CN110807525B - Neural Network Flight Guarantee Service Time Estimation Method Based on Improved Genetic Algorithm - Google Patents

Abstract

A neural network flight guarantee service time estimation method based on an improved genetic algorithm is disclosed. The method comprises the steps of dividing flight guarantee services into three types; selecting a standard action node, reducing the dimension of an original variable, and constructing a BP neural network flight guarantee service time estimation model; preprocessing and normalizing the data, and training and testing the model; optimizing the model using an improved genetic algorithm; and actually applying the model and the like. The neural network flight guarantee service time estimation method based on the improved genetic algorithm has the following beneficial effects: on the basis of a traditional genetic algorithm, a chromosome structure, a fitness function, a selection operator, a crossover operator, a mutation operator and crossover mutation probability are respectively designed, so that accurate estimation of flight guarantee service time is realized, and further flight guarantee service efficiency can be improved.

Description

Neural network flight support service time estimation method based on improved genetic algorithm

Technical Field

The invention belongs to the technical field of civil aviation, and in particular relates to a neural network flight support service time estimation method based on an improved genetic algorithm (AMGA).

Background Art

Managers of airport support services must not only consider the constraints of flight support service operation time, but also the exclusivity of aircraft to equipment occupancy. When dispatching special vehicle equipment, factors such as aircraft model, flight schedule, and vehicle route need to be considered, so the flight support process is nonlinear, non-stationary, and dynamic. Once a certain link in the overall flight support service process is disturbed and the operation is not completed within the time specified by the standard, it will directly affect the subsequent flight support service links, forming a ripple effect, and may even cause serious consequences. Therefore, it is necessary to establish a mathematical model and computer simulation technology for flight support services, and improve flight operation efficiency by detailed characterization of ground operation status, so as to ensure maximum resource utilization and equipment use.

When BP neural network solves complex nonlinear problems, one part is the forward propagation of samples, and the other part is the back propagation of errors. The connection weights and thresholds of each layer are adjusted to reduce the actual error. This cycle is repeated until the output error of the model is less than the accuracy requirement, that is, the model output result is infinitely close to the actual result. However, there are few methods for estimating flight support service time using BP neural network.

Summary of the invention

In order to solve the above problems, the purpose of the present invention is to provide a neural network flight support service time estimation method based on an improved genetic algorithm.

In order to achieve the above object, the neural network flight support service time estimation method based on improved genetic algorithm provided by the present invention comprises the following steps performed in sequence:

Step 1: Analyze the flight support service process of airport surface operation, and divide the flight support services into three categories according to the different objects being controlled, namely aircraft services, passenger services and baggage, cargo and mail services;

Step 2: Select n standard action nodes from the above three types of flight support services, take the time variable of each standard action node as an original variable, reduce the dimension of the original variable to obtain m principal components and use the m principal components as the input variables of the BP neural network flight support service time estimation model, and then build the BP neural network flight support service time estimation model according to the number of principal components;

Step 3: Preprocess and normalize the data related to the above input variables, and then randomly divide all the processed data into training samples and test samples, and then input the training samples into the BP neural network flight support service time estimation model to train the model, and finally input the test samples into the trained BP neural network flight support service time estimation model;

Step 4: Use the improved genetic algorithm to optimize the trained BP neural network flight support service time estimation model to obtain an optimized improved genetic algorithm neural network flight support service time estimation model;

Step 5: Process the flight support service data of any airport according to the above steps 1 and 2 and input it into the optimized neural network flight support service time estimation model of the improved genetic algorithm. The output of the optimized neural network flight support service time estimation model of the improved genetic algorithm is the estimated value of the flight support service time of the airport.

In step 1), the flight support service process of the airport field operation is analyzed, and the flight support service is divided into three categories according to different controlled objects, namely, aircraft service, passenger service and baggage cargo and mail service. The method is:

The aircraft services mentioned above mainly include: adding wheel chocks, jet bridge docking, passenger ladder docking, cabin cleaning, garbage disposal, sewage cleaning, adding in-flight meals, refueling, jet bridge evacuation, passenger ladder evacuation, maintenance inspection, and removing wheel chocks; the passenger services mainly include opening the cabin door, passengers disembarking, passengers boarding, and closing the cabin door; the baggage, cargo and mail services mainly include opening the cargo door, unloading cargo and mail baggage, loading cargo and mail baggage, and closing the cargo door.

In step 2), n standard action nodes are selected from the above three types of flight support services, the time variable of each standard action node is used as an original variable, the original variable is reduced in dimension to obtain m principal components and the m principal components are used as input variables of the BP neural network flight support service time estimation model, and then the method of constructing the BP neural network flight support service time estimation model according to the number of principal components is: first, according to step 1), the standard action nodes are selected from the required flight support services as the original variables, and then the principal component analysis method is used to reduce the dimension of the above original variables, and the original variables after dimension reduction are used as the input layer nodes of the BP neural network, and the number of nodes of the corresponding hidden layer output layer is calculated, and finally the BP neural network flight support service time estimation model is constructed.

In step 3), the data related to the above-mentioned input variables are preprocessed and normalized, and then all the processed data are randomly divided into training samples and test samples. The training samples are then input into the BP neural network flight support service time estimation model to train the model. Finally, the test samples are input into the trained BP neural network flight support service time estimation model. The method is: first, data related to the input variables of the BP neural network flight support service time estimation model are selected from the real data of the actual flight support service process of the airport, and the data is processed using a normalization method to reduce the impact of missing data. Secondly, the normalized data is divided into training samples and test samples, and then successively input into the BP neural network flight support service time estimation model to train and test the model, and finally a trained BP neural network flight support service time estimation model is obtained.

In step 4), the method of using the improved genetic algorithm to optimize the above-trained BP neural network flight support service time estimation model to obtain the optimized improved genetic algorithm neural network flight support service time estimation model is: first, improve the two-layer structure chromosome of the genetic algorithm, improve the fitness function of the genetic algorithm, and then select the operation methods of "best individual preservation strategy" and "random league selection strategy" to improve the selection operator of the genetic algorithm to avoid the probability of selection being too large, resulting in the reduction of genetic algorithm efficiency and the destruction of excellent individuals in the population, and at the same time avoid the probability of selection being too small and falling into "premature maturity" and designing adaptive crossover probability and mutation probability, and finally completing the improved genetic algorithm neural network flight support service time estimation model.

The neural network flight support service time estimation method based on improved genetic algorithm provided by the present invention has the following beneficial effects: on the basis of the traditional genetic algorithm, the chromosome structure, fitness function, selection operator, crossover operator, mutation operator and crossover mutation probability are designed respectively to achieve accurate estimation of flight support service time, thereby improving the efficiency of flight support service.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a schematic diagram of a specific flight support service.

FIG. 2 is a schematic diagram of the basic structure of the BP neural network provided by the present invention.

FIG. 3 is a flow chart of a flight support service operation according to an example of the present invention.

FIG. 4 is a two-layer ladder structure chromosome and its coding diagram designed by the present invention.

FIG5 is a flow chart of the improved genetic algorithm (AMGA) designed by the present invention.

FIG6 is a graph showing the variation of the number of hidden layer nodes designed by the present invention.

Figure 7 is a comparison chart of the estimation errors of the three models using the improved genetic algorithm (AMGA), the traditional genetic algorithm (GA), and no algorithm.

Figure 8 is a comparison chart of the estimated and actual values of flight support service time.

DETAILED DESCRIPTION

The neural network flight support service time estimation method based on the improved genetic algorithm provided by the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

The neural network flight support service time estimation method based on the improved genetic algorithm provided by the present invention comprises the following steps performed in sequence:

Step 1: Analyze the flight support service process of airport surface operation, and divide the flight support services into three categories according to the different objects being controlled, namely aircraft services, passenger services and baggage, cargo and mail services;

Among them, aircraft services mainly include: wheel chock installation, jet bridge docking, passenger staircase docking, cabin cleaning, garbage disposal, sewage cleaning, in-flight food addition, aviation fuel filling, jet bridge evacuation, passenger staircase evacuation, maintenance inspection, and wheel chock removal; passenger services mainly include cabin door opening, passenger disembarkation, passenger boarding, and cabin door closing; baggage and cargo services mainly include cargo door opening, cargo and mail baggage unloading, cargo and mail baggage loading, and cargo door closing. It can be seen that flight support services are very complex. The specific flight support service diagram is shown in Figure 1. When the aircraft is parked on the apron, various flight support service operations sometimes need to be performed simultaneously.

Flight support services include several operations, each of which requires the use of several resources and equipment. The following is an analysis of the equipment involved in the main flight support services. Passenger disembarkation, passenger boarding, and crew boarding operations mainly use jet bridges, shuttle buses, and passenger elevators. Jet bridges are used when the aircraft is close to the aircraft position, and passenger elevators and shuttle buses are used when the aircraft is far away. In-flight catering operations mainly use in-flight catering vehicles to provide meals for aircraft and serve different types of aircraft. Cabin cleaning and sewage cleaning operations mainly use garbage trucks to clean up food waste, water trucks to replenish water consumed during the flight of the aircraft, and sewage trucks to discharge sewage generated during the flight. Aviation fuel filling operations mainly use aviation fuel trucks, which are divided into fuel tankers and fuel plug trucks, which are used to refuel aircraft. Baggage loading and unloading and cargo and mail operations mainly use conveyor belt trucks, tractors, and platform trucks to facilitate the loading and unloading of baggage, cargo, and mail. Aircraft push-out operations use tractors, which are also called trailers and are used to tow aircraft service equipment.

Through the study of the above flight support services, the characteristics of the overall flight support service operation and the characteristics of the single flight support service operation are analyzed. From an overall perspective, the flight support operation process of a single flight can be roughly divided into four parallel work processes, namely, maintenance inspection service, cabin service, cargo hold service, and aviation fuel filling service. Some parallel work processes also contain some serial flight support service operations, among which the cargo hold service includes opening the cargo door, unloading luggage and mail, loading luggage and mail, closing the cargo door and other support operations; the cabin service includes not only serial operations but also parallel operations. Through in-depth analysis and research, the flight support service process can be abstracted into a complex network topology diagram. The relationship between each flight support service operation is intricate and strongly coupled, with not only a temporal sequence, but also a logical relationship. According to the provisions of the "Airport Flight Operation Support Standards", each specific flight support service operation has its constraints and must be completed within the time specified by the standard. Among them, cabin cleaning and in-flight meal addition should only start after passengers and crew get off the plane; sewage cleaning operations should be completed 15 minutes before the flight is scheduled to close the cabin door, etc. Managers of flight support services must not only consider the constraints of flight support service operation time, but also the exclusivity of aircraft to equipment occupancy. When dispatching special vehicle equipment, factors such as aircraft model, flight schedule, and vehicle route need to be considered, so the flight support service process is nonlinear, non-stationary, and dynamic. If a certain link in the overall flight support service process is disturbed and the operation is not completed within the time specified by the standard, it will directly affect the subsequent flight support service operations, forming a ripple effect, and may even cause serious consequences.

Step 2: Select n standard action nodes from the above three types of flight support services, take the time variable of each standard action node as an original variable, reduce the dimension of the original variable to obtain m principal components and use the m principal components as the input variables of the BP neural network flight support service time estimation model, and then build the BP neural network flight support service time estimation model according to the number of principal components;

From the above analysis, we can see that the flight support service process is a complex system with nonlinear and dynamic characteristics. The BP neural network is suitable for solving this type of problem. Its basic structure is shown in Figure 2.

Select n standard action nodes from the above three types of flight support services. In the present invention, as shown in FIG3 , the number of standard action nodes is 21, which are: wheel chock, bridge docking, passenger elevator docking, cabin door opening, cargo door opening, passenger disembarkation, unloading of mail and luggage, cabin cleaning, garbage disposal, sewage cleaning, in-flight food addition, aviation fuel filling, maintenance inspection, crew boarding, loading of mail and luggage, passenger boarding, cabin door closing, cargo door closing, bridge evacuation, passenger elevator evacuation and wheel chock removal. In the complex mesh topology of FIG3 , proceed in sequence according to the direction indicated by the arrow. There is both a time sequence and a logical order between each standard action node, which cannot be reversed, so as to ensure the rationality of the flight support service process. The time variables of these 21 standard action nodes are represented by X=(x ₁ ,x ₂ ,...x _n )(n=1,2,…,21), and the time variable of each standard action node is used as an original variable.

In order to simplify the complexity of the BP neural network flight support service time estimation model, the present invention adopts the principal component analysis (PCA) method to reduce the dimension of the above original variables. The principal component analysis (PCA) method uses the idea of dimensionality reduction (linear transformation) to transform multiple original variables into several unrelated principal components with little information loss. Each principal component is a linear combination of the original variables and the principal components are unrelated to each other. The specific steps are as follows:

(1) The above original variables are standardized according to the following formula to obtain normalized variables;

n表示航班保障服务标准动作节点个数；p为航班保障服务各个标准动作节点的时间变量的输入次数；x_ij为第i个航班保障服务标准动作节点的时间变量的第j次输入；

为第j次输入的航班保障服务的时间变量的均值；in,

n represents the number of standard action nodes of flight support services; p is the number of inputs of the time variable of each standard action node of flight support services; x _ij is the jth input of the time variable of the i-th standard action node of flight support services;

is the mean value of the time variable of the flight support service input for the jth time;

(2) Calculate the elements of the standardized matrix based on the normalized variables obtained in the previous step, and construct the standardized matrix Z from all the elements;

in,

Construct the normalized matrix Z from all elements;

(3) Using the above-mentioned standardized matrix Z, the correlation coefficient matrix R is solved;

Z^T表示标准化矩阵Z的转置操作；in,

Z ^T represents the transpose operation of the normalized matrix Z;

计算出p个贡献率η，并按照累计贡献率公式

计算出p个累计贡献率，然后从大到小对累计贡献率进行排序，最后选择出m个累计贡献率≥0.85的原始变量作为主成分X＝(x₁,x₂,...x_m)，并将这些主成分作为BP神经网络航班保障服务时间估计模型的输入。(4) Solve the characteristic equation of the correlation coefficient matrix R above |R-λI _p |=0 to obtain p eigenvalues λ _j (j=1,2,…,p). According to the formula

Calculate p contribution rates η and use the cumulative contribution rate formula

Calculate p cumulative contribution rates, then sort the cumulative contribution rates from large to small, and finally select m original variables with cumulative contribution rates ≥ 0.85 as principal components X = (x ₁ , x ₂ , ... x _m ), and use these principal components as inputs to the BP neural network flight support service time estimation model.

其中α为根据经验取的常数，在本发明中求得的隐含层节点个数k＝14；根据所求问题，确定输出层节点个数q＝1，即航班保障服务时间。因此在本发明中确定的BP神经网络航班保障服务时间估计模型的结构是10-14-1，最终由此构建成BP神经网络航班保障服务时间估计模型。The number of principal components m calculated by the PCA method is used as the number of input layer nodes to construct a BP neural network flight support service time estimation model. In the present invention, the number of input layer nodes m is 10, which are respectively the time for refueling, the time for loading and unloading blocks, the time for disembarking passengers, the time for boarding passengers, the time for loading and unloading cargo, mail and baggage, the time for docking bridge, the time for unloading blocks, the time for evacuating bridges, the time for handling garbage and the time for cleaning and preparing meals. The number of hidden layer nodes is assumed according to empirical formulas. There are three common empirical formulas for the number of hidden layer nodes: k = log ₂ m,

Where α is a constant obtained based on experience. The number of hidden layer nodes k=14 obtained in the present invention is determined according to the problem being solved. The number of output layer nodes q=1 is determined, i.e., the flight support service time. Therefore, the structure of the BP neural network flight support service time estimation model determined in the present invention is 10-14-1, and the BP neural network flight support service time estimation model is finally constructed.

输入层的输出变量为

隐含层的输入变量为

隐含层的输出变量为

输入层至隐含层的连接权重为w_ij(i＝1,2,…,m；j＝1,2,…,k)；隐含层至输出层的连接权重v_jt(j＝1,2,…,k；t＝1,2,…,q)；隐含层各单元的输出阈值θ_j(j＝1,2,…,q)；输出层各单元的输出阈值γ_j(j＝1,2,…,q)；p是输入次数。在本发明中，输入变量中的x₁表示航油加注时间，x₂表示上轮挡时间，x₃表示旅客下机时间，x₄表示旅客登机时间，x₅表示装卸货邮行李时间，x₆表示廊桥对接时间，x₇表示撤轮挡时间，x₈表示廊桥撤离时间，x₉表示垃圾处理时间，x₁₀表示清洁配餐时间。用c_i(i＝1,2,…,14)表示隐含层中的各个节点，y代表航班保障服务时间。In this three-layer BP neural network flight support service time estimation model, the m input variables of the input layer are the m principal components selected by the PCA method.

The output variable of the input layer is

The input variables of the hidden layer are

The output variable of the hidden layer is

The connection weight from the input layer to the hidden layer is w _ij (i=1,2,…,m; j=1,2,…,k); the connection weight from the hidden layer to the output layer is v _jt (j=1,2,…,k; t=1,2,…,q); the output threshold value of each unit in the hidden layer is θ _j (j=1,2,…,q); the output threshold value of each unit in the output layer is γ _j (j=1,2,…,q); p is the number of input times. In the present invention, x ₁ in the input variable represents the time for refueling, x ₂ represents the time for on-block loading, x ₃ represents the time for passengers to get off the plane, x ₄ represents the time for passengers to board the plane, x ₅ represents the time for loading and unloading cargo, mail and baggage, x ₆ represents the time for docking with the bridge, x _{7 represents the time for off-block loading, x 8 represents} the time for evacuating the bridge, x ₉ represents the time for garbage disposal, and x ₁₀ represents the time for cleaning and catering. Use c _i (i=1,2,…,14) to represent each node in the hidden layer, and y represents the flight support service time.

Step 3: Preprocess and normalize the data related to the above input variables, and then randomly divide all the processed data into training samples and test samples, and then input the training samples into the BP neural network flight support service time estimation model to train the model, and finally input the test samples into the trained BP neural network flight support service time estimation model;

The data related to the input variables in the present invention are selected from the real data recorded in the actual flight support service process of a certain airport during a certain period of time. However, in the actual recording process of the airport, it is inevitable that some data will be missing due to special circumstances, so it is necessary to pre-process these data to reduce the impact of the missing data. Some flight support service data before processing are shown in Table 1.

Table 1. Data of some flight support services before processing

First, remove the start and end time codes of the flight support service operation, such as [06] in Table 1; then delete the parking position number after the flight arrives at the station, because the parking position number has little to do with the content to be studied, and finally find the duration of each flight support service operation, that is, the end time of each operation minus the start time. In Table 1, the wheel block operation is taken as an example, and the processing of other flight support service operations is similar to the wheel block operation. The processed flight support service data is shown in Table 2.

Table 2. Processed security service data

In order to improve the accuracy of the BP neural network flight support service time estimation model, the output variable _y1 and input variables _x1 , _x2 , x3, _x4 , _{x5 , x6} , _x7 , _x8 , _x9 , _x10 in the model are normalized. The formula is as follows:

Among them, u represents the input variable data after processing, _xi represents the input variable data before processing, _xmax represents the maximum input variable data; _xmin represents the minimum input variable data; _umax represents the upper limit data after processing, and _umin represents the lower limit data after processing. Assuming that the data range after processing is controlled in [0, 1], then _umax = 1, _umin = 0.

The preprocessed and normalized data are randomly divided into training samples and test samples in a ratio of 3:1. The training samples are then input into the BP neural network flight support service time estimation model to train the model. The test samples are then input into the trained BP neural network flight support service time estimation model, and whether the model meets the accuracy requirements is determined based on the error size.

Step 4: Use the improved genetic algorithm to optimize the trained BP neural network flight support service time estimation model to obtain an optimized improved genetic algorithm neural network flight support service time estimation model;

Among them, the improved genetic algorithm is to represent the chromosome in the traditional genetic algorithm as a two-layer structure and improve the corresponding operator, design the traditional mutation probability as an adaptive crossover mutation probability, and optimize the network structure and network connection weights and thresholds. The specific steps are as follows:

(1) Improved genetic algorithm two-layer structure chromosome design:

Improve the traditional chromosome. The structure of the chromosome is arranged in layers by many genes. The chromosome gene design is divided into two layers, including control genes and parameter genes. The control gene is in the upper layer, which is used to control the number of nodes in the hidden layer and optimize the structure of the BP neural network; the parameter gene is in the lower layer, which is used to optimize the connection weight and threshold of the BP neural network, and the parameter gene string of the lower layer is controlled by the control gene of the upper layer. Encode the gene, the coding of the control gene is binary, "1" represents that the corresponding gene is in an activated state, and the low-level gene string associated with this gene is valid; "0" represents that the corresponding gene is in an inactivated state, and the low-level gene string associated with this gene is invalid; the parameter gene is encoded as a real number. The designed two-layer structure chromosome and its coding diagram are shown in Figure 4. The chromosome designed by the present invention can be divided into two levels. The coding length of the control gene should be equal to the number of hidden layer nodes m, and its position should be in the upper layer of the chromosome; the position of the parameter gene should be in the lower layer of the chromosome, and its coding length should be equal to the total number of connection weights and thresholds in the chromosome (n+1)*m+(m+1)*p.

(2) Improve the fitness function design of genetic algorithm

The improved genetic algorithm should not only optimize the BP neural network structure, but also optimize the connection weights and thresholds of the BP neural network, so as to minimize the estimation error of the flight support service time and optimize the complexity of the established model. This is a dual-objective optimization problem. The designed fitness function should be able to reflect both the complexity of the BP neural network structure and the estimation accuracy of the BP neural network structure. The estimation accuracy is determined by the overall estimation error of the actual training sample data of each flight support service operation time, while the network complexity is determined by the number of hidden layer nodes of the designed BP neural network structure. The fitness function is designed as follows:

f＝αf _rmse +βf _com 0＜α,β＜1

是航班保障服务时间的估算值，y_i是航班保障服务时间的实际值，这两个值均由训练好的BP神经网络航班保障服务时间估计模型得出；N(1)、N(0)分别表示对照基因中活化和失活的神经元数量，即对照基因中1和0的数量；β和α分别表示BP神经网络结构复杂度的调整系数和BP神经网络估计精度的调整系数。Where _frmse∈ [0,1] is the root mean square error (RMSE) of the actual training sample data at each operation time of the flight support service, _fcom is the complexity of the network structure,

is the estimated value of the flight support service time, and _yi is the actual value of the flight support service time. Both values are obtained by the trained BP neural network flight support service time estimation model; N(1) and N(0) represent the number of activated and inactivated neurons in the control gene, that is, the number of 1 and 0 in the control gene; β and α represent the adjustment coefficient of the BP neural network structure complexity and the adjustment coefficient of the BP neural network estimation accuracy, respectively.

(3) Improve the selection operator of genetic algorithm

Traditional roulette and some methods based on fitness ratio usually appear "premature maturity" or "closed competition", making it impossible to retrieve, and eventually it is easy to get stuck in local extreme points instead of the maximum points. In view of this limitation, the operation methods of "best individual preservation strategy" and "random league selection strategy of size 2" are selected.

1) Best individual preservation strategy: Select the most adaptable individual in the parent population and directly select the selected individual into the next generation population. This not only preserves the best individuals in the previous generation population, but also ensures the global convergence of the genetic algorithm.

2) League selection strategy of size 2: For all individuals except the best solution in the previous generation population, two individuals are randomly selected to compare their fitness, and the individuals with better fitness are selected to enter the next generation population, and the individuals with poor fitness are eliminated until a complete offspring group is generated, which can ensure that individuals with relatively high quality can enter the next generation population.

(4) Improve the crossover operator and mutation operator of genetic algorithm

In the genetic algorithm, the control gene layer on the upper chromosome uses a single-point crossover operator and a simple mutation operator; the parameter gene layer on the lower chromosome uses a global arithmetic crossover operator and a non-uniform mutation operator. The global arithmetic crossover operator uses the superposition principle of geometric vectors to calculate each component of the intersecting previous generation vector, thereby expanding the search range of the algorithm; the non-uniform mutation operator associates mutations with the evolutionary generations of the population, and in the early stages of the evolutionary process, the number of elite individuals is relatively small, and the range of use is relatively large. In the later stages of the evolutionary process, in order to prevent the destruction of excellent individuals, the range of allowed changes is relatively narrow, so that the local optimal value can be obtained.

(5) Improve the adaptive crossover probability and mutation probability of genetic algorithm

Because the selection of crossover and mutation probabilities will lead to a decrease in the efficiency of the genetic algorithm, if the selection probability is too high, it will easily destroy the excellent individuals in the population; if the selection probability is too low, the speed of individual updates will become much slower, and it is easy to fall into "premature maturity". The calculation formula for adaptive crossover probability is as follows:

In the formula, f _c represents the crossover individual with a smaller fitness value, f _min represents the minimum fitness value in the current population, and f _a represents the average fitness value of the current population 0＜k ₁ ,k ₂ ≤1. To ensure that the algorithm can search globally, a larger adaptive crossover probability needs to be selected, generally k ₁ =1, k ₂ =1. The calculation formula for the adaptive mutation probability is as follows:

In the formula, _fm represents the fitness value of the individual to be mutated, _fmin represents the minimum fitness value in the current group, _fa represents the average fitness value of the current group, 0＜ _k3 , _k4≤1 . Since the crossover mutation operation in the genetic algorithm is a simulation of the mutation of living objects in nature during the genetic process, the probability of adaptive mutation should be relatively small, generally _k3 ＝0.5, _k4 ＝0.5.

The above steps complete the design of the improved genetic algorithm, set the operation parameters of the population, the total number is N, the maximum evolutionary generation number G, the number of hidden layer nodes assumed at the beginning (usually a larger value), etc. Randomly generate N individuals to form the initial population, divide the population into 2 offspring groups and encode the chromosome into a two-layer structure. Decode the individual to count the number of 1s in a single gene string in the upper offspring group, which is the number of hidden layer nodes of the corresponding BP neural network; decompose the actual parameter string of the lower parameter gene associated with the upper control gene into the value 1 to obtain the initial connection weight and threshold of the hidden layer node; at the same time, train the BP neural network according to the calculation formula in step (2) to calculate the fitness value. Assign the obtained connection weight and threshold to the BP neural network, use the training sample to train the BP neural network, and then use the test sample to test the BP neural network to calculate the test error. According to the different fitness values of each individual, according to the selection operator designed in step (4), enter the next generation of elite individuals, and the selected individuals use adaptive probability to perform mutation operations to generate offspring groups. Decode the upper and lower individuals in the offspring population, obtain the structure and initial connection weights and thresholds of the BP neural network, train the network multiple times, and calculate their fitness values. Determine whether the best individual fitness value can meet the set value, or whether it can be increased to the maximum evolution number, decode the individual with the best fitness value, and obtain the optimal number of hidden layer nodes and their optimal initial connection weights and thresholds. The flowchart of the improved genetic algorithm is shown in Figure 5.

The optimal number of hidden layer nodes and their optimal initial connection weights and thresholds calculated by the improved genetic algorithm are assigned to the BP neural network flight support service time estimation model. The structural parameters of the BP neural network after optimization by the improved genetic algorithm are determined as follows: population size N = 100; maximum evolutionary generations G = 200; fitness function of estimation accuracy adjustment coefficient α = 0.9, network complexity adjustment coefficient β = 0.1; coefficients k ₁ = 1, k ₂ = 1, k ₃ = 0.5, k ₄ = 0.5, k ₃ = 0.5 in adaptive crossover and mutation probability P _c , P _m. Finally, the number of input layer nodes m = 10, the number of output layer nodes is 1, and the number of hidden layer nodes is m = 14; the value range of connection weights w _ij , v _jt and thresholds θ _j , γ is [-3, 3]. As shown in Figure 6, after 103 generations of evolution, the average fitness reaches the minimum, and the average fitness basically does not change with the increase of evolutionary generations. When the evolutionary generation is 103, the corresponding number of hidden layer nodes is 7, so the optimal number of hidden layer nodes of the BP neural network is m=7.

Table 3. Optimal initial connection weights and thresholds

The corresponding optimal initial connection weights and thresholds of the improved genetic algorithm neural network flight support service time estimation model are shown in Table 3, where _x1 , _x2 , _x3 , x4, _x5 , _x6 , _x7 , _x8 , _x9 , _x10 represent 10 input variables, _c1 , _{c2 , c3} , _c4 , _c5 , _c6 , _c7 represent 7 hidden layer variables, y represents the output variable, θ represents the threshold of the hidden layer, and γ represents the threshold of the output layer.

A BP neural network flight support service time estimation model with a network structure of 10-7-1 is established, and the optimal initial weights and thresholds in Table 3 are assigned. The Sigmoid function is selected as the excitation function, the learning rate is 0.01, the step size is 0.9, the maximum number of training times is 2000, and the training expected value is 0.01. Finally, the optimized improved genetic algorithm neural network flight support service time estimation model is formed.

Step 5: Process the flight support service data of any airport according to the above steps 1 and 2 and input it into the optimized neural network flight support service time estimation model of the improved genetic algorithm. The output of the optimized neural network flight support service time estimation model of the improved genetic algorithm is the estimated value of the flight support service time of the airport.

In order to verify the effect of the present invention, the inventors used three algorithms, namely the neural network flight support service time estimation model (AMGA-BP model) of the above-mentioned optimized improved genetic algorithm, the neural network flight support service time estimation model (GA-BP model) of the traditional genetic algorithm and the neural network flight support service time estimation model (BP model) without adding an algorithm, to estimate the flight support service time.

其公式为：The absolute value of relative error RE, mean absolute error MAE and Hilton coefficient TIC are used as the standard to measure the advantages and disadvantages of the three algorithms. Among them, the absolute value RE is used to calculate the estimated error of each flight support service time; the mean absolute error MAE and Hilton coefficient TIC measure the accuracy of the algorithm by calculating the overall error between the estimated value and the true value of the flight support service time. Assume that there are a total of N flight support service estimates in the test set, the actual value of the flight support service time is y _i , and the estimated value is

The formula is:

Table 4. Estimated effect evaluation index values

45 groups of samples are randomly selected from the test set. The evaluation index values of the estimated effect are shown in Table 4. The comparison of the guarantee service time errors estimated by different algorithms is shown in Figure 7.

It can be seen from Table 4 that the mean absolute error MAE value of the AMGA-BP model is smaller than that of the GA-BP model and the BP model, and the Hilton coefficient TIC value is smaller than that of the GA-BP model and the BP model, indicating that the AMGA-BP model has a higher estimation accuracy. It can be seen from Figure 7 that the estimation error curve of the AMGA-BP model is basically lower than the error curves of the GA-BP model and the BP model; the error estimation curves of the GA-BP model and the BP model have basically the same change trend. The estimated values of the guaranteed service time of the three models are compared with the actual values, as shown in Figure 8.

(AMGA-BP model), the traditional genetic algorithm neural network flight support service time estimation model (GA-BP model) and the neural network flight support service time estimation model without algorithm (BP model)

In summary, compared with the GA-BP model and the BP model, the AMGA-BP model can better estimate the flight support service time and better handle nonlinear problems. It is worth noting that the AMGA-BP model is not very accurate in estimating the flight support service time for each flight, but its overall estimation is relatively robust. As can be seen from Figure 7, for some flights, the accuracy of the AMGA-BP model in estimating the flight support service time is much higher than that of the GA-BP model and the BP model, but the error is still not particularly small, which indicates that the AMGA-BP model needs to further improve the adaptability of the flight support service time under different conditions. Figure 8 compares the estimated values of the flight support service time of the AMGA-BP model, the GA-BP model and the BP model with the actual flight support service time. It can be clearly seen that the estimated value of the BP model is the most different from the actual value, followed by the GA-BP model. This shows that for the complex nonlinear problem of flight support service, the traditional genetic algorithm alone cannot meet the expected requirements, and it must be improved on the basis of the neural network flight support service time estimation model. The flight support service time estimated by the AMGA-BP model is closest to the actual flight support service time.

Claims (5)

1. A neural network flight support service time estimation method based on an improved genetic algorithm is characterized by comprising the following steps: the method comprises the following steps which are carried out in sequence:

step 1: analyzing flight support service flow of airport scene operation, dividing flight support service into three types according to different controlled objects, and respectively serving an aircraft, passengers and luggage, goods and mails;

and 2, step: selecting n standard action nodes from the three types of flight support services, taking the time variable of each standard action node as an original variable, reducing the dimensions of the original variable to obtain m main components, taking the m main components as input variables of a BP neural network flight support service time estimation model, and then constructing the BP neural network flight support service time estimation model according to the number of the main components;

and step 3: preprocessing and normalizing the data related to the input variables, randomly dividing all the processed data into training samples and testing samples, inputting the training samples into a BP neural network flight guarantee service time estimation model to train the model, and finally inputting the testing samples into the trained BP neural network flight guarantee service time estimation model;

and 4, step 4: optimizing the trained BP neural network flight guarantee service time estimation model by using an improved genetic algorithm to obtain an optimized neural network flight guarantee service time estimation model of the improved genetic algorithm;

and 5: and (3) inputting the flight guarantee service data of any airport into the optimized neural network flight guarantee service time estimation model of the improved genetic algorithm after being processed according to the steps 1 and 2, wherein the output of the optimized neural network flight guarantee service time estimation model of the improved genetic algorithm is the estimated value of the flight guarantee service time of the airport.

2. The neural network flight support service time estimation method based on the improved genetic algorithm as claimed in claim 1, wherein: in step 1), the method for analyzing the flight support service flow of the airport scene operation, and dividing the flight support service into three types according to the difference of controlled objects, wherein the three types are respectively the service of an aircraft, the service of a passenger and the service of luggage and luggage postal service, and the method comprises the following steps:

the aircraft service mainly comprises: the method comprises the following steps of performing operations including wheel block mounting, gallery bridge butt joint, passenger ladder vehicle butt joint, cabin cleaning, garbage treatment, sewage cleaning, feeding, aviation oil filling, gallery bridge evacuation, passenger ladder vehicle evacuation, locomotive inspection and wheel block removal; the service of the passenger mainly comprises the operations of opening a passenger cabin door, taking off the passenger, boarding the passenger and closing the passenger cabin door; the service of the luggage goods and the mails mainly comprises the operations of opening a goods cabin door, unloading the mail luggage, loading the mail luggage and closing the goods cabin door.

3. The neural network flight support service time estimation method based on the improved genetic algorithm as claimed in claim 1, wherein: in step 2), the method for selecting n standard action nodes from the three types of flight support services, using the time variable of each standard action node as an original variable, performing dimension reduction on the original variable to obtain m principal components, using the m principal components as input variables of a BP neural network flight support service time estimation model, and then constructing the BP neural network flight support service time estimation model according to the number of the principal components comprises the following steps: firstly, selecting standard action nodes as original variables from flight support services required by analysis according to the step 1), then using a principal component analysis method to reduce the dimensions of the original variables, using the original variables after the dimension reduction as input layer nodes of a BP neural network, calculating the number of nodes of a corresponding hidden layer output layer, and finally constructing a BP neural network flight support service time estimation model.

4. The neural network flight support service time estimation method based on the improved genetic algorithm as claimed in claim 1, wherein: in step 3), the method for preprocessing and normalizing the data related to the input variables, then randomly dividing all the processed data into training samples and testing samples, then inputting the training samples into the BP neural network flight guarantee service time estimation model to train the model, and finally inputting the testing samples into the trained BP neural network flight guarantee service time estimation model comprises the following steps: firstly, selecting data related to input variables of a BP neural network flight guarantee service time estimation model from actual data of an airport flight guarantee service process, processing the data by using a normalization method to reduce influence caused by partial data loss, dividing the data after normalization into training samples and testing samples, then inputting the training samples and the testing samples into the BP neural network flight guarantee service time estimation model in sequence to train and test the model, and finally obtaining the trained BP neural network flight guarantee service time estimation model.

5. The neural network flight support service time estimation method based on the improved genetic algorithm as claimed in claim 1, wherein: in step 4), the method for obtaining the optimized neural network flight guarantee service time estimation model of the improved genetic algorithm by optimizing the trained BP neural network flight guarantee service time estimation model by using the improved genetic algorithm is as follows: firstly, improving two-layer structure chromosomes of a genetic algorithm, improving a fitness function of the genetic algorithm, secondly, selecting an operation method of an optimal individual preservation strategy and a random tournament selection strategy to improve a selection operator of the genetic algorithm, avoiding the reduction of the efficiency of the genetic algorithm caused by too high selection probability from damaging excellent individuals in a population, simultaneously avoiding the selection probability from falling into 'premature' due to too low selection probability to design self-adaptive cross probability and variation probability, and finally completing a neural network flight guarantee service time estimation model for improving the genetic algorithm.

Images