CN116797043A - Auxiliary intelligent decision method for airspace running performance evaluation and sector division - Google Patents
Auxiliary intelligent decision method for airspace running performance evaluation and sector division Download PDFInfo
- Publication number
- CN116797043A CN116797043A CN202310675279.2A CN202310675279A CN116797043A CN 116797043 A CN116797043 A CN 116797043A CN 202310675279 A CN202310675279 A CN 202310675279A CN 116797043 A CN116797043 A CN 116797043A
- Authority
- CN
- China
- Prior art keywords
- aircraft
- airspace
- data
- sector
- state
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 72
- 238000011156 evaluation Methods 0.000 title claims abstract description 26
- 238000004088 simulation Methods 0.000 claims abstract description 36
- 238000013439 planning Methods 0.000 claims abstract description 17
- 230000006399 behavior Effects 0.000 claims abstract description 16
- 238000012549 training Methods 0.000 claims abstract description 16
- 238000010801 machine learning Methods 0.000 claims abstract description 12
- 230000006870 function Effects 0.000 claims description 24
- 210000002569 neuron Anatomy 0.000 claims description 24
- 230000009471 action Effects 0.000 claims description 18
- 230000008859 change Effects 0.000 claims description 18
- 230000008569 process Effects 0.000 claims description 18
- 239000011159 matrix material Substances 0.000 claims description 15
- 230000004913 activation Effects 0.000 claims description 12
- 230000007704 transition Effects 0.000 claims description 10
- 238000012544 monitoring process Methods 0.000 claims description 9
- 238000005516 engineering process Methods 0.000 claims description 7
- 238000005457 optimization Methods 0.000 claims description 7
- 230000001105 regulatory effect Effects 0.000 claims description 7
- 102000005717 Myeloma Proteins Human genes 0.000 claims description 6
- 108010045503 Myeloma Proteins Proteins 0.000 claims description 6
- 102100029469 WD repeat and HMG-box DNA-binding protein 1 Human genes 0.000 claims description 6
- 101710097421 WD repeat and HMG-box DNA-binding protein 1 Proteins 0.000 claims description 6
- 230000008878 coupling Effects 0.000 claims description 6
- 238000010168 coupling process Methods 0.000 claims description 6
- 238000005859 coupling reaction Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 6
- 230000003068 static effect Effects 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 5
- 239000000203 mixture Substances 0.000 claims description 4
- 230000001133 acceleration Effects 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 claims description 3
- 230000001174 ascending effect Effects 0.000 claims description 3
- 230000000694 effects Effects 0.000 claims description 3
- 239000000446 fuel Substances 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000007619 statistical method Methods 0.000 claims description 3
- 238000002945 steepest descent method Methods 0.000 claims description 3
- 238000000547 structure data Methods 0.000 claims description 3
- 238000012876 topography Methods 0.000 claims description 3
- 230000009194 climbing Effects 0.000 claims description 2
- 238000011161 development Methods 0.000 description 5
- 238000007726 management method Methods 0.000 description 4
- 238000007418 data mining Methods 0.000 description 3
- 230000002787 reinforcement Effects 0.000 description 3
- 230000001186 cumulative effect Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000007405 data analysis Methods 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
Landscapes
- Traffic Control Systems (AREA)
Abstract
The invention relates to an auxiliary intelligent decision method for airspace running performance evaluation and sector planning, which comprises the following steps: step one, constructing an empty pipe basic data platform and collecting data from the empty pipe basic data platform; training an AI controller according to the selected reference airspace, classifying command behavior characteristics of the controller by adopting a machine learning algorithm, and autonomously generating simulated flight plans of different categories according to classification results; inputting a simulated flight plan into an empty pipe simulation system, and autonomously implementing intelligent aircraft state adjustment by combining the trained 'AI controllers' and generating a virtual workload; and step four, based on the virtual workload obtained by statistics, applying a machine learning algorithm to implement sector division. The empty pipe simulation and airspace planning method is high in accuracy, high in efficiency and good in generalization.
Description
Technical Field
The invention relates to a flight plan generation, air management simulation and sector planning method, in particular to an auxiliary intelligent decision method for airspace operation performance evaluation and sector planning based on big data.
Background
With the rapid development of national economy, the requirements of users in the military aviation domain on airspace resources are vigorous, the air traffic flow is continuously and rapidly increased, the airspace structure and the using mode tend to be complex and diversified, the airspace operation pressure is continuously increased, the contradiction between the requirements and the resource supply is increasingly apparent, and the system becomes a key bottleneck for restricting the development of civil aviation. Therefore, starting from rich technical means, the effective improvement of scientificity and rationality of airspace planning and management has become urgent matters for guaranteeing the healthy development and safe operation of civil aviation. At present, the airspace planning and management work mainly takes experience as a main part, and combines a computer auxiliary means to carry out simulation evaluation on a planning scheme. However, the simulation evaluation method of the simulator still depends too much on human experience in the aspects of training planning, airspace scheme comparison and the like, and accordingly, the airspace planning, simulation and evaluation results lack scientificity, authenticity and authority.
The development of machine learning technology enables air-space planning and management to have the condition of transition from extensive decision making based on experience to refined decision making based on mass data; the real-time simulation evaluation work of the airspace needs to be carried out by means of a data analysis means, a scientific real-time simulation evaluation method of the airspace is provided, the accurate simulation of the operation of the airspace is completed, the requirements of the airspace are accurately reflected, the optimization of the airspace structure is realized, the safety capacity is refined and quantized, the operation efficiency is strengthened and improved, the bottleneck of the airspace is broken, and the sustainable and healthy development of civil aviation is ensured.
Disclosure of Invention
The invention aims to provide the empty pipe simulation and airspace planning method which is higher in accuracy, higher in efficiency and better in generalization.
The technical scheme for realizing the aim of the invention is to provide an auxiliary intelligent decision method for airspace operation performance evaluation and sector division, which comprises the following steps:
firstly, constructing an air-traffic control basic data platform and collecting monitoring data and airspace structure data of each aircraft from the air-traffic control basic data platform;
step two, according to the selected reference airspace, carrying out track data preprocessing to obtain the running track of the aircraft in the airspace, further extracting the state change condition of the aircraft in the airspace to describe the command behavior characteristics of an air traffic controller and train the AI controller, classifying the command behavior characteristics of the controller by adopting a machine learning algorithm, and generating simulated flight plans of different categories independently according to the classification result;
inputting a simulated flight plan into an empty pipe simulation system, and autonomously implementing intelligent aircraft state adjustment by combining the trained 'AI controllers' and generating a virtual workload;
and step four, based on the virtual workload obtained by statistics, applying a machine learning algorithm to implement sector division.
Further, the first step specifically includes the following steps:
1.1, developing various data connection interfaces based on the operation multi-source data of an aircraft; the data types comprise aircraft monitoring data, control voice data, airspace information and weather information, and a basic data source is provided for airspace simulation;
1.2, aiming at various empty pipe operation data, forming various data into normalized unified representation according to data types.
Further, the second step specifically includes the following steps:
2.1 the raw aircraft trajectory data set contains various noise and missing data points, for a selected airspace, aircraft that does not pass through that airspace are first filtered out, in Ω and Θ i Respectively representing the connection line formed by the space range of the reference airspace and the track point of the ith aircraft, ifThen the ith aircraft is deleted;
2.2 the total number n of trajectory points of the ith aircraft in airspace Ω i <5, then delete the ith aircraft;
2.3, according to the historical flight information of the aircraft, screening out the change time and adjustment quantity values of all state quantities of the aircraft, namely the altitude H, the heading H and the speed v, and splicing the state change time period and the state unchanged time period to form the historical flight state sequence of the aircraft. If the altitude, heading and speed of the ith aircraft at time t and (t+1) are H respectively t And H t+1 ，h t And h t+1 And v t And v t+1 List 1 respectivelyClimbing indicator (H) t -H t+1 <o), right turn (h) t -h t+1 <o) and acceleration (v) t -υ t+1 <o) with 0 indicating the height invariance (H) t -H t+1 O), heading invariance (h t -h t+1 =o) and velocity invariance (v) t -υ t+1 Let-1 denote the drop (H) t -H t+1 >o), turn left (h t -h t+1 >o) and deceleration (v) t -υ t+1 >o), thereby converting the state change process of each aircraft in the specific space domain into codes comprising a series of numbers 0,1 and-1, and matching the track data and the weather information with the voice recognition data by combining the control voice data and the weather information;
2.4, acquiring a coupling relation between an initial state parameter and a state sequence of the aircraft in the airspace when the aircraft enters the airspace and a coupling relation between the initial state parameter and time taken by the aircraft to fly through the airspace when the aircraft enters the airspace by using matched aircraft operation data aiming at a specific airspace based on a supervised learning method.
Further, step 2.3 specifically includes the following steps:
2.3.1 dividing a specific airspace into grids of 1km by 1km, and regarding the control voice data and the weather information together as a hidden state sequencem is the maximum value of the number of the regulated voice data and the number of the weather information dataThe hidden state sequence is an observation value sequence generated by the hidden state sequence, and because the number of track points of each aircraft is different, the aircraft tracks are clustered based on a dynamic time alignment method and a k-nearest neighbor algorithm, and the dynamic time alignment method can measure the similarity between time sequence data with different lengths; for lengths of +.>And M' are two time series dataAnd->The dynamic time alignment method can be applied to obtain +.>Distance matrix of dimensions, wherein element (i ', J') represents a point +.>And (4) point->Distance between->Further, in order to acquire the optimal regular path, the corrected cumulative distance γ (i ', J') is described as follows:
2.3.2 reamCoordinate set representing intersection of trajectory of aircraft with discrete grid points,/->Representing the number of coordinates of the discrete track points, elements in the a priori matrix A +.>Representation->And->Transition probability between, element +.>Representation and->Associated regulated speech data and weather information parameters, element +.pi.in initial distribution probability ∈>The initial distribution probability of the hidden state is represented, and the optimal state sequence delta is obtained by adopting the following method to complete the initial matching process:
further, step 2.4 specifically includes the following steps:
2.4.1 static and dynamic State parameters such as the model (x 1 ) Speed (x) 2 ) Temperature (x) 3 ) Wind speed (x) 4 ) Wind direction (x) 5 ) Height (x) 6 ) And heading (x) 7 ) All are regarded as input variables, the state change process of the aircraft in the sector, namely 0,1 and-1 sequences are regarded as output variables (y), so that a multi-layer perceptron model is designed, an AI controller is constructed, and the AI controller module is embedded into an airspace simulation evaluation system; constructing a multi-layer perceptron model comprising L layers, and setting the number of neurons X in a first layer as N l That is to sayAnd the activation function of each neuron is f l Neurons in layer l receive signals from neurons in layer (l-1) above, neuron +.>From->Receives signals and the connection weight between them is +.>From which one dimension is N l-1 *N l Weight matrix W of (2) l Wherein the element is ∈ ->Furthermore, the->Also comprises 1 deviation->And its activation amount is The deviation of (2) is +.>Use->Representing network variables associated with neurons:
thus (2)
Taking the initial input layer as layer 0 input, if the input vector x contains N components, then the input layerN neurons possess activation amountsThe L layer of the network is the output layer, if the output vector y contains M components, each component y of the output vector g Can be expressed as +.>Given the connection weight and the offset of the network and the input vector, the activation amount of each neuron can be calculated using the above formula;
based on the generated aircraft trajectory pattern pair (ζ q′ ，T q′ )(q′＝1，2，...，χ′)，ξ q′ And T q′ Respectively representing the q 'th training vector and the target value, χ' represents the total track number, and the training vector ζ q′ Comprising N components and their target values T q′ Comprising M components, in the t 'th training step, if the input vector x (t')=ζ q′ Then the corresponding network output value y (T ') can be obtained according to the above formula, so that the mean square error E= |y (T ') -T is caused to be the same as the mean square error E= |y (T ') -T q′ || 2 To minimize the mean square error, the weight and bias values of the network are updated using the steepest descent method as follows:
where α' represents the learning rate.
2.4.2 static and dynamic State parameters such as the model (x 1 ) Speed (x) 2 ) Temperature (x) 3 ) Wind speed (x) 4 ) Wind direction (x) 5 ) Height (x) 6 ) And heading (x) 7 ) All are regarded as input variables, the flight time of the aircraft in the air space is regarded as output variables (y'), and the same model structure and algorithm as those in the step 2.4.1 are adopted, so that different types of simulated flight plans are generated independently according to the modeling analysis result and by combining with the aircraft state sequence;
setting eachThe total number of state change sequences for aircraft is p and k (k < Z) flight plans must be generated based on these aircraft, Z being the maximum number of flight plans, the first of the state sequence for aircraft iThe variables are defined asDistance between aircraft i and aircraft J state sequence values +.>Called "edit distance", i.e. the sum of the number of times the variable is added, the number of times the variable is subtracted and the number of times the variable is replaced, which needs to be implemented on the original sequence in order to make the two sequences of values identical; the following methods are used to generate simulated flight plans of different categories:
step 2.4.2.1: selecting an initial center point:
2.4.2.1-1: calculating an edit distance between pairs of different aircraft state adjustment value sequences;
2.4.2.1-2: acquiring for aircraft j its corresponding value in the form
2.4.2.1-3: arranging all aircraft in ascending orderValue and let k minimum +.>The aircraft corresponding to the value is set as an initial center point;
2.4.2.1-4: acquiring an initial clustering result by matching each aircraft to a center point nearest to the aircraft;
2.4.2.1-5: acquiring the sum of the distances of all the aircrafts from the central point of the aircrafts;
step 2.4.2.2: updating the center point:
aiming at the initial clustering result, taking the total distance between the minimum aircraft in the cluster and the central point corresponding to the cluster as an objective function, updating the central point and replacing the original central point with the new central point;
step 2.4.2.3: allocating aircraft to respective central points
2.4.2.3-1: assigning each aircraft to a center point nearest to the aircraft and acquiring a clustering result;
2.4.2.3-2: and calculating the sum of the distances of all the aircrafts from the central point of the aircrafts, stopping the algorithm if the sum is the same as the previous value, and otherwise returning to the step 2.4.2.2.
Further, the third step specifically includes the following steps:
firstly, leading airspace data and topography data from an empty pipe basic data platform for airspace simulation and planning, and associating a sector division intelligent auxiliary decision module with an airspace simulation evaluation system;
3.2, leading the flight performance data, the monitoring data, the flight plan data, the control voice data, the airspace information, the airport information and the weather information data to perform airspace microscopic simulation;
3.2.1, acquiring aircraft flow distribution probability by adopting a maximum entropy theory, and processing and analyzing actual flight data to jointly form initial input data of the simulation aircraft flow module by combining aircraft performance and control rule conditions;
3.2.2"AI controllers" issue altitude, speed or heading commands for the aircraft based on real-time air traffic dynamics, while taking command time and content information as "virtual workload" for statistical analysis.
Further, the fourth step specifically includes the following steps:
4.1, analyzing the clustering result of the aircraft track data reflecting the control behavior, combining the control sector handover protocol and the operation mode, and taking the fuzzy comprehensive decision result in each airspace state as the prior knowledge of Q learning by using a fuzzy comprehensive decision method, wherein the number of the sectors which can be processed simultaneously is 10 by using a machine learning-based sector optimization dividing technology;
4.1.1 constructing a decision set V and a factor set U of a fuzzy comprehensive decision method, setting an action set Xi and a state set lambda of a Q learning algorithm, wherein the factor sets are as follows: the area of the airspace subarea, the workload in the airspace subarea, the shape of the airspace subarea and the space connection attribute of the route and the route, and the decision set is an action set selected by the airspace subarea system;
4.1.2 referring to expert experience, a fuzzy evaluation matrix R and a weight set Θ are constructed. Method for calculating all states in state set lambda by utilizing fuzzy comprehensive decisionThe superiority vector of taking the respective action>
4.1.3 the respective statesLower->Performing normalization adjustment and taking the normalization adjustment as prior knowledge of Q learning;
4.2, respectively designing a continuous return function and a discrete return function based on the air traffic behavior region distribution characteristics and the air traffic behavior recognition and based on the control handover activities, wherein the composition of the continuous return function comprises the flight section length, the fuel economy and the maneuverability demand factors of the aircraft; to ensure flight safety, a discrete penalty term, i.e., a discrete return function, is applied for the process of causing a dangerous condition or exceeding the sector capacity state transition;
4.3 generating an initial Q matrix according to all the determined states and action sets, if the total number of sectors is 10, giving the current state of the system at the time tReinforced learning rate alpha t And a prize value discount factor beta, if the 10 sector selection actions are +.>They will transition to the new system state +.>And each receives a prize valueAfter all the handover location points or sub-areas perform the actions, a new sector configuration result is reached, so that the Q value of each sector at the (t+1) moment is updated:
and->Respectively represent sector system status->Sector +.>And->Cost value paid,/->Representing status->10 Nash equalization strategies, respectively adopted by 10 sectors in the case of +.>And->Respectively indicate status->Sector +.>And->Is a function of the state value of (c),and->Respectively indicate (t+1) time status +.>Sector +.>And->State-behavior value functions of (2); and after the Q value is stable, finishing the training process of the model to obtain a sector division result.
The invention has the positive effects that: (1) The invention introduces a flight path big data mining technology in the step two, which can provide more scientific basis for random aircraft flow generation in flight plan formulation, flight running time setting and other parameters, and can further improve the accuracy of flight preset flight time. (2) The existing various simulation software mainly presets a state adjustment strategy of the aircraft in the simulation process before the simulation from the angle of a control rule, the dynamic adaptability is not realized, and in addition, the controller command rule of a specific airspace cannot be fully excavated based on a flight history track to provide a reference for flight fine simulation; (3) Traditional sector design schemes based on mathematical optimization methods or heuristic methods are used for implementing modeling processes from scratch according to each sector design requirement, and the modeling processes need complete various information. When the sector division scene changes, the re-modeling and the re-optimization are needed, and the constructed division models have no reusability. The reinforcement learning is used as a model-free algorithm driven by data training, and after model training is finished, the reinforcement learning can implement self-adaption and autonomous learning aiming at various scenes, so that an optimal optimization scheme is obtained, and the generalization of the model is better. In contrast, the sector division method based on the reinforcement learning technology provided in the step four of the invention has better generalization, lower requirements on the completeness of the modeled data, lower complexity of the modeling process, and capability of rapidly obtaining sector division results aiming at different airspace structures, improving airspace planning efficiency and being capable of adaptively solving the problem of different sector division.
Detailed Description
Example 1
The auxiliary intelligent decision method for airspace operation performance evaluation and sector planning in the embodiment comprises the following steps:
firstly, constructing an air-traffic control basic data platform and collecting monitoring data and airspace structure data of each aircraft from the air-traffic control basic data platform; the method specifically comprises the following steps:
1.1 constructing an airspace simulation track data processing technology based on the running multisource data of an aircraft, simultaneously processing 200 track pairs, developing various data leading interfaces for each pair of tracks comprising a boeing flight and an air passenger flight, and designing an airspace basic big data mining platform and associating the airspace basic big data mining platform with an airspace simulation evaluation system. The data types comprise aircraft monitoring data, control voice data, airspace information, weather information and the like, and a basic data source is provided for airspace simulation;
1.2, aiming at various empty pipe operation data, forming various data into normalized unified representation according to data types.
Step two, according to the selected reference airspace, carrying out track data preprocessing to obtain the running track of the aircraft in the airspace, further extracting the state change condition of the aircraft in the airspace to describe the command behavior characteristics of an air traffic controller and train the AI controller, classifying the command behavior characteristics of the controller by adopting a machine learning algorithm, and generating simulated flight plans of different categories independently according to the classification result; the method specifically comprises the following steps:
2.1 the raw aircraft trajectory data set contains various noise and missing data points, for a selected airspace, aircraft that does not pass through that airspace are first filtered out, in Ω and Θ i Respectively representing the connection line formed by the space range of the reference airspace and the track point of the ith aircraft, ifThen the ith aircraft is deleted;
2.2 the total number n of trajectory points of the ith aircraft in airspace Ω i <5, then delete the ith aircraft;
2.3, according to the historical flight information of the aircraft, screening out the change time and the adjustment quantity value of each state quantity (height H, heading H and speed v) of the aircraft, and splicing the state change time period and the state unchanged time period to form the historical flight state sequence of the aircraft. Altitude, heading and speed of the ith aircraft at times t and (t+1), respectivelyIs H t And H t+1 ，h t And h t+1 And v t And v t+1 Climbing (H) is denoted by 1 t -H t+1 < o), right turn (h t -h t+1 < o) and acceleration (v) t -υ t+1 < o), each of 0 is denoted as highly invariant (H t -H t+1 O), heading invariance (h t -h t+1 =o) and velocity invariance (v) t -υ t+1 Let-1 denote the drop (H) t -H t+1 > o), turn left (h t -h t+1 > o) and deceleration (v t -υ t+1 O), thereby converting the state change process of each aircraft in a specific space domain into a code comprising a series of numbers 0,1 and-1, and matching the flight path data and the weather information with the voice recognition data in combination with the regulated voice data and the weather information;
2.3.1 dividing a specific airspace into grids of 1km by 1km, and regarding the control voice data and the weather information together as a hidden state sequencem is the maximum value of the number of the regulated voice data and the number of the weather information dataThe hidden state sequence is an observation value sequence generated by the hidden state sequence, and because the number of track points of each aircraft is different, the aircraft tracks are clustered based on a dynamic time alignment method and a k-nearest neighbor algorithm, and the dynamic time alignment method can measure the similarity between time sequence data with different lengths; for lengths of +.>And M' are two time series dataAnd->Using dynamic time warpingWhole method can obtain->Distance matrix of dimensions, wherein element (i ', J') represents a point +.>And (4) point->Distance between->Further, in order to acquire the optimal regular path, the corrected cumulative distance γ (i ', J') is described as follows:
2.3.2 reamCoordinate set representing intersection of trajectory of aircraft with discrete grid points,/->Representing the number of coordinates of the discrete track points, elements in the a priori matrix A +.>Representation->And->Transition probability between, element +.>Representation and->Associated regulated speech data and weather information parameters, element +.pi.in initial distribution probability ∈>The initial distribution probability of the hidden state is represented, and the optimal state sequence delta is obtained by adopting the following method to complete the initial matching process:
2.4, acquiring a coupling relation between an initial state parameter and a state sequence of the aircraft in the airspace when the aircraft enters the airspace and a coupling relation between the initial state parameter and time for the aircraft to fly through the airspace when the aircraft enters the airspace by using matched aircraft operation data according to a supervised learning method;
2.4.1 static and dynamic State parameters such as the model (x 1 ) Speed (x) 2 ) Temperature (x) 3 ) Wind speed (x) 4 ) Wind direction (x) 5 ) Height (x) 6 ) And heading (x) 7 ) All are regarded as input variables, the state change process of the aircraft in the sector, namely 0,1 and-1 sequences are regarded as output variables (y), so that a multi-layer perceptron model is designed, an AI controller is constructed, and the AI controller module is embedded into an airspace simulation evaluation system; constructing a multi-layer perceptron model comprising L layers, and setting the number of neurons X in a first layer as N l That is to sayAnd the activation function of each neuron is f l Neurons in layer l receive signals from neurons in the preceding layer (l-1), such as neurons +.>From the slaveReceives signals and the connection weight between them is +.>From which one dimension is N l-1 *N l Weight matrix W of (2) l Wherein the element is ∈ ->Furthermore, the->Also comprises 1 deviation->And its activation amount is +.> The deviation of (2) is +.>Use->Representing network variables associated with neurons:
thus (2)
Taking the initial input layer as the layer 0 input, if the input vector x comprises N components, N neurons in the input layer have activation amountsThe L layer of the network is the output layer, if the output vector y contains M components, each component y of the output vector g Can be expressed as +.>Given the connection weight and the offset of the network and the input vector, the activation amount of each neuron can be calculated using the above formula;
based on the generated aircraft trajectory pattern pair (ζ q′ ，T q′ )(q′＝1，2，...，χ′)，ξ q′ And T q′ Respectively representing the q 'th training vector and the target value, χ' represents the total track number, and the training vector ζ q′ Comprising N components and their target values T q′ Comprising M components, in the t 'th training step, if the input vector x (t')=ζ q′ Then the corresponding network output value y (T ') can be obtained according to the above formula, so that the mean square error E= |y (T ') -T is caused to be the same as the mean square error E= |y (T ') -T q′ || 2 To minimize the mean square error, the weight and bias values of the network are updated using the steepest descent method as follows:
where α' represents the learning rate.
2.4.2 static and dynamic State parameters such as the model (x 1 ) Speed (x) 2 ) Temperature (x) 3 ) Wind speed (x) 4 ) Wind direction (x) 5 ) Height (x) 6 ) And heading (x) 7 ) All are regarded as input variables, the flight time of the aircraft in the air space is regarded as output variables (y'), and the same model structure and algorithm as those in the step 2.4.1 are adopted, so that different types of simulated flight plans are generated independently according to the modeling analysis result and by combining with the aircraft state sequence;
setting the total number of state change sequences of each aircraft to be p andgenerating k (k < Z) flight plans based on the aircraft, Z being the maximum of the number of flight plans, and sequencing the state of aircraft iThe variables are defined as X ij (i＝1，2，...，Z；/>) Distance between aircraft i and aircraft J state sequence values +.>Called "edit distance", i.e. the sum of the number of times the variable is added, the number of times the variable is subtracted and the number of times the variable is replaced, which needs to be implemented on the original sequence in order to make the two sequences of values identical; the following methods are used to generate simulated flight plans of different categories:
step 2.4.2.1: selecting an initial center point:
2.4.2.1-1: calculating an edit distance between pairs of different aircraft state adjustment value sequences;
2.4.2.1-2: acquiring for aircraft j its corresponding value in the form
2.4.2.1-3: arranging all aircraft in ascending orderValue and let k minimum +.>The aircraft corresponding to the value is set as an initial center point;
2.4.2.1-4: acquiring an initial clustering result by matching each aircraft to a center point nearest to the aircraft;
2.4.2.1-5: acquiring the sum of the distances of all the aircrafts from the central point of the aircrafts;
step 2.4.2.2: updating the center point:
aiming at the initial clustering result, taking the total distance between the minimum aircraft in the cluster and the central point corresponding to the cluster as an objective function, updating the central point and replacing the original central point with the new central point;
step 2.4.2.3: allocating aircraft to respective central points
2.4.2.3-1: assigning each aircraft to a center point nearest to the aircraft and acquiring a clustering result;
2.4.2.3-2: and calculating the sum of the distances of all the aircrafts from the central point of the aircrafts, stopping the algorithm if the sum is the same as the previous value, and otherwise returning to the step 2.
Inputting a simulated flight plan into an empty pipe simulation system, and autonomously implementing intelligent aircraft state adjustment by combining the trained 'AI controllers' and generating a virtual workload; the method specifically comprises the following steps:
firstly, leading airspace data and topography data from an empty pipe basic data platform for airspace simulation and planning, and associating a sector division intelligent auxiliary decision module with an airspace simulation evaluation system;
3.2, leading in data such as flight performance data, monitoring data, flight plan data, control voice data, airspace information, airport information, air-phase information and the like, and carrying out airspace microscopic simulation;
3.2.1, acquiring aircraft flow distribution probability by adopting a maximum entropy theory, and processing and analyzing actual flight data to jointly form initial input data of the simulation aircraft flow module by combining the conditions of aircraft performance, control rules and the like;
3.2.2"AI controller" according to the real-time air traffic dynamics, issue various instructions such as altitude, speed or heading adjustment to the aircraft, send information such as time and content of instruction as "virtual workload" to carry on the statistical analysis at the same time;
and step four, based on the virtual workload obtained by statistics, applying a machine learning algorithm to implement sector division. The method specifically comprises the following steps:
4.1, analyzing the clustering result of the aircraft track data reflecting the control behavior, combining the control sector handover protocol and the operation mode, and taking the fuzzy comprehensive decision result in each airspace state as the prior knowledge of Q learning by using a fuzzy comprehensive decision method, wherein the number of the sectors which can be processed simultaneously is 10 by using a machine learning-based sector optimization dividing technology;
4.1.1 constructing a decision set V and a factor set U of a fuzzy comprehensive decision method, setting an action set Xi and a state set lambda of a Q learning algorithm, wherein the factor sets are as follows: the area of the air domain subarea, the workload in the air domain subarea, the shape of the air domain subarea, the space connection attribute of the air route and the like, and the decision set is an action set selected by the air domain subarea system;
4.1.2 referring to expert experience, a fuzzy evaluation matrix R and a weight set Θ are constructed. Method for calculating all states in state set lambda by utilizing fuzzy comprehensive decisionThe superiority vector of taking the respective action>
4.1.3 the respective statesLower->Performing normalization adjustment and taking the normalization adjustment as prior knowledge of Q learning;
4.2, respectively designing a continuous return function and a discrete return function based on the air traffic behavior region distribution characteristics and the air traffic behavior recognition and based on the control handover activities, wherein the composition of the continuous return function comprises factors such as the length of an aircraft flight section, the fuel economy, the maneuverability requirement and the like; in order to ensure flight safety, discrete penalty items, namely a discrete return function, are applied to the state transition processes such as dangerous conditions or exceeding sector capacity and the like;
4.3 generating an initial Q matrix according to all the determined states and action sets, if the total number of sectors is 10, giving the current state of the system at the time tReinforced learning rate alpha t And a prize value discount factor beta, if the 10 sector selection actions are +.>They will transition to the new system state +.>And each receives a prize valueAfter all the handover location points or sub-areas perform the actions, a new sector configuration result is reached, so that the Q value of each sector at the (t+1) moment is updated:
and->Respectively represent sector system status->Sector +.>And->Cost value paid,/->Representing status->10 Nash equalization strategies, respectively adopted by 10 sectors in the case of +.>And->Respectively indicate status->Sector +.>And->Is a function of the state value of (c),and->Respectively indicate (t+1) time status +.>Sector +.>And->State-behavior value functions of (2); and after the Q value is stable, finishing the training process of the model to obtain a sector division result.
It is apparent that the above examples are merely illustrative of the present invention and are not limiting of the embodiments of the present invention. Other variations or modifications of the above teachings will be apparent to those of ordinary skill in the art. It is not necessary here nor is it exhaustive of all embodiments. While remaining within the scope of the invention, obvious variations or modifications are incorporated by reference herein.
Claims (7)
1. The auxiliary intelligent decision method for airspace operation performance evaluation and sector planning is characterized by comprising the following steps:
firstly, constructing an air-traffic control basic data platform and collecting monitoring data and airspace structure data of each aircraft from the air-traffic control basic data platform;
step two, according to the selected reference airspace, carrying out track data preprocessing to obtain the running track of the aircraft in the airspace, further extracting the state change condition of the aircraft in the airspace to describe the command behavior characteristics of an air traffic controller and train the AI controller, classifying the command behavior characteristics of the controller by adopting a machine learning algorithm, and generating simulated flight plans of different categories independently according to the classification result;
inputting a simulated flight plan into an empty pipe simulation system, and autonomously implementing intelligent aircraft state adjustment by combining the trained 'AI controllers' and generating a virtual workload;
and step four, based on the virtual workload obtained by statistics, applying a machine learning algorithm to implement sector division.
2. The airspace operation performance evaluation and sector assignment assisting intelligent decision-making method according to claim 1, wherein the method comprises the following steps of: the first step specifically comprises the following steps:
1.1, developing various data connection interfaces based on the operation multi-source data of an aircraft; the data types comprise aircraft monitoring data, control voice data, airspace information and weather information, and a basic data source is provided for airspace simulation;
1.2, aiming at various empty pipe operation data, forming various data into normalized unified representation according to data types.
3. The airspace operation performance evaluation and sector assignment assisting intelligent decision-making method according to claim 1, wherein the method comprises the following steps of: the second step specifically comprises the following steps:
2.1 the raw aircraft trajectory data set contains various noise and missing data points, for a selected airspace, aircraft that does not pass through that airspace are first filtered out, in Ω and Θ i Respectively representing the connection line formed by the space range of the reference airspace and the track point of the ith aircraft, ifThen the ith aircraft is deleted;
2.2 the total number n of trajectory points of the ith aircraft in airspace Ω i <5, then delete the ith aircraft;
2.3, according to the historical flight information of the aircraft, screening out the change time and adjustment quantity values of all state quantities of the aircraft, namely the altitude H, the heading H and the speed v, and splicing the state change time period and the state unchanged time period to form the historical flight state sequence of the aircraft. If the altitude, heading and speed of the ith aircraft at time t and (t+1) are H respectively t And H t+1 ，h t And h t+1 And v t And v t+1 Climbing (H) is denoted by 1 t -H t+1 <0) Turn right (h) t -h t+1 <o) and acceleration (v) t -υ t+1 <o) with 0 indicating the height invariance (H) t -H t+1 O), heading invariance (h t -h t+1 =o) and velocity invariance (v) t -υ t+1 Let-1 denote the drop (H) t -H t+1 >o), turn left (h t -h t+1 >o) and deceleration (v) t -υ t+1 >o), thereby converting the course of the state change of the individual aircraft in a specific space into a code comprising a series of numbers 0,1 and-1, in combinationControlling voice data and meteorological information, and matching the track data and the meteorological information with voice recognition data;
2.4, acquiring a coupling relation between an initial state parameter and a state sequence of the aircraft in the airspace when the aircraft enters the airspace and a coupling relation between the initial state parameter and time taken by the aircraft to fly through the airspace when the aircraft enters the airspace by using matched aircraft operation data aiming at a specific airspace based on a supervised learning method.
4. A airspace runnability assessment and sector assignment assist intelligent decision method according to claim 3, characterized in that: step 2.3 specifically includes the following steps:
2.3.1 dividing a specific airspace into grids of 1km by 1km, and regarding the control voice data and the weather information together as a hidden state sequencem is the maximum value of the number of the regulated voice data and the number of the weather information dataThe hidden state sequence is an observation value sequence generated by the hidden state sequence, and because the number of track points of each aircraft is different, the aircraft tracks are clustered based on a dynamic time alignment method and a k-nearest neighbor algorithm, and the dynamic time alignment method can measure the similarity between time sequence data with different lengths; for lengths of +.>And M' are two time series dataAnd->The dynamic time alignment method can be applied to obtain +.>Distance matrix of dimensions, wherein element (i ', j') represents a point +.>And (4) point->Distance between->Further, in order to acquire the optimal regular path, the corrected accumulated distance γ (i ', j') is described as follows:
2.3.2 reamCoordinate set representing intersection of trajectory of aircraft with discrete grid points,/->Representing the number of coordinates of the discrete track points, elements in the a priori matrix A +.>Representation->And->Transition probability between, element +.>Representation and->Associated regulated speech data and weather information parameters, element +.pi.in initial distribution probability ∈>The initial distribution probability of the hidden state is represented, and the optimal state sequence delta is obtained by adopting the following method to complete the initial matching process:
5. a airspace runnability assessment and sector assignment assist intelligent decision method according to claim 3, characterized in that: step 2.4 specifically includes the following steps:
2.4.1 static and dynamic State parameters such as the model (x 1 ) Speed (x) 2 ) Temperature (x) 3 ) Wind speed (x) 4 ) Wind direction (x) 5 ) Height (x) 6 ) And heading (x) 7 ) All are regarded as input variables, the state change process of the aircraft in the sector, namely 0,1 and-1 sequences are regarded as output variables (y), so that a multi-layer perceptron model is designed, an AI controller is constructed, and the AI controller module is embedded into an airspace simulation evaluation system; constructing a multi-layer perceptron model comprising L layers, and setting the number of neurons X in a first layer as N l That is to sayAnd the activation function of each neuron is f l Neurons in layer l receive signals from neurons in layer (l-1) above, neuron +.>From->Receives signals and the connection weight between them is +.>From which one dimension is N l-1 *N l Weight matrix W of (2) l Wherein the elements areFurthermore, the->Also comprises 1 deviation->And its activation amount is +.> The deviation of (2) is +.>Use->Representing network variables associated with neurons:
thus (2)
Taking the initial input layer as the layer 0 input, if the input vector x comprises N components, N neurons in the input layer have activation amountsThe L layer of the network is the output layer, if the output vector y contains M components, each component y of the output vector g Can be expressed as +.>Given the connection weight and the offset of the network and the input vector, the activation amount of each neuron can be calculated using the above formula;
based on the generated aircraft trajectory pattern pair (ζ q′ ,T q′ )(q′＝1,2,...,χ′)，ξ q′ And T q′ Respectively representing the q 'th training vector and the target value, χ' represents the total track number, and the training vector ζ q′ Comprising N components and their target values T q′ Comprising M components, in the t 'th training step, if the input vector x (t')=ζ q′ Then the corresponding network output value y (T ') can be obtained according to the above formula, so that the mean square error E= |y (T ') -T is caused to be the same as the mean square error E= |y (T ') -T q′ || 2 To minimize the mean square error, the weight and bias values of the network are updated using the steepest descent method as follows:
where α' represents the learning rate.
2.4.2 static and dynamic State parameters such as the model (x 1 ) Speed (x) 2 ) Temperature (x) 3 ) Wind speed (x) 4 ) Wind direction (x) 5 ) Height (x) 6 ) And heading (x) 7 ) All are taken as input variables, the flight time of the aircraft in the space is taken as output variable (y'), and the same model structure as that of the step 2.4.1 is adoptedThe building algorithm autonomously generates simulated flight plans of different categories according to the modeling analysis result and by combining an aircraft state sequence;
the total number of state change sequences per aircraft is set to p and k (k) must be generated based on these aircraft<Z) flight plans, Z being the maximum number of flight plans, the first of the state sequence of aircraft iThe variables are defined asDistance between state sequence values of aircraft i and aircraft j +.>Called "edit distance", i.e. the sum of the number of times the variable is added, the number of times the variable is subtracted and the number of times the variable is replaced, which needs to be implemented on the original sequence in order to make the two sequences of values identical; the following methods are used to generate simulated flight plans of different categories:
step 2.4.2.1: selecting an initial center point:
2.4.2.1-1: calculating an edit distance between pairs of different aircraft state adjustment value sequences;
2.4.2.1-2: acquiring for aircraft j its corresponding value in the form
2.4.2.1-3: arranging all aircraft in ascending orderValue and let k minimum +.>The aircraft corresponding to the value is set as an initial center point;
2.4.2.1-4: acquiring an initial clustering result by matching each aircraft to a center point nearest to the aircraft;
2.4.2.1-5: acquiring the sum of the distances of all the aircrafts from the central point of the aircrafts;
step 2.4.2.2: updating the center point:
aiming at the initial clustering result, taking the total distance between the minimum aircraft in the cluster and the central point corresponding to the cluster as an objective function, updating the central point and replacing the original central point with the new central point;
step 2.4.2.3: allocating aircraft to respective central points
2.4.2.3-1: assigning each aircraft to a center point nearest to the aircraft and acquiring a clustering result;
2.4.2.3-2: and calculating the sum of the distances of all the aircrafts from the central point of the aircrafts, stopping the algorithm if the sum is the same as the previous value, and otherwise returning to the step 2.4.2.2.
6. The airspace operation performance evaluation and sector assignment assisting intelligent decision-making method according to claim 1, wherein the method comprises the following steps of: the third step specifically comprises the following steps:
firstly, leading airspace data and topography data from an empty pipe basic data platform for airspace simulation and planning, and associating a sector division intelligent auxiliary decision module with an airspace simulation evaluation system;
3.2, leading the flight performance data, the monitoring data, the flight plan data, the control voice data, the airspace information, the airport information and the weather information data to perform airspace microscopic simulation;
3.2.1, acquiring aircraft flow distribution probability by adopting a maximum entropy theory, and processing and analyzing actual flight data to jointly form initial input data of the simulation aircraft flow module by combining aircraft performance and control rule conditions;
3.2.2"AI controllers" issue altitude, speed or heading commands for the aircraft based on real-time air traffic dynamics, while taking command time and content information as "virtual workload" for statistical analysis.
7. The airspace operation performance evaluation and sector assignment assisting intelligent decision-making method according to claim 1, wherein the method comprises the following steps of: the fourth step comprises the following steps:
4.1, analyzing the clustering result of the aircraft track data reflecting the control behavior, combining the control sector handover protocol and the operation mode, and taking the fuzzy comprehensive decision result in each airspace state as the prior knowledge of Q learning by using a fuzzy comprehensive decision method, wherein the number of the sectors which can be processed simultaneously is 10 by using a machine learning-based sector optimization dividing technology;
4.1.1 constructing a decision set V and a factor set U of a fuzzy comprehensive decision method, setting an action set Xi and a state set lambda of a Q learning algorithm, wherein the factor sets are as follows: the area of the airspace subarea, the workload in the airspace subarea, the shape of the airspace subarea and the space connection attribute of the route and the route, and the decision set is an action set selected by the airspace subarea system;
4.1.2 referring to expert experience, a fuzzy evaluation matrix R and a weight set Θ are constructed. Method for calculating all states lambda in state set lambda by utilizing fuzzy comprehensive decision l Vector of superiority of taking various actions
4.1.3 the respective states Λ l Lower part (C)Performing normalization adjustment and taking the normalization adjustment as prior knowledge of Q learning;
4.2, respectively designing a continuous return function and a discrete return function based on the air traffic behavior region distribution characteristics and the air traffic behavior recognition and based on the control handover activities, wherein the composition of the continuous return function comprises the flight section length, the fuel economy and the maneuverability demand factors of the aircraft; to ensure flight safety, a discrete penalty term, i.e., a discrete return function, is applied for the process of causing a dangerous condition or exceeding the sector capacity state transition;
4.3 generating an initial Q matrix according to all the determined states and action sets, if the total number of sectors is 10, giving the current state of the system at the time tReinforced learning rate alpha t And a prize value discount factor beta, if the 10 sector selection actions are +.>They will transition to a new system state ζ' and each receive a prize valueAfter all the handover location points or sub-areas perform the actions, a new sector configuration result is reached, so that the Q value of each sector at the (t+1) moment is updated:
and->Respectively represent sector system status->Sector +.>And->Cost value paid,/->Representing status->10 Nash equalization strategies, respectively adopted by 10 sectors in the case of +.>And->Respectively indicate status->Sector +.>And->Is a function of the state value of (c),and->Respectively indicate (t+1) time status +.>Sector +.>And->State-behavior value functions of (2); and after the Q value is stable, finishing the training process of the model to obtain a sector division result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310675279.2A CN116797043A (en) | 2023-06-08 | 2023-06-08 | Auxiliary intelligent decision method for airspace running performance evaluation and sector division |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310675279.2A CN116797043A (en) | 2023-06-08 | 2023-06-08 | Auxiliary intelligent decision method for airspace running performance evaluation and sector division |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116797043A true CN116797043A (en) | 2023-09-22 |
Family
ID=88040848
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310675279.2A Pending CN116797043A (en) | 2023-06-08 | 2023-06-08 | Auxiliary intelligent decision method for airspace running performance evaluation and sector division |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116797043A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117391526A (en) * | 2023-11-01 | 2024-01-12 | 中国民航科学技术研究院 | Method, system, electronic device and storage medium for measuring and calculating workload of controller |
| CN117939503A (en) * | 2024-03-21 | 2024-04-26 | 中国电子科技集团公司第二十八研究所 | Low-altitude airspace information collection processing method and system based on cloud edge cooperative architecture |
-
2023
- 2023-06-08 CN CN202310675279.2A patent/CN116797043A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117391526A (en) * | 2023-11-01 | 2024-01-12 | 中国民航科学技术研究院 | Method, system, electronic device and storage medium for measuring and calculating workload of controller |
| CN117391526B (en) * | 2023-11-01 | 2024-05-28 | 中国民航科学技术研究院 | Method, system, electronic device and storage medium for measuring and calculating workload of controller |
| CN117939503A (en) * | 2024-03-21 | 2024-04-26 | 中国电子科技集团公司第二十八研究所 | Low-altitude airspace information collection processing method and system based on cloud edge cooperative architecture |
| CN117939503B (en) * | 2024-03-21 | 2024-05-31 | 中国电子科技集团公司第二十八研究所 | Low-altitude airspace information collection processing method and system based on cloud edge cooperative architecture |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116797043A (en) | Auxiliary intelligent decision method for airspace running performance evaluation and sector division | |
| CN109508812B (en) | Aircraft track prediction method based on deep memory network | |
| Silva Arantes et al. | Heuristic and genetic algorithm approaches for UAV path planning under critical situation | |
| CN112162564B (en) | Unmanned aerial vehicle flight control method based on simulation learning and reinforcement learning algorithm | |
| Zhou et al. | Smart train operation algorithms based on expert knowledge and reinforcement learning | |
| CN108564136B (en) | A kind of airspace operation Situation Assessment classification method based on fuzzy reasoning | |
| CN110989626B (en) | Unmanned aerial vehicle path planning method based on control parameterization | |
| CN111680882B (en) | Multi-level risk assessment method for unmanned aerial vehicle system under probability coupling relation | |
| US20200202723A1 (en) | Machine learning in avionics | |
| CN112327903B (en) | Aircraft trajectory generation method based on deep mixed density network | |
| CN111882047B (en) | Rapid empty pipe anti-collision method based on reinforcement learning and linear programming | |
| CN112818599A (en) | Air control method based on reinforcement learning and four-dimensional track | |
| CN116560406A (en) | Unmanned aerial vehicle cluster collaborative planning and autonomous scheduling method | |
| CN113988408A (en) | Multi-objective planning-based artificial guidance method for passenger flow evacuation in subway station | |
| CN113743666B (en) | Flight outgoing mission planning method, device, equipment and medium | |
| CN114969978A (en) | Airplane offshore platform recovery scheduling method based on improved wolf optimization algorithm | |
| Gao et al. | Hybrid swarm intelligent algorithm for multi-UAV formation reconfiguration | |
| CN113093568A (en) | Airplane automatic driving operation simulation method based on long-time and short-time memory network | |
| Xie et al. | Estimating a civil aircraft’s development cost with a GM (1, N) model and an MLP neural network | |
| Xiao et al. | Model-based dynamic shielding for safe and efficient multi-agent reinforcement learning | |
| Li et al. | Traffic fatalities prediction based on support vector machine | |
| CN110428666A (en) | A kind of aircarrier aircraft on-air collision solution decision-making technique for evolving intelligent based on man-machine coordination | |
| CN115493599A (en) | Safety-based aircraft path planning method | |
| Cody et al. | Applying learning systems theory to model cognitive unmanned aerial vehicles | |
| Neidhoefer et al. | Intelligent control for near-autonomous aircraft missions |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication |