An evolutionary neural network, weight updating method and device for intelligent evaluation of controllers

By adopting an evolutionary neural network in the controller intelligent evaluation system, combining backpropagation and NEAT neural network, the problem of poor adaptability in the existing technology is solved, and intelligent and intelligent evaluation of controllers is realized, and the evaluation is improved. Applicability and adaptability.

CN114492743BActive Publication Date: 2024-11-22AIR TRAFFIC ADMINISTRATION OF CIVIL AVIATION ADMINISTRATION OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210053521.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-18
Publication Date
2024-11-22
Estimated Expiration
2042-01-18

AI Technical Summary

Technical Problem

There are differences in adaptability and accuracy of the existing controller intelligent evaluation system, especially due to the differences in characteristics of different regions and units, which leads to the problems of unfamiliarity with details, missing details and long-term tracking and recording difficulties in traditional expert assessment methods. .

Method used

An intelligent evaluation method based on evolutionary neural networks is adopted, including sequentially set input layer, artificial weight layer and NEAT neural network hidden layer, and the initial artificial empowerment default value is obtained through backpropagation neural network training, and in the evolutionary neural network Use the ReLU activation function and the mean square error loss function, combined with expert subjective evaluation and objective evaluation, to achieve intelligent assessment of controllers.

Benefits of technology

This method improves the applicability and adaptability of the algorithm, can perform intelligent evaluation in different scenarios and stages, provide universal weight calculation results, support controller assessments in different regions and units, and combines expert knowledge and computer intelligence, Achieve a comprehensive assessment of controllers.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114492743B_ABST
    Figure CN114492743B_ABST
Patent Text Reader

Abstract

The present application discloses an evolutionary neural network, a weight updating method and a device for intelligent evaluation of air traffic controllers. The structure of the evolutionary neural network includes the following layers arranged in sequence: a first input layer, including a plurality of nodes corresponding to secondary indicators; a second input layer, including 5 node surfaces corresponding to primary indicators; an artificial weight layer, wherein the values ​​in the nodes of the artificial weight layer are obtained by multiplying the data in the nodes in the second input layer by artificial weights; a NEAT neural network hidden layer; a first output layer and a second output layer, including 5 output nodes corresponding to the primary indicators and 1 comprehensive result node; the evolutionary neural network is trained by taking the computer operations recorded in the control intelligent simulator as input and the corresponding scoring results as output; wherein the initial artificial weights of the artificial weight layer during training are obtained by back propagation neural network training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of air traffic control, and in particular to a weight optimization method for an intelligent evaluation system for air traffic controllers based on an evolutionary neural network. Background Art

[0002] The controller intelligent evaluation system is an intelligent assessment system based on the control simulator. It is an important auxiliary equipment for controller assessment in various scenarios and stages. It can not only be used to simulate dangerous / emergency / far beyond the general workload scenarios that cannot or have a very low probability of occurring in actual work scenarios, to achieve the purpose of testing the general ability limit and emergency handling capabilities of controllers; it can also be used as a teaching device to provide a good simulation environment for new control trainees to familiarize themselves with basic control work processes, work scenarios, work methods, etc.

[0003] There are two main assessment methods used in the corresponding controller training assessment. Their limitations in the use of indicator weights are as follows:

[0004] One is to score the controllers based on the subjective assessment of the control experts. This assessment can be written, based on actual work scenarios, or based on the assessment results of ordinary control simulators. The limitations of this method are: when experts assess different assessment indicators, the weights have a certain subjective component when scoring, and the evaluation results of different experts are different, and the weight results they give are often different. In addition, the subjective expression of the importance of indicators is difficult to express with specific numbers.

[0005] The other is to conduct assessment based on intelligent control simulators. The advantage of this type of assessment is that it saves the human resources of control experts, uses intelligent algorithms to assess trainees, and generates evaluation results for trainees to check themselves and teachers to view, such as Chinese patent application 202011199912.8. There are two main ways to calculate the weights of assessment indicators currently used in this intelligent simulator system. One is random generation, which requires users to manually adjust it based on subjective judgment; the other is artificial neural network, but its network structure is simple, and the trained neural network is poor in universal applicability and user participation. The calculation model is complex and cannot incorporate the evaluator's preference for different indicators. In some cases, there is a big difference between the artificial neural network determination and the actual importance of the indicator.

[0006] For example, there are many types of control work, and the controls of different natures such as airport control, terminal area control and regional radar control are quite different from each other. Secondly, even if the control units of the same nature of work provide air traffic services within the area under their jurisdiction, the complexity of air traffic activities will be different due to the differences in airport distribution, route / airway direction and airspace environment in the region. For the above reasons, traditional expert assessment and manual evaluation methods are still widely used. Since the evaluation and assessment of the controller's work is necessary and frequent, and there are different depths of assessment methods according to different assessment purposes. However, when implementing this assessment method, most of the assessment instructors are selected from each other among the regional control units, which leads to the following problems: 1) Inspectors who are not from the unit may not be familiar with the specific details of the work of the person to be assessed, and cannot conduct a detailed and comprehensive assessment; 2) Inspectors may rely too much on fixed assessment forms to evaluate the assessment objects one by one, and details not mentioned on the assessment forms may be missed during the evaluation process; 3) Due to the mobility of instructors and the frequency of evaluations, it is often impossible to keep long-term tracking records for the assessment of a single controller, thus losing the function of mutual verification between multiple assessment results. Summary of the invention

[0007] The embodiments of the present application provide an evolutionary neural network, an intelligent evaluation method and a device for intelligent evaluation of air traffic controllers, so as to at least solve the technical problem of poor adaptability of existing air traffic controller evaluation systems.

[0008] According to one aspect of the present application, an evolutionary neural network for intelligent evaluation of controllers is provided, wherein the structure of the evolutionary neural network comprises the following layers arranged in sequence:

[0009] The first input layer includes a plurality of nodes corresponding to the secondary indicators one by one, wherein the secondary indicators refer to indicators used to evaluate the control operation, and any computer operation recorded in the control intelligent simulator corresponds to at least one secondary indicator among all the secondary indicators; the secondary indicators are indicators selected from the assessment indicators of multiple regional air traffic control bureaus using the Delphi method and are computer coded and quantifiable; the data in the nodes of the first input layer are obtained by normalizing the data in the computer operation recorded in the corresponding control intelligent simulator;

[0010] The second input layer includes five nodes corresponding to the first-level indicators, namely rules, communication, safety, efficiency, and human-computer interface;

[0011] The artificial weight layer includes 5 nodes corresponding to the first-level indicators one by one, and the values ​​in the nodes of the artificial weight layer are obtained by multiplying the data in the nodes in the second input layer by the artificial weights;

[0012] NEAT neural network hidden layer, used for adaptive learning and optimization of the topological structure of the neural network;

[0013] The first output layer includes 5 nodes corresponding to the first-level indicators. The value of each node is the score result of the corresponding first-level indicator.

[0014] The second output layer includes 1 node, which is a comprehensive scoring result obtained by weighting the nodes of the first output layer;

[0015] The evolutionary neural network uses the computer operations recorded in the control intelligent simulator as input and the corresponding scoring results as output to train the evolutionary neural network; wherein the weights of the initial artificial weighted default values ​​used in the artificial weight layer during training are obtained through back-propagation neural network training.

[0016] Furthermore, in the present invention, the weights of the initial artificially weighted default values ​​used in the training are obtained through back propagation neural network training and include:

[0017] A back propagation neural network is trained using computer operations recorded in a control intelligent simulator as input and corresponding scoring results as output, wherein the input layer structure of the back propagation neural network is the same as the first input layer of the evolutionary neural network, and the output layer structure of the back propagation neural network is the same as the output layer structure of the evolutionary neural network.

[0018] Furthermore, in the present invention, the weights of the nodes in the second input layer obtained by back propagation neural network training that correspond one-to-one to the primary indicators are: rules 0.20405669, calls 0.135801045, security 0.27565417, efficiency 0.18206525, and human-computer interface 0.20021345.

[0019] Furthermore, in the present invention, the activation function used by the back propagation neural network is the ReLU function, the mean square error MSE is used as the loss function, and the optimization method of regularization and gradient descent is adopted.

[0020] Furthermore, in the present invention, when the evolutionary neural network is used to perform intelligent evaluation of controllers after training, the weight range of each node during manual weighting is 0 to 0.4.

[0021] Furthermore, in the present invention, when the evolutionary neural network is trained, the activation function used is the ReLU function, the loss function used is the mean square error MES, and the optimization method is the gradient descent including regularization.

[0022] Furthermore, in the present invention, when the evolutionary neural network is trained, the hidden layer of the NEAT neural network uses a crossover mutation method, the individual selection method is a roulette method, and the evaluation method is a fitness function.

[0023] Furthermore, in the present invention, the encoding initialization network method of the NEAT neural network uses a genetic encoding scheme.

[0024] The second aspect of the present application is to provide a weight updating method for intelligent evaluation of controllers based on an evolutionary neural network, using the neural network described in the first aspect to perform intelligent evaluation of controllers, wherein the artificial weighting of the artificial weight layer is adjusted according to the different needs of the training stage, region, and unit, thereby providing the user with the function of free weighting.

[0025] The third aspect of the present application is to provide an intelligent evaluation device for controllers based on an evolutionary neural network, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the second aspect when executing the program.

[0026] Beneficial effects:

[0027] The present invention provides a new neural network structure for intelligent evaluation of air traffic controllers. The algorithm can combine the advantages of expert subjective evaluation and objective evaluation and make a trade-off between the two, thereby increasing the algorithm's optimization and self-adaptation capabilities. The algorithm can be intelligently applied to different scenarios, trainees at different stages, different regions and different control units to carry out evaluation work for air traffic controller assessment. The weight calculation method and calculation results involved in the algorithm are universal when applied to basic training and assessment of air traffic controllers. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0029] Figure 1 is a flow chart of an intelligent evaluation method according to an embodiment of the present application;

[0030] Figure 2 is a structural diagram of a back propagation neural network according to an embodiment of the present application;

[0031] Figure 3 is a structural diagram of an evolutionary neural network according to an embodiment of the present application;

[0032] Figure 4 is a flow chart of the NEAT algorithm according to an embodiment of the present application;

[0033] Figure 5 is a schematic diagram of the operation of the NEAT algorithm according to an embodiment of the present application;

[0034] Figure 6 is a five-dimensional output energy radar diagram according to an embodiment of the present application;

[0035] Figure 7 is a near-near environment simulation diagram according to an embodiment of the present application;

[0036] Figure 8 is a back propagation neural network loss function according to an embodiment of the present application;

[0037] Fig. 9 is the evolutionary neural network loss function according to an embodiment of the present application;

[0038] Fig.10 It is the prediction accuracy of the evolutionary neural network according to the embodiment of the present application. DETAILED DESCRIPTION

[0039] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0040] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0041] The specific working process in this patent embodiment is as follows Figure 1 As shown: By using the control intelligence simulator and quantitative indicators representing the characteristics of control work, the weights of the indicators are calculated according to the following three steps:

[0042] The first step is to use a back-propagation neural network, such as Figure 2 As shown, the network is trained to learn the training samples and output the weights of the first-level indicators;

[0043] In the second step, we use an evolutionary neural network, such as Figure 3 As shown, the evolutionary neural network is trained to learn the training samples, generate the optimal neural network structure, and output the weights of the secondary indicators;

[0044] In the third step, the test set is used to verify the optimal evolved neural network generated in the second step.

[0045] The above optimal evolutionary neural network is as follows Figure 3 As shown, the structure of the evolved neural network includes the following layers arranged in sequence:

[0046] The first input layer includes a plurality of nodes corresponding to the secondary indicators one by one, wherein the secondary indicators refer to indicators used to evaluate the control operation, and any computer operation recorded in the control intelligent simulator corresponds to at least one secondary indicator among all the secondary indicators; the secondary indicators are indicators selected from the assessment indicators of multiple regional air traffic control bureaus using the Delphi method and are computer coded and quantifiable; the data in the nodes of the first input layer are obtained by normalizing the data in the computer operation recorded in the corresponding control intelligent simulator;

[0047] The second input layer includes five nodes corresponding to the first-level indicators, namely rules, communication, safety, efficiency, and human-computer interface;

[0048] The artificial weight layer includes 5 nodes corresponding to the first-level indicators one by one, and the values ​​in the nodes of the artificial weight layer are obtained by multiplying the data in the nodes in the second input layer by the artificial weights;

[0049] The third input layer includes multiple nodes corresponding to the secondary indicators one by one;

[0050] NEAT neural network hidden layer, used for adaptive learning and optimization of the topological structure of the neural network;

[0051] The first output layer includes 5 nodes corresponding to the first-level indicators. The value of each node is the score result of the corresponding first-level indicator.

[0052] The second output layer includes 1 node, which is a comprehensive scoring result obtained by weighting the nodes in the first output layer;

[0053] The evolutionary neural network uses the computer operations recorded in the control intelligent simulator as input and the corresponding scoring results as output to train the evolutionary neural network; wherein, the weights of the initial artificial weighted default values ​​used in the training of the artificial weight layer are obtained through back propagation neural network training, which are 0.20405669 for rules, 0.135801045 for calls, 0.27565417 for safety, 0.18206525 for efficiency, and 0.20021345 for human-machine interface. When the evolutionary neural network is used for intelligent evaluation of controllers after training, the weight range of each node during artificial weighting is 0 to 0.4, and the weight can be adjusted within the above range according to actual conditions.

[0054] The above-mentioned back propagation neural network and evolutionary neural network are used in the intelligent evaluation system for controllers, especially the weights and weight update methods used in the embodiments of this patent, which can support the completion of the assessment of regional, terminal, and airport controllers in different regions. The system uses a control intelligent simulator as the implementation object and a control intelligent simulator indicator set as the assessment indicator. For the first-level indicators, the system provides a manual subjective weight adjustment function, and for the second-level indicators, an objective weight calculation method is provided. This patented system combines the advantages of expert knowledge and computer intelligence, and the provided assessment indicator weights can achieve the purpose of intelligent assessment of controllers.

[0055] The structure of the above back propagation neural network can be seen in Figure 2 A back propagation (BP) neural network is established for training the primary indicator weights. The input layer of the BP neural network is N data input nodes, the hidden layer is 5 primary indicator nodes, and the output layer has two layers, the first layer is 5 output nodes, and the second layer is 1 output node.

[0056] Each input node corresponds to a computer operation that can be recorded in the intelligent control simulator, and any operation in the control simulator corresponds to at least one of all output nodes. Five hidden nodes correspond to each primary indicator, and the primary indicators include five indicators, namely rules, calls, safety, efficiency, and human-computer interface. The five output nodes of the first layer of the output layer use neural network calculations to obtain five output values, corresponding to five different dimensions of capability quantification assessment purposes. The one output node of the second layer of the output layer uses neural network calculations to obtain one output value, corresponding to one-dimensional comprehensive capability quantification assessment purpose.

[0057] The experimental data used by the BP neural network are: m groups of experimental data are obtained, each group of experimental data is the controller's operation on the simulator and the 5-dimensional score and 1-dimensional comprehensive score;

[0058] The training method of BP neural network using training set is as follows: m groups of experimental data are divided into training set, test set and development set, and multiple neural networks with the same structure are trained using training set, the number of input layer nodes of each neural network is the number of computer operations that can be recorded in the above-mentioned control intelligent simulator, the output layer 1 node of the neural network has 5 nodes, and the output layer 2 node has 1. A hidden layer is set between the input layer and the output layer, and the hidden layer is set to 5 neural nodes; the test set is used to evaluate the training results of the neural network, and the development set is used to evaluate the prediction effect of the neural network, and finally the optimal BP neural network structure is determined, and the weight results of all neural networks are recorded.

[0059] According to the intelligent evaluation index system structure of controller training effect, the actual values ​​of 6 outputs including 5 different dimension ability assessment results and 1 comprehensive ability assessment result are obtained from the m groups of simulator training results collected, which are recorded as Divide these m samples into training set m according to the ratio of 4:4:2 train , test set m test , and development set m dev . train training samples are used to train N neural networks with different structures. test training samples are used to screen the advantages among these N neural networks, and finally m dev The prediction effect of the neural network model left by the screening is evaluated by a training sample. After the accuracy meets the requirements, the parameters in the neural network are the weights of the first-level indicators.

[0060] Using neural network, the method of training the weights of various indicators is as follows: design an alternative neural network structure, including 132 input layer nodes [x1, x2, …, x 132 ] T , 5 in output layer 1 + 1 in output layer 2, a total of 6 output nodes [y1, y2, ..., y6] T There are n hidden layers (n=1, 2, ...., 10), and each layer has 5 neural nodes. i represents the i-th hidden layer, represents the j-th neuron, the activation function used is the ReLU function, the mean square error MSE is used as the loss function, and the regularization and gradient descent optimization methods are used. The number of hidden layers n = 1, 2, ...., 10. Different n determines the different structures of the neural network.

[0061] Finally, it is determined that the input layer has 132 nodes, the two output layers have 6 nodes, the hidden layer n=1 has 1 layer, and there are 5 neural nodes. The activation function used in this method is the ReLU function, and the loss function loss is:

[0062]

[0063] in represents m input nodes, j = 1, 2, ..., 6, where j = 1, 2, ..., 5 represents the prediction results of 5-dimensional ratings, j = 6 represents the prediction results of comprehensive ability assessment, λ is the regularization hyperparameter, w [l] is the weight w of the lth layer, ||·|| F is the F-norm.

[0064] Combine the following Figure 3The structure of the evolutionary neural network is further described in detail. An evolutionary neural network is established on the basis of the BP neural network structure, and an artificial weight layer and a NEAT hidden layer are added. The specific evolutionary neural network is divided into a data input layer, input layer 1, input layer 2, an artificial weight layer and a NEAT neural network hidden layer and output layer 1 and output layer 2. The data input layer is N data input nodes, input layer 1 is 5 first-level indicator nodes, the artificial weight layer is 5 weighted first-level indicator nodes, the hidden layer is a NEAT neural network random structure, and the output layer has two layers, the first layer is 5 output nodes, and the second layer is 1 output node.

[0065] Among the N data input nodes, each data input node corresponds to a computer operation that can be recorded in the control intelligent simulator. After obtaining the results of the recorded calculation operations, the data input node will normalize the data and feed it into the neural network. Each node of the input layer 1 corresponds to each secondary indicator one by one. Any operation in the control intelligent simulator corresponds to at least one secondary indicator among all secondary indicators. The secondary indicators are selected from the assessment indicators of multiple regional air traffic control bureaus using the Delphi method and meet the requirements of computer coding and quantification. The experimental data of the controller on the simulator is obtained, and the number of each secondary indicator corresponding to the experimental data is obtained and normalized as the input data of the neural input layer 1.

[0066] The 5 input layer 2 nodes correspond to each first-level indicator one by one. The first-level indicators include 5, namely rules, calls, safety, efficiency, and human-computer interface. The input nodes of the 5 artificial weight layers are the 5 nodes in the input layer 2 and the artificial weights are multiplied and then input into the artificial weight layer. The artificial weight layer is normalized again and the data is input into the input layer 3. The input layer 3 is M second-level indicators and serves as the input layer of the hidden layer NEAT neural network. The output layer 1 is 5 output nodes and serves as the output layer of the hidden layer NEAT neural network. The output layer 2 is 1 output node.

[0067] The artificial weight layer is the weighted and normalized input value of the five first-level indicators, and combined with the data of each second-level indicator obtained by input layer 1, they will serve as the input layer of the NEAT neural network. Output layer 1 is 5 output nodes, which will serve as the output layer of the hidden layer NEAT neural network and are used for structural evolution of the NEAT neural network. The five output nodes of the first layer of the output layer use neural network calculations to obtain five output values, corresponding to the purpose of quantitative assessment of capabilities in five different dimensions. The one output node of the second layer of the output layer uses neural network calculations to obtain one output value, corresponding to the purpose of quantitative assessment of comprehensive capabilities in one dimension.

[0068] The experimental data set used by the evolutionary neural network is the same as that of the BP neural network.

[0069] The training method of the evolutionary neural network using the training set is as follows: m groups of experimental data are divided into a training set and a test set, and an evolvable neural network is trained using the training set. The number of input layer nodes of each neural network is the number of computer operations that can be recorded in the above-mentioned control intelligent simulator, and the output layer 1 of the neural network has 5 nodes and the output layer 2 has 1 node. A NEAT neural network is set between the input layer and the output layer as a hidden layer. The input layer and output layer of the NEAT neural network structure are determined as the artificial weight layer and output layer 1, and the number of hidden nodes and the number of hidden layers will be obtained by the NEAT algorithm. The test set is used to evaluate the results of the neural network training, and the optimal evolutionary neural network structure is finally determined, and the weight results of all neural networks are recorded.

[0070] When training the evolutionary neural network, the activation function used is the ReLU function, the loss function used is the mean square error MES, the optimization method is the gradient descent including regularization, the NEAT evolutionary neural network uses the crossover mutation method, the individual selection method is the roulette method, and the evaluation method is the fitness function.

[0071] After the BP neural network and evolutionary neural network training is completed, the weights in the network are saved.

[0072] exist Figure 4 In the NEAT neural network, the operation process is as follows:

[0073] The NEAT neural network partial encoding initialization network method uses a genetic encoding scheme.

[0074] The crossover operation is based on the evolutionary training crossover operation, and its mathematical expression is:

[0075] x′ i =x i +σ i N(0,1), where x is the gene DNA, i represents the i-th gene point, N(0,1) is subject to normal distribution, N i (0,1) is the random number generated by the i-th base point that conforms to the normal distribution, and γ is the global system

[0076] The mutation operation is weight mutation and adding connection mutation. The fitness function is F, and its mathematical expression is:

[0077] F=R 2 =1-SSE / SST, where for y i The estimated value of for y i The average value of .

[0078] The selection operation is (μ+λ)-ES, that is, the parent selects μ individuals and the offspring selects μ individuals with higher fitness to form a new population of the next generation.

[0079] The algorithm operation process is: set the population size P(m) = M, the evolutionary generation R, the number of iterations I, and the number of network layers L.

[0080] Initialization: Randomly generate an initial population P(m) and set R=1;

[0081] The fitness of each individual in P(m) is evaluated through the fitness function;

[0082] Perform crossover and mutation operations to obtain P”(m);

[0083] Perform individual selection and replication operations to obtain P(m+1)

[0084] Termination condition judgment: When the termination condition is not met, set m=m+1 and go to (c) to continue evolution; if the termination condition is not met, output the current optimal individual and the algorithm ends.

[0085] exist Figure 5 In NEAT evolutionary operation, crossover and mutation operation process is as follows:

[0086] NEAT's mutation operation can change the weight of the connection and the network structure. Structural mutations include adding connections and adding nodes. Each mutation operation expands the genome by adding genes. When adding connections as a mutation operation, it is achieved by adding new connection genes to connect nodes that are not connected; when adding new nodes as a mutation operation, the existing connection is split, the new node is placed in the old connection position, and the old connection gene is disabled. When two new connection genes are placed in the genome, the new node is located between the two new connections, which means that the new node gene is also added to the genome. During the mutation process, when a new gene appears, an innovation number is added and assigned to the gene.

[0087] In such Figure 5 The two parent genes, parent 1 and parent 2, are affected by the appearance of new gene node 6 when the algorithm performs mutation operations on the parent 2 gene. The original connection between gene node 1 and gene node 4 no longer exists, and is replaced by new connections 1->7 and 6->4. The weight of the original connection 1->4 will be inherited by the new connection 1->7, and the weight of the connection 6->4 will be assigned the default value 1 by the algorithm.

[0088] Figure 5The article also briefly demonstrates the process of crossover operation in the hidden layer of the evolutionary neural network NEAT described in this article, and the number of neurons is simplified. As shown in the figure, there are two parent genes, parent 1 and parent 2, where parent 1 contains 5 neurons, gene nodes 1 to 3 come from the input layer 3 of the evolutionary neural network, gene node 4 represents the number of layers and neurons in the NEAT hidden layer, and neuron 5 comes from the output layer 1 of the evolutionary neural network; the new gene node 6 contained in parent 2 is also a NAET hidden layer, that is, the NEAT network structure represented by parent 2 contains two hidden layers and 2 hidden layer gene nodes. The offspring genes generated by the crossover operation merge and inherit the network structure represented by the two parent genes.

[0089] exist Figure 6 This is an example of the 5-dimensional output result of the evolutionary neural network in this embodiment. The 5-dimensional output result represents the results of the assessment in the five aspects of rules, calls, safety, efficiency, and human-machine interface. Each result is a number between 0 and 1. The results are divided into 1 to 5 levels according to percentages, thereby forming a 5-dimensional capability radar chart model.

[0090] When using evolutionary neural networks to conduct intelligent assessment of air traffic controllers, the 5-dimensional and 1-dimensional output results generated can be manually reviewed and modified by experts to determine whether to adopt the prediction results. If adopted, the prediction results are output; otherwise, the 5-dimensional and 1-dimensional output results are manually modified, and the corresponding experimental data and the re-scored results are used as a set of training data to update the evolutionary neural network.

[0091] When using evolutionary neural networks in different scenarios, experts can manually adjust the weights of the five level 1 indicators of the "artificial weight" layer according to different experimental objectives. At the same time, this patent provides initial setting values, multiple sets of reference values, and a recommended range for weight modification.

[0092] The following is a specific implementation description using the Guangzhou approach control simulation as an example. The main indicators used are approach control related indicators, including 5 first-level indicators, A rules, B communication, C safety, D efficiency, E human-machine interface, and 132 second-level indicators.

[0093] Analyze the airspace environment and sector characteristics; such as Figure 6 As shown in the figure, Guangzhou approach is divided into five areas according to the geographical level. They are the Approach North Sector (AN), Approach West Sector (AW), Approach East Sector (AE), Approach Sector (AA) and Departure Sector (AD). The functions of each sector are as follows:

[0094] Departure sector: mainly responsible for the flight take-off altitude to 2100 meters;

[0095] The approach sector is mainly responsible for the final separation of incoming flights, from 2400 meters to the establishment of blind landing;

[0096] Approach North Sector: Responsible for arranging the order of flights arriving at Guangzhou from ATAGA, IGONO, and GYA, assigning runway numbers and approach procedures. Also responsible for flights departing from YIN and P268, and handing them over to regional control at the specified altitude;

[0097] Approach West Sector: Responsible for controlling flights departing from NOLON and GYA and landing in Shenzhen, Zhuhai and Macao, as well as departing flights departing from Guangzhou and flying to VIBOS and P269;

[0098] Approach East Sector: Responsible for controlling flights departing from Guangzhou and departing from LMN, flights departing from Shenzhen, Zhuhai and Macau and flying over IDUMA, as well as departing flights from IDUMA and P270;

[0099] Collect experimental data and conduct 100 sets of simulation training for the experimenters. In each experiment, every instruction issued by the experimenter, every operation of the automation system and the aircraft data are triggered and counted according to the secondary indicators of the present invention. If the behavior meets the secondary indicators, it is recorded and the parameters are input into the neural network. Before that, the neural network is trained with 100 or more sets of data collected, and the hyperparameter learning rate α is set to 0.0075. The back propagation neural network is trained in sequence, and the training round is 5000 rounds. The loss value changes during training are as follows: Figure 8 .

[0100] The neural network is trained with 5000 rounds and 50 iterations. The loss changes during training are as follows: Fig. 9 , the one-dimensional output accuracy is as follows Fig.10 .

[0101] Compared with pure neural network algorithms, the loss function value of the neural network structure and weights described in this patent is smaller under the same data set (that is, closer to the data label given by experts). And the algorithm described in this patent can use the same structure in different scenarios to achieve better prediction results. And the five-dimensional output that can be output by this article can better provide teachers with an evaluation of the students' various abilities, and the one-dimensional output provides teachers with a comprehensive evaluation result.

[0102] The innovative method of this paper is mainly based on the advantages of the fusion of "expert evaluation method", "artificial neural network method" and "genetic algorithm evaluation". A weight optimization method for the controller intelligent evaluation system using an improved evolutionary neural network using a combined weighting method is formed.

[0103] The NEAT neural network was used and improved, using a hierarchical neural network that combines the advantages of expert subjective scoring and objective evaluation;

[0104] The weights formed have the advantages of self-adaptation, self-learning, self-adjustment, self-optimization, self-organization and self-planning; they can fully integrate the basic training of air traffic controllers and the characteristics of air traffic control training in various regions and units. These advantages are not possessed by the weight optimization method in the current air traffic controller intelligent assessment system.

[0105] The system and method provided in this embodiment can meet the needs of individual controllers for personalized autonomous learning and assessment, as well as the needs of control units to integrate and comprehensively track and evaluate the controllers' learning progress, skill level, and business capabilities, thereby becoming a system for continuous capability evaluation of controllers.

[0106] The weight optimization method of the intelligent evaluation system for air traffic controllers created by the present invention uses a neural network to fit the nonlinear relationship between control indicators. It can intelligently (adaptively and self-adjust) update the weights according to the assessment conditions (referring to various training assessments for air traffic controllers at different training stages, as well as the different personalized assessment scenarios of various units in different regions of the country) and output assessment results that meet the requirements of the "Civil Aviation Air Traffic Control Training Management Rules" (CCAR-70TM-R1) manual.

[0107] According to another aspect of the present application, a weight updating method for intelligent evaluation of controllers based on an evolutionary neural network is provided. In the process of intelligent evaluation of controllers using the above-mentioned neural network, artificial weighting of the artificial weight layer is performed according to the purpose of use by the user, specifically including adjustments according to the different needs of training stages, regions, and units.

[0108] According to another aspect of the present application, a controller intelligent evaluation device based on evolutionary neural network includes a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the method described in the second aspect when executing the program.

[0109] It should be noted that the above-mentioned computer-implemented intelligent evaluation method for controllers based on an evolutionary neural network is the same as the previously described intelligent evaluation method for controllers based on an evolutionary neural network, and will not be described in detail here.

[0110] These computer programs can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1The steps of the functions specified in one or more blocks can be implemented by different modules corresponding to different steps.

[0111] The above program can be run in the processor, or it can be stored in the memory (or computer-readable medium), which includes permanent and non-permanent, removable and non-removable media. Information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined in this article, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0112] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. An evolutionary neural network for intelligent evaluation of air traffic controllers, Features: The structure of the evolved neural network includes the following layers arranged in sequence: The first input layer includes a plurality of nodes corresponding to the secondary indicators one by one, wherein the secondary indicators refer to indicators used to evaluate the control operation, and any computer operation recorded in the control intelligent simulator corresponds to at least one secondary indicator among all the secondary indicators; the secondary indicators are indicators selected from the assessment indicators of multiple regional air traffic control bureaus using the Delphi method and are computer coded and quantifiable; the data in the nodes of the first input layer are obtained by normalizing the data in the computer operation recorded in the corresponding control intelligent simulator; The second input layer includes five nodes corresponding to the first-level indicators, namely rules, calls, safety, efficiency, and human-computer interface; The artificial weight layer includes 5 nodes corresponding to the first-level indicators one by one, and the values ​​in the nodes of the artificial weight layer are obtained by multiplying the data in the nodes in the second input layer by the artificial weights; NEAT neural network hidden layer, used for adaptive learning and optimization of the topological structure of the neural network; The first output layer includes 5 nodes corresponding to the first-level indicators. The value of each node is the score result of the corresponding first-level indicator. The second output layer includes 1 node, which is a comprehensive scoring result obtained by weighting the nodes in the first output layer; The evolutionary neural network uses the computer operations recorded in the control intelligent simulator as input and the corresponding scoring results as output to train the evolutionary neural network; wherein the weights of the initial artificial weighted default values ​​used in the artificial weight layer during training are obtained through back-propagation neural network training.

2. The evolved neural network according to claim 1, Features: The weights of the initial artificial weighted default values ​​used in the training are obtained through back propagation neural network training and include: A back propagation neural network is trained using computer operations recorded in a control intelligent simulator as input and corresponding scoring results as output, wherein the input layer structure of the back propagation neural network is the same as the first input layer of the evolutionary neural network, and the output layer structure of the back propagation neural network is the same as the output layer structure of the evolutionary neural network.

3. The evolved neural network according to claim 2, Features: The weights of the nodes in the second input layer obtained by the back propagation neural network training that correspond one to one to the primary indicators are: rules 0.20405669, calls 0.135801045, safety 0.27565417, efficiency 0.18206525, and human-computer interface 0.20021345.

4. The evolved neural network according to claim 1, Features: The activation function used by the back propagation neural network is the ReLU function, the mean square error MSE is used as the loss function, and the optimization method of regularization and gradient descent is adopted.

5. The evolved neural network according to claim 1, Features: When the evolutionary neural network is used to perform intelligent evaluation of controllers after training, the weight range of each node during artificial weighting is 0 to 0.

4.

6. The evolved neural network according to claim 1, Features: When the evolutionary neural network is trained, the activation function used is the ReLU function, the loss function used is the mean square error MES, and the optimization method is the gradient descent including regularization.

7. The evolved neural network according to claim 1, Features: When the evolutionary neural network is trained, the hidden layer of the NEAT neural network uses a crossover mutation method, the individual selection method is a roulette method, and the evaluation method is a fitness function.

8. The evolved neural network according to claim 1, Features: The encoding initialization network method of the NEAT neural network uses a genetic encoding scheme.

9. A weight updating method for intelligent evaluation of controllers based on evolutionary neural networks. Features: In the process of using the neural network described in any one of claims 1 to 8 to perform intelligent evaluation of controllers, the artificial weighting of the artificial weight layer is adjusted according to the different needs of the training stage, region, and unit.

10. An intelligent evaluation device for controllers based on evolutionary neural networks, Features: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the method according to claim 9 are implemented when the processor executes the program.