takeoff authorization????????, ???? ∈ ????active. Perform conflict detection on drones that have passed the takeoff authorization. If there is a conflict drone, output the conflict drone????????, collision and possible conflict time; otherwise, go to step 5;
Step 3 : Solve the optimization problem;
Step 4 : If there is a solution, output the new speed of the UAV on the current route V n????,????????????−1,n????,????? ???????,new; if there is no solution, the output event will stimulate the UAV to start its own anti-collision algorithm;
Step 5 : After the interval of ????max, go to step 2.
-
Exception handling algorithm
Consider the conflicts caused by abnormal weather and traffic control (external factors) for diversion.
Step 1 : Obtain air route network information ????, affected air route ???? ban, affected node ???? ban, all drone plan information that has passed takeoff authorization??????? ? ,???? ∈ ????active;
Step 2 : Remove the affected route ???ban and node ???ban, and update the route network information ???;
Step 3 : By comparing the plan information????????, ???? ∈ ????active waypoint ℎ????, filter out the set of affected drones???? ban, if????ban = ∅, the output has no effect and the program is terminated; otherwise, step 4 is executed;
Step 4 : Apply Dijkstra's algorithm to the affected drones, output new waypoints ℎ????, ???? ∈ ????ban and update????????, ???? ∈ ????ban.
3.2.2 Distributed Air Traffic Control
If a certain aircraft in the centralized air traffic control system needs to change its flight plan, all related UAVs need to be re-planned online and the flight plan updated, and the planning complexity increases with the increase of aircraft. Therefore, the computational complexity of the centralized framework is too high, and we hope to have another framework. Just like driving a car, we need to navigate where to go, and the map tells us how to get from point a to point b. This plan is that the map is designed for us before we drive. For an aircraft, the system will determine the flight plan based on the air traffic before takeoff, but once it takes off, the aircraft autonomously decides what to do. This is a distributed overall framework. The distributed framework transfers a lot of control from the ground station to the aircraft. Each aircraft takes care of itself and is organized as a whole. However, during the flight, the aircraft will avoid obstacles with other aircraft in accordance with certain agreements. In this part, we propose a concept called Sky Highway. We have a paper " Sky Highway Design for Dense Traffic " that briefly explains our whole idea. Interested students can take a look.
Regarding route obstacle avoidance, the aircraft can directly perform some obstacle avoidance flights on the route. In order to increase the bandwidth of the entire route network, we have made some designs at the node. For example, this node is to change the direction, and we also hope that the aircraft It can pass directly, so there is no need to wait. If multiple waterways intersect, this is called an intersection node, or intersection. Usually when we pass an intersection, the most common traffic light is the traffic light, but the traffic light means that the plane has to wait here. So at present we have adopted a round-island structure to deal with inefficient waiting strategies like traffic lights. Then control, we generally carry out the idea like the artificial potential field method to ensure that the drone will fly forward in the channel without being stuck. The artificial potential field method has some shortcomings, which may lead to stuck. For example, if we all go to a point, it may be that no one can reach this point. Everyone wants to reach it, but cannot reach at the same time. This is the stuck problem. We have solved these problems now. Then there is the roundabout. We have done some detailed research. The aircraft can enter the channel and go out smoothly. This roundabout design can be regarded as our innovation.
3.3 Simulation and experiment
Finally, let’s introduce our simulation. We built a simulation environment of MATLAB by ourselves, which contains the information of the airway network, the information of the drone to be reviewed, the information of the no-fly zone entered, etc., as shown in the following figure:
We have also built some corresponding platforms in the laboratory, using this positioning facility to do it, as shown in the following figure:
4. Summary and Outlook
UAV air traffic or urban air movement is the general trend. Wireless networks and space-time big data are the basis of traffic. At the same time, traffic design also places new demands on the network. We have done a series of work on the air highway:
-
Designed a route network model and a UAV model, a centralized low-altitude traffic control algorithm, and the most important distributed-based low-altitude traffic control algorithm, which increased traffic while ensuring the safe flight of UAVs.
-
A simulation and experimental platform was built, and the feasibility of a centralized low-altitude traffic control algorithm and a distributed low-altitude traffic control algorithm was verified through use case tests.
In the future, we hope to continue to work in the following areas:
-
Improve the efficiency of flight state estimation algorithm.
-
Efficient dispatch at the airport.
-
The scheduling algorithm of fixed-wing aircraft.
-
Robust scheduling algorithm for heterogeneous systems (hybrid airspace of rotorcraft and fixed wing).
-
Development of semi-physical simulation ATC test system.
-
Flight verification based on real scenarios.
---------- END ----------
Job Offers
The Information Security Department of Meituan is hiring risk control strategy engineers, senior product managers (risk control strategies), mobile security engineers/experts, system security engineers, cloud native security engineers/experts, blue army experts/senior blue army experts, business blue forces Expert/senior business blue army expert, account security expert, front-end architect/expert, Java development engineer, senior Java engineer, senior security management engineer, security integrated solution expert, data security expert, product manager-anti-climbing direction, audit strategy Management, intrusion countermeasure expert, data security product expert, data mining senior expert, big data research and development engineer, big data senior architect, deep learning algorithm engineer, senior data mining/expert (anti-cheating algorithm direction) and other positions.
Welcome to send your resume to [email protected]. Please indicate the subject of the email (intentional post-city-Meituan technical team official account).
Maybe you still want to watch
| Good news! Meituan-Tsinghua big data course is open to the public!
| Meituan OCTO trillion-level data center computing engine technology analysis