Airport single flight noise prediction method based on BP neural network

Abstract

The invention relates to the technical field of airport noise prediction, and particularly discloses an airport single flight noise prediction method based on a BP neural network, and the method comprises the following steps: firstly screening airport single flight noise original data, and carrying out the data normalization processing; establishing a training set and a verification set for the screened original data, and constructing a prediction model training set; training the training set to obtain a noise prediction model meeting the precision; performing secondary training on the noise prediction model obtained by training to obtain a final noise prediction model; and predicting the noise influence of a specific model by using the obtained noise prediction model. According to the airport single flight noise prediction method based on the BP neural network, LA prediction can be carried out based on the airplane state, and LWECPN and YLDN distribution prediction can be carried out according to the flight plan, so that flight operation is optimized, and the noise influence is reduced.

Description

Technical Field

[0001] The invention relates to the technical field of airport noise prediction, and in particular to an airport single flight noise prediction method based on BP neural network. Background Art

[0002] Civil aviation noise is one of the main factors hindering the development of civil aviation. With the increase in global air traffic demand, aviation noise pollution will become more and more serious. Relevant medical research shows that long-term exposure to aviation noise can cause hearing damage, sleep disorders, and increase the risk of cardiovascular disease. Not only that, aviation noise can also hinder the development of children's learning and cognitive abilities. The prediction method of aviation noise will help us quantify the negative impact of aviation noise on the environment and provide a basis for the control of aviation noise.

[0003] In order to quantify the negative impact of aircraft noise on the surrounding environment, it is urgent to establish a rapid prediction method for aircraft noise. At present, the most representative aircraft noise calculation method is the INM model (Integrated Noise Model) developed by the Federal Aviation Administration (FAA). Traditional calculation methods represented by INM are usually based on noise theory and have strong versatility. However, since the INM calculation method requires detailed information (such as airport information, runway information, aircraft model, engine type, flight operation information, flight information, etc.), it is difficult to meet the input data requirements in actual use. Summary of the invention

[0004] The purpose of the present invention is to provide an airport single flight noise prediction method based on BP neural network. Taking into account reducing the difficulty of applying the prediction model and improving the prediction speed under the condition of ensuring the prediction accuracy, the prediction of the regional ground A sound pressure level distribution is realized by traversing the grid points in the area around the aircraft. Compared with the traditional model, the model proposed by the present invention is easier to use and has lower cost.

[0005] To achieve the above object, the present invention provides an airport single flight noise prediction method based on BP neural network, comprising the following steps:

[0006] S1. Filter the original noise data of a single flight at the airport and perform data normalization processing;

[0007] S2, establish a training set and a validation set with the screened original data, and construct a prediction model training set S;

[0008] S3, training the training set S in step S2 to obtain a noise prediction model that meets the accuracy requirements;

[0009] S4, performing secondary training on the noise prediction model obtained by training in step S3 to obtain a final noise prediction model;

[0010] S5. Using the noise prediction model obtained in step S4, predict the noise impact of a specific aircraft model.

[0011] Preferably, in step S1, the track data and noise monitoring data are correlated and preprocessed, and data that cannot be used for model training due to low quality are cleaned up to obtain higher quality noise point monitoring data and aircraft track data.

[0012] Preferably, in step S2, the filtered raw data is used to establish a training set and a validation set, and the prediction model training set S is specifically constructed as follows:

[0013] Half of the original data after screening is used as the training set, and the other half is used as the validation set to construct the prediction model training set S, S = (x 1 ,x 2 ,…,x n ); in the prediction model structure, n is the total number of training set samples, x is the input variable, y is the output variable, and the neurons are activated by the sigmoid activation function. The mapping relationship is:

[0014] Preferably, step S3 specifically includes the following steps:

[0015] S31. First, a noise distribution prediction problem is set;

[0016] S32, assigning a weight to each piece of sample data in the training set S in step S2, assuming that each piece of sample data is equally important;

[0017] S33, assuming that k prediction models are trained, different hidden layer structures are set for screening, and MATLAB software and Neural Net Fitting toolbox are used for training respectively to obtain multiple prediction models, and the appropriate number of hidden layers is selected by observing the prediction model results;

[0018] S34, setting neural grid parameters;

[0019] Let TF1 be the transfer function from the input to the first hidden layer, TF2 be the transfer function from the first hidden layer to the second hidden layer, TF3 be the output layer transfer function, set TF1 and TF2 to logical sigmoid functions; set TF3 to linear activation function;

[0020] S35, setting model training parameters, including training times, training targets and learning rates, and using a gradient descent algorithm to update model parameters;

[0021] S36, neural network training;

[0022] S37, performing data denormalization;

[0023] S38. Calculate the root mean square error (MSE) of the validation set and select the model with the smaller MES value as the noise prediction model.

[0024] Preferably, in step S4, the optimal model obtained by training step S3 is obtained, the prediction problem in step S2 is simulated and compared with the actual noise monitoring data, and the screened data is subjected to a second round of model training to obtain the final noise prediction model.

[0025] Preferably, the screened data is to remove bad pixels greatly affected by environmental noise, that is, to remove data with monitoring values ​​higher than predicted values, and use the remaining data for a second round of model training.

[0026] Therefore, the present invention adopts the above-mentioned airport single flight noise prediction method based on BP neural network, and the beneficial effects are as follows:

[0027] (1) The present invention proposes an airport noise prediction agent model based on BP neural network. The model is trained using noise monitoring data around the airport and aircraft flight data. The model can predict the A sound pressure level of the target point based on four parameters, namely, aircraft altitude, aircraft speed, aircraft flight direction, and horizontal distance between the aircraft and the target point. Compared with the traditional model, the model proposed by the present invention is easier to use and has lower cost.

[0028] (2) The method provided by the present invention has few input variables and fast calculation speed. The sound pressure level A value calculated by the prediction model can be used to calculate the weighted equivalent continuous perceived noise level L of the noise evaluation index specified in the current standard. WECPN and the annual average day and night equivalent sound level YL DN The calculation of the contour lines of the present invention facilitates noise assessment.

[0029] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the prediction model structure of an embodiment of a method for predicting noise of a single flight at an airport based on a BP neural network according to the present invention;

[0031] Figure 2 It is a system diagram of airport single flight noise prediction according to an embodiment of an airport single flight noise prediction method based on BP neural network of the present invention;

[0032] Figure 3 It is a flow chart of a training learner of an embodiment of a method for predicting airport single flight noise based on BP neural network of the present invention;

[0033] Figure 4It is a schematic diagram of a prediction model structure actually used in an embodiment of a method for predicting noise of a single flight at an airport based on a BP neural network of the present invention;

[0034] Figure 5 It is a noise level data diagram of the Xilihe aircraft noise monitoring point at 07:00-08:00 on April 24, 2023 in an embodiment of an airport single flight noise prediction method based on a BP neural network of the present invention;

[0035] Figure 6 It is a noise level data diagram of the Gangqiying aircraft noise monitoring point at 06:00-07:00 on April 24, 2023 in an embodiment of an airport single flight noise prediction method based on a BP neural network of the present invention;

[0036] Figure 7 It is a noise level data diagram of the Gangqiying aircraft noise monitoring point at 07:00-08:00 on April 24, 2023 in an embodiment of an airport single flight noise prediction method based on a BP neural network of the present invention;

[0037] Figure 8 This is an example of a method for predicting noise of a single flight at an airport based on a BP neural network, a noise level data diagram of the Nanshunmin aircraft noise monitoring point at 07:00-08:00 on April 24, 2023;

[0038] Fig. 9 It is a noise level data diagram of the Nanshunmin aircraft noise monitoring point at 08:00-09:00 on April 24, 2023 in an embodiment of an airport single flight noise prediction method based on a BP neural network of the present invention;

[0039] Fig.10 It is a noise level data diagram of the Dafa Village aircraft noise monitoring point at 07:00-08:00 on April 24, 2023 in an embodiment of an airport single flight noise prediction method based on a BP neural network of the present invention;

[0040] Fig.11 It is a schematic diagram of the aircraft noise distribution prediction range of an embodiment of a method for predicting noise of a single flight at an airport based on a BP neural network of the present invention;

[0041] Fig.12It is a sound level distribution diagram obtained by training a prediction model with a 3×2×1 hidden layer structure in an embodiment of an airport single flight noise prediction method based on a BP neural network of the present invention, wherein (a) is a sound level distribution diagram obtained by training the prediction model with a 3×2×1 hidden layer structure for the first time, (b) is a sound level distribution diagram obtained by training the prediction model with a 3×2×1 hidden layer structure for the 14th time, (c) is a sound level distribution diagram obtained by training the prediction model with a 3×2×1 hidden layer structure for the 25th time, and (d) is a sound level distribution diagram obtained by training the prediction model with a 3×2×1 hidden layer structure for the 31st time;

[0042] Fig.13 It is a sound level distribution diagram obtained by training a prediction model with a 4×2×1 hidden layer structure in an embodiment of a method for predicting noise of a single flight at an airport based on a BP neural network of the present invention, wherein (a) is a sound level distribution diagram obtained by training the prediction model with a 4×2×1 hidden layer structure for the 14th time, (b) is a sound level distribution diagram obtained by training the prediction model with a 4×2×1 hidden layer structure for the 16th time, (c) is a sound level distribution diagram obtained by training the prediction model with a 4×2×1 hidden layer structure for the 21st time, and (d) is a sound level distribution diagram obtained by training the prediction model with a 4×2×1 hidden layer structure for the 42nd time;

[0043] Fig.14 It is a sound level distribution diagram obtained by training a prediction model with a 5×2×1 hidden layer structure in an embodiment of an airport single flight noise prediction method based on a BP neural network of the present invention, wherein (a) is a sound level distribution diagram obtained by training a prediction model with a 5×2×1 hidden layer structure, (b) is a sound level distribution diagram obtained by training the prediction model with a 5×2×1 hidden layer structure for the 16th time, (c) is a sound level distribution diagram obtained by training the prediction model with a 5×2×1 hidden layer structure for the 21st time, and (d) is a sound level distribution diagram obtained by training the prediction model with a 5×2×1 hidden layer structure for the 42nd time;

[0044] Fig.15 1 is a diagram of an embodiment of an airport single flight noise prediction method based on a BP neural network of the present invention, wherein (a) is a sound level distribution diagram obtained by training with the prediction model for the first time, (b) is a sound level distribution diagram obtained by training with the prediction model for the 24th time, (c) is a sound level distribution diagram obtained by training with the prediction model for the 8th time, and (d) is a sound level distribution diagram obtained by training with the prediction model for the 18th time;

[0045] Fig.16It is the first round training result of the B738 noise model in an embodiment of an airport single flight noise prediction method based on BP neural network of the present invention;

[0046] Fig.17 It is a schematic diagram of the absolute value of the prediction deviation of the B738 noise model in the first round of validation set according to an embodiment of the airport single flight noise prediction method based on the BP neural network of the present invention;

[0047] Fig.18 It is the prediction result of each noise monitoring point in the first round of the B738 noise model in an embodiment of the airport single flight noise prediction method based on BP neural network of the present invention;

[0048] Fig.19 It is the second round training result of the B738 noise model in an embodiment of an airport single flight noise prediction method based on BP neural network of the present invention;

[0049] Fig. 20 It is the prediction result of each noise monitoring point in the second round of the B738 noise model in an embodiment of the airport single flight noise prediction method based on BP neural network of the present invention;

[0050] Fig.21 It is a schematic diagram of a departure track of a B738 noise prediction model used as an example in an embodiment of a method for predicting noise of a single flight at an airport based on a BP neural network according to the present invention;

[0051] Fig. 22 The present invention is a method for predicting single-flight noise at an airport based on a BP neural network. The ...

[0052] Fig.23 It is the equivalent continuous A sound level distribution prediction result of the B738 noise prediction model of the airport single flight noise prediction method embodiment based on BP neural network of the present invention;

[0053] Fig.24 The weighted equivalent continuous perceived noise level L of an airport single flight noise prediction method embodiment based on BP neural network of the present invention is WECPN The prediction results;

[0054] Fig.25The annual average day and night equivalent sound level YL of an embodiment of a method for predicting airport single flight noise based on BP neural network of the present invention is DN Forecast results (5 flights per hour);

[0055] Fig.26 The annual average day and night equivalent sound level YL of an embodiment of a method for predicting airport single flight noise based on BP neural network of the present invention is DN Prediction results (10 flights per hour);

[0056] Fig. 27 The annual average day and night equivalent sound level YL of an embodiment of a method for predicting airport single flight noise based on BP neural network of the present invention is DN Forecast results (15 flights per hour). DETAILED DESCRIPTION

[0057] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0058] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0059] A BP neural network airport single flight noise prediction method comprises the following steps:

[0060] S1. Filter the original noise data of a single flight at the airport and perform data normalization processing;

[0061] First, the track data and noise monitoring data are correlated and preprocessed to clean up the data that cannot be used for model training due to low quality, so as to obtain higher-quality noise point monitoring data and aircraft track data.

[0062] S2. Use the filtered raw data to establish a training set and a validation set, and construct the prediction model training set S as follows:

[0063] Half of the original data after screening is used as the training set, and the other half is used as the validation set to construct the prediction model training set S, S = (x 1 ,x 2 ,…,x n ); the prediction model structure is as follows Figure 1 , where n is the total number of training set samples, x is the input variable, y is the output variable, and the neurons are activated using the sigmoid activation function. The mapping relationship is: v ij represents the weight of the i-th input variable and the j-th hidden neuron, and w is the weight of the connection between the neuron and y.

[0064] S3, such as Figure 3As shown, the training set S of step S2 is trained to obtain a noise prediction model that meets the accuracy requirements, specifically including the following steps:

[0065] S31. First, a noise distribution prediction problem is set;

[0066] S32, assigning a weight to each piece of sample data in the training set S in step S2, assuming that each piece of sample data is equally important;

[0067] S33, assuming that k prediction models are trained, different hidden layer structures are set for screening, and MATLAB software and Neural Net Fitting toolbox are used for training respectively to obtain multiple prediction models, and the appropriate number of hidden layers is selected by observing the prediction model results;

[0068] S34, using a gradient descent algorithm to adjust the prediction model weight parameters;

[0069] S34, setting neural grid parameters;

[0070] Let TF1 be the transfer function from the input to the first hidden layer, TF2 be the transfer function from the first hidden layer to the second hidden layer, and TF3 be the output layer transfer function. If the training structure is not ideal, you can try to change the transfer function.

[0071] Set TF1 and TF2 as logistic sigmoid functions; set TF3 as linear activation function;

[0072] S35. Set model training parameters. Usually, the number of training times and training targets are considered as two key factors. The learning rate is also an important parameter.

[0073] 1. Number of training times: You need to decide the number of iterations for model training. Too many iterations may lead to overfitting, while too few iterations may cause the model to fail to fully learn.

[0074] 2. Training target: You can set a target error value, such as 0.05. Training will continue until the model error reaches or is less than this target value.

[0075] 3. Learning rate: The learning rate controls the step size of parameter updates. If the learning rate is set too large, although it will speed up the convergence speed at the beginning, it may cause oscillations when approaching the optimal point, resulting in failure to converge.

[0076] Specify that the training function uses the gradient descent algorithm to update the model's parameters, such as weights and biases, to minimize this difference.

[0077] S36, neural network training process;

[0078] 1. Set up an extended sample input matrix, input the input features of each sample, and plan to calculate the first hidden layer, the second hidden layer, and the network output during forward propagation. In each training iteration, first perform forward propagation, then calculate the error to determine whether the target error is reached.

[0079] 2. If the error is still greater than the target error, perform backpropagation to calculate the impact of the error on the weights.

[0080] 3. Update the weights and continue to the next iteration.

[0081] S37, performing data denormalization;

[0082] When calling the mapminmax function, the data is normalized to the range of [0, 1] by default. This function returns the normalized sample data and also provides a processing setting (ps), which is mainly used to denormalize the result or use the same setting when normalizing another set of data.

[0083] S38. Calculate the root mean square error (MSE) of the validation set and select the model with the smaller MES value as the noise prediction model.

[0084] S4. Perform secondary training on the noise prediction model trained in step S3 to obtain the final noise prediction model. According to the optimal model trained in step S3, simulate the prediction problem in step S2 and compare it with the actual noise monitoring data. Perform a second round of model training on the selected data. The selected data is to remove bad points that are greatly affected by environmental noise, that is, remove data whose monitoring values ​​are higher than the predicted values. Perform a second round of model training on the remaining data to obtain the final noise prediction model.

[0085] S5. Using the noise prediction model obtained in step S4, predict the noise impact of a specific aircraft model.

[0086] Embodiment 1

[0087] First, determine the airport noise evaluation index

[0088] (1) Weighted equivalent continuous perceived noise level L WECPN

[0089] ICAO recommends to all member states the effective perceived noise level L EPN The weighted equivalent continuous perceived noise level includes the number of aircraft take-offs and landings and the time period, which can objectively calculate the airport noise level. my country's airport noise standards GB9660-1988 and GB9661-1988 stipulate that the weighted equivalent continuous perceived noise level L WECPN As an evaluation quantity, its expression is:

[0090]

[0091] Where N 1 Indicates the number of flights during the daytime from 07:00 to 19:00;

[0092] N 2 Indicates the flight schedule from 19:00 to 22:00 in the evening;

[0093] N 3 Indicates the number of flights between 22:00 and 07:00 at night;

[0094] It represents the average effective perceived noise level of all flights.

[0095] L EPN Substituting the calculation formula into (1) and replacing 39.4 with 10lg8640, we can get:

[0096]

[0097] T 0 =10Substitute into (2), we get:

[0098]

[0099] Weighted equivalent continuous perceived noise level (L WECPN ) and equivalent sound level (L eq ) has the following approximate relationship:

[0100] L WECPN ≈L eq +13 (4)

[0101] The "Aircraft Noise Environment Standards around Airports" GB96060-88 stipulates L WECPN The standard values ​​are shown in Table 1:

[0102] Table 1 Aircraft weighted equivalent continuous perceived noise level standard values

[0103]

[0104] Category I areas: special residential areas; residential, cultural and educational areas.

[0105] Category II area: living area other than Category I area.

[0106] (2) Annual average day and night equivalent sound level YL DN

[0107] Different from the weighted equivalent continuous perceived noise level, the daytime equivalent sound level is the noise exposure level L in the aircraft. AEBased on the noise energy obtained by combining the number of flights during the day and at night, the formula is as follows:

[0108]

[0109] Where L AEi represents the noise exposure level of the i-th aircraft during the day;

[0110] L AEj represents the noise exposure level of the jth aircraft at night;

[0111] N d Indicates the number of aircraft movements during the day;

[0112] N n Indicates the number of aircraft movements at night.

[0113] The day is divided into two periods, 07:00-22:00 during the day and 22:00-07:00 at night, and the night period is weighted 10 times to indicate the impact of night noise on people.

[0114] The annual average day and night equivalent sound level is the annual energy average value of the day and night equivalent sound level every day, expressed in YL DN Indicated by dB(A).

[0115]

[0116] The "Aircraft Noise Environmental Quality Standards for Areas Surrounding Airports" (the third draft for comments) stipulates the annual average daytime and nighttime equivalent sound level limits for aircraft noise as follows:

[0117] Table 2 Annual average day and night equivalent sound level limits of aircraft noise

[0118]

[0119] Class I land: urban and rural land that is sensitive to aircraft noise, including residential, educational and scientific research, medical and health and other similar land.

[0120] Class II land: urban and rural land that is more sensitive to aircraft noise, including administrative offices, culture and art, commercial services and other similar land.

[0121] Class III land: urban and rural land that is less sensitive to aircraft noise, including industrial production, logistics warehousing, sports and entertainment, parks and squares, and other similar land.

[0122] Category IV land: urban and rural land that is not sensitive to aircraft noise, including agricultural production, mining production, transportation facilities, public facilities and other similar land.

[0123] like Figure 2As shown in the figure, the airport single flight noise prediction process based on BP neural network can be divided into four steps: data preparation, model training, model verification and engineering application:

[0124] In the data preparation stage, first, the track data and noise monitoring data are correlated and preprocessed to clean up the data that cannot be used for model training due to low quality; then, a training dataset for training the model and a test dataset for evaluating the model performance are constructed.

[0125] During the model training and validation phase, the training dataset is used for model training, and the test dataset is used for model performance testing.

[0126] In the engineering application stage, the actual track data is input into the trained and verified noise prediction model to obtain the predicted noise data and perform data visualization.

[0127] 1. Model type selection

[0128] The BP neural network model is used for prediction model training. The BP neural network model is a multi-layer feedforward neural network trained according to the error back propagation algorithm. Its core idea is gradient descent. It calculates the error between the predicted result and the actual result, and propagates the error back from the output layer to the hidden layer and input layer. The weight value is adjusted according to the error, and the predicted result and the actual result are made closer and closer through continuous iteration.

[0129] The prediction model structure used in this embodiment is as follows Figure 4 As shown in the figure, the input variables are: aircraft altitude, aircraft speed, angle between aircraft direction and the line connecting the aircraft and the monitoring point, and horizontal distance between the aircraft and the monitoring point. The output variable is the instantaneous A sound level at the monitoring point.

[0130] A. Build a model training dataset

[0131] There are 28 noise monitoring points in the airport. According to the quality analysis results of the airport noise monitoring data, this embodiment selects the data of the four points with the best monitoring quality from among the 28 monitoring points as the training and testing data sets for the prediction model.

[0132] Compared with daytime, the ambient noise and rail traffic noise at night are relatively small. Therefore, the monitoring data between 01:00-09:00 on April 24, 2023 are selected as the training data set. In addition, the monitoring data of each noise monitoring point is divided into three categories according to the distance between the aircraft and the point, comparing the flight track data. Figure 5 As shown, the track data within 5000m from the monitoring point is represented by yellow dots, and the track data within 2500m from the monitoring point is represented by red dots. It can be clearly seen that the closer the aircraft is to the monitoring point, the greater the instantaneous value of the noise, which is in line with the law of aircraft noise propagation. Figure 5-Figure 10 They respectively represent the difference in the number of aircraft tracks at the same monitoring point at different times.

[0133] Table 3 shows the statistical results of the types of flights arriving and departing between 1:00 and 9:00 on April 24, 2023. In this embodiment, the B738 model is selected for training the noise prediction model.

[0134] Table 3 Statistical results of flights arriving and departing (1:00-8:00, April 24)

[0135] A319 A320 A321 A332 A333 B737 B738 B773 B788 2 21 23 4 6 4 36 1 1

[0136] Model training steps

[0137] (1) Noise prediction problem setting

[0138] Since the proxy model is a black box model, the user cannot directly judge the model performance. Therefore, this embodiment first sets a noise distribution prediction problem. The trained proxy model is used to predict the ground noise distribution under the assumed aircraft state, so that the user can intuitively judge whether the model is reasonable.

[0139] Assume the aircraft status: the aircraft is directly above the Gangqiying monitoring point, the flight direction is -7°, the flight altitude is 20m, and the flight speed is 0.049 nautical miles / s. Prediction target: The instantaneous A sound level distribution within a square area with a side length of 5km and the Gangqiying monitoring point as the center. The grid size is 50m×50m, and there are 100×100 grids in the area, such as Fig.11 shown.

[0140] (2) Model structure selection

[0141] First, we need to select a suitable model structure. We select three different structures with hidden layer neurons of 3×2×1, 4×2×1, and 5×3×1 for model training, and use the trained model to predict the instantaneous A sound level distribution in the hypothetical problem in (1). We train the model of each structure 40 times. Based on the 40 A sound level distribution prediction results, we intuitively select a reasonable model structure.

[0142] (3) Dataset rationality verification

[0143] This embodiment analyzes the monitoring data of 28 monitoring points and selects the noise data of the 4 points with the best detection quality as the model training data set. Therefore, it is necessary to verify the rationality of the data set. At the same time, among the 22 monitoring points other than the 4 best points, 7 points with better monitoring quality are selected. The data set consisting of the noise monitoring data of the 4 best points is numbered 1, and the data set consisting of the noise monitoring data of the 7 points with slightly worse quality is numbered 2. The model structure selected in (2) is used to perform model training based on these two data sets respectively. Based on the A sound level distribution prediction results, the rationality of the data set is intuitively judged.

[0144] (4) Model training and verification

[0145] Take the B738 model as an example for model training and validation. The data set is randomly divided into two parts, namely the training set and the validation set. The training set is used for model training, and the validation set is used to calculate the mean square error (MSE). The number of training times is 40.

[0146] (5) Example Application

[0147] Among the trained models, the model with the lowest mean square error is selected for example application. A departure track and an approach track are selected to calculate the weighted equivalent continuous perceived noise level L WECPN and the annual average day and night equivalent sound level YL DN Predictions of the distributions of these two noise indicators.

[0148] 3. Analysis of prediction results

[0149] Model structure selection

[0150] Figure 12-14 The results of using models with different neuron structures to predict the distribution of sound levels in A are shown. a, b, c, and d in the figure represent four types of simulation results obtained from the A training set. Figure b has the best simulation effect and fits the law of the data. Even if the data set is changed later, the effect will not be very bad. The fitting of figure d does not fit the general law, which is underfitting. Compared with figure b, the degree of fitting of figures a and c is greater. The model is too complex and there is overfitting. Fig.13 and Fig.14 The A-level distributions calculated using the 4×2×1 and 5×3×1 training models are shown. Fig.12 Similar, but overall forecast performance is not as good as Fig.12 .

[0151] As the complexity of the model structure increases, the A-level distribution becomes more and more bizarre and more likely to fall into the local optimum. This shows that when the model complexity increases, the overfitting phenomenon becomes more serious. Therefore, in the following work, 3×2×1 is selected as the hidden layer neuron structure.

[0152] Verify the rationality of the data set.

[0153] The model was trained 40 times using the No. 2 dataset, and the trained model was used to predict the A sound level distribution. Four typical results were selected as follows: Fig.15 In the calculation of aircraft noise propagation, the aircraft can be considered as a point sound source, so the instantaneous A-level distribution contour of the ground should be approximately a circle. Fig.15 The prediction result in the example deviates seriously from the common sense that the contour line should be approximately a circle. Therefore, this also verifies the effectiveness of the data cleaning work in the present invention. In the following work, data set No. 1 will be used for model training.

[0154] 3. Model training and verification

[0155] There are 2537 sets of data in the dataset used for the B738 noise prediction model. The dataset is randomly divided into two parts, namely the training set and the validation set. The training set is used for model training, and the validation set is used to calculate the mean square error (MSE). The number of training times is 40. The A-level distribution calculated by the training model and the mean square error calculated by the validation set are shown in the figure below. Fig.16 As shown, it is summarized in the following Table 4:

[0156] exist Fig.16 In the present invention, four training trend features in the validation set are selected. Among them, model a has an overfitting phenomenon, while the prediction results of models c and d show a phenomenon of extending to both sides.

[0157] Table 4 B738 noise model training results (first round)

[0158]

[0159] The model with the lowest mean square error in the validation set in Table 4 is the one obtained by training for the 12th time. The absolute value of the prediction deviation of the validation set is as follows Fig.17 As shown, the proportion of samples with an absolute value of prediction deviation below 10% is 95.43%.

[0160] The trained prediction model was used to predict the A-level monitoring values ​​of 12 B738 flights (11 peaks were at Gangqiying and 1 peak was at Xilihe). The prediction results are as follows: Fig.18As shown in the figure, a represents the departure noise of JDL2618 at the just-positioned location; b represents the departure noise of CCA8383 at the just-positioned location; c represents the departure noise of CUA5507 at the just-positioned location; d represents the departure noise of CUA5515 at the just-positioned location; e represents the departure noise of CUA5515 at the just-positioned location; f represents the departure noise of CUA5667 at the just-positioned location.

[0161] from Fig.18 It can be found that the prediction results can basically reflect the changing trend of the monitoring values. However, when the aircraft is far away from the point, the noise level caused by the aircraft is low, so the impact of environmental noise will be more obvious, resulting in high predicted values ​​for these points; the prediction curves for some locations are not smooth. Therefore, the monitoring data where the aircraft is more than 2500m away from the monitoring point are screened to remove bad points that are greatly affected by environmental noise. The specific method is to remove the data where the monitoring value is more than 3dB higher than the predicted value, and use the remaining data for the second round of model training. The training results are as follows: Fig.19 As shown in Table 5, the following training results are obtained:

[0162] Table 5 selects the four training results with the smallest mean square error, which are more consistent with the noise propagation logic than the first round in terms of image quality.

[0163] Table 5 B738 noise model training results (second round)

[0164]

[0165] The optimal model obtained from the second round of training is used to predict the A sound level at the monitoring point again. The prediction results are as follows: Fig. 20 As shown in the figure, a represents the departure noise prediction result of JDL2618 at Gangqiying point; b represents the departure noise prediction result of CCA8383 at Gangqiying point; c represents the departure noise prediction result of CCA8615 at Gangqiying point; d represents the departure noise prediction result of CCA8665 at Gangqiying point; e represents the departure noise prediction result of CUA2935 at Gangqiying point; f represents the departure noise prediction result of CUA5507 at Gangqiying point.

[0166] The prediction results of the second round of training model can effectively solve the problem of uneven prediction curve. In addition, the peak position is closer to the monitoring value.

[0167] Model Example Application

[0168] This section uses the B738 noise prediction model with the lowest mean square error as an example. Fig.21The figure shows a departure track with flight number JDL2618 and aircraft model B738. The track starts at 5:23:45 and ends at 5:27:22. This study predicts the instantaneous ground A-level distribution in a square area with a side length of 40 km around the track at 2s intervals.

[0169] The input data is shown in Table 6. Due to the large amount of data, the data is extracted at intervals of 10 seconds for display.

[0170] Table 6 Noise prediction input data

[0171]

[0172] (1) A sound level distribution prediction results (L A )

[0173] A sound level distribution prediction results are as follows Fig. 22 As the aircraft altitude increases, the impact range of aircraft noise on the ground will gradually decrease.

[0174] Will Fig. 22 The instantaneous A sound level distribution prediction results are superimposed to obtain the equivalent continuous A sound level distribution (L eq ),like Fig.23 As shown:

[0175] (2) Weighted equivalent continuous perceived noise level (L WECPN ) Prediction results

[0176] like Fig.24 The weighted equivalent continuous perceived noise level L is shown. WECPN The prediction results can be used to determine whether they exceed the standard based on the limit values ​​in the table.

[0177] (3) Annual average day and night equivalent sound level (YL DN ) Prediction results

[0178] Can be L eq Calculate L AE , by L AE Calculate L DN , by L DN Calculate YL DN Assuming that there are 5, 10, and 15 flights per hour, YL DN Distribution (such as Figure 25-27 As shown), and determine YL according to the limit values ​​in Table 2 DN Whether it exceeds the standard.

[0179] Therefore, the present invention adopts the above-mentioned airport single flight noise prediction method based on BP neural network, and the L AThe proportion of samples with an absolute value of the predicted relative deviation below 10% is 95.43%. The prediction accuracy can meet the engineering requirements; L WECPN and Y.L. DN The contour prediction results can be used to evaluate the noise indicators in the area around the airport; in future research, the monitoring quality of the remaining monitoring points will be optimized to increase the combination types of input variables and improve the accuracy and credibility of the model.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting airport single flight noise based on BP neural network, characterized in that: The following steps are involved: S1. Filter the original noise data of a single flight at the airport and perform data normalization processing; S2, establish a training set and a validation set with the screened original data, and construct a prediction model training set S; S3, training the training set S in step S2 to obtain a noise prediction model that meets the accuracy requirements; S4, performing secondary training on the noise prediction model obtained by training in step S3 to obtain a final noise prediction model; S5. Using the noise prediction model obtained in step S4, predict the noise impact of a specific aircraft model.

2. The method for predicting airport single flight noise based on BP neural network according to claim 1 is characterized in that: In step S1, the track data and noise monitoring data are correlated and preprocessed, and the data that cannot be used for model training due to low quality are cleaned up to obtain high-quality noise point monitoring data and aircraft track data.

3. The method for predicting airport single flight noise based on BP neural network according to claim 2 is characterized in that: In step S2, the filtered raw data is used to establish a training set and a validation set, and the prediction model training set S is specifically constructed as follows: Half of the filtered original data is used as the training set, and the other half is used as the validation set to construct the prediction model training set S, S = (x1, x2, ..., x n ); in the prediction model structure, n is the total number of training set samples, x is the input variable, y is the output variable, and the neurons are activated by the sigmoid activation function. The mapping relationship is:

4. The method for predicting airport single flight noise based on BP neural network according to claim 3 is characterized in that: Step S3 specifically includes the following steps: S31. First, a noise distribution prediction problem is set; S32, assigning a weight to each piece of sample data in the training set S in step S2, assuming that each piece of sample data is equally important; S33, assuming that k prediction models are trained, different hidden layer structures are set for screening, and MATLAB software and Neural Net Fitting toolbox are used for training respectively to obtain multiple prediction models, and the appropriate number of hidden layers is selected by observing the prediction model results; S34, setting neural grid parameters; Let TF1 be the transfer function from the input to the first hidden layer, TF2 be the transfer function from the first hidden layer to the second hidden layer, TF3 be the output layer transfer function, set TF1 and TF2 to logical sigmoid functions; set TF3 to linear activation function; S35, setting model training parameters, including training times, training targets and learning rates, and using a gradient descent algorithm to update model parameters; S36, neural network training; S37, performing data denormalization; S38. Calculate the root mean square error (MSE) of the validation set and select the model with the smaller MES value as the noise prediction model.

5. The method for predicting airport single flight noise based on BP neural network according to claim 4 is characterized in that: In step S4, the optimal model obtained by training step S3 is simulated, the prediction problem in step S2 is simulated and compared with the actual noise monitoring data, and the screened data is subjected to a second round of model training to obtain the final noise prediction model.

6. The method for predicting airport single flight noise based on BP neural network according to claim 5, characterized in that: The screened data is to remove bad pixels greatly affected by environmental noise, that is, to remove data whose monitoring values ​​are higher than the predicted values, and use the remaining data for the second round of model training.

Images