Airport noise prediction method based on back propagation neural network and support vector machine

Abstract

The invention provides an airport noise prediction method based on a back propagation neural network and a support vector machine, and belongs to the technical field of air traffic transportation noise, and the method comprises the steps: collecting airport noise historical data, obtaining flight historical data of airplanes with different powers, and obtaining a prediction input variable and a prediction output variable; according to the input and output variables and an elastic back propagation algorithm, in combination with a gradient time sequence dynamic rarefaction method and an adaptive structure adjustment method, establishing a back propagation neural network prediction model; constructing an airport noise kernel function, and establishing a support vector machine prediction model in combination with a cross validation method; predicting by using a back propagation neural network prediction model and a support vector machine prediction model, comparing and analyzing prediction results in combination with evaluation indexes, and selecting an optimal prediction result as an airport noise prediction result; the problems that an existing noise prediction algorithm depends on a large amount of historical data, the calculated amount is huge, and the noise prediction efficiency is low are solved.

Description

Technical Field

[0001] This invention belongs to the field of air traffic noise technology, and particularly relates to an airport noise prediction method based on backpropagation neural network and support vector machine. Background Technology

[0002] With the development of my country's civil aviation industry, noise and interference problems are becoming increasingly serious. Accurate and timely airport noise information is crucial for the successful application of the Air Traffic Management System (ATMS). It can help track users make better travel decisions, mitigate the impact of airport noise, reduce airport noise complaints, and improve air traffic operation efficiency. The prerequisite and key to achieving all of this is the ability to accurately predict airport noise, and the accuracy of the prediction directly determines the efficiency of air traffic operation.

[0003] Airport noise prediction models are mainly divided into time series models, nonparametric regression models, and neural network models. Early researchers used parametric models based on time series analysis to predict airport noise. Unlike other time series methods, this model does not require fixed initialization simulations. It treats the airport noise at a certain moment as a more general non-stationary random sequence, usually with 3 or 6 model parameters. Nonparametric regression is a nonparametric modeling method suitable for uncertain and nonlinear dynamic systems. It does not require prior knowledge, only sufficient historical data, to find "nearest neighbors" similar to the current point in the historical data, and uses those "nearest neighbors" to predict the value at the next moment. Therefore, when special events occur, the prediction results are more accurate than parametric modeling.

[0004] Existing technologies summarize four challenging directions in the field of airport noise prediction and point out that researchers need to combine statistical data and artificial intelligence methods in a complementary manner, and provide a unified public dataset to solve the challenges in the field of airport noise prediction, thereby improving the performance of airport noise prediction. Among the existing technologies, neural network models are used to train relevant noise data, and the neural network models are adjusted during the training process. Finally, the adjusted neural network models are used to predict noise, thereby improving the performance and accuracy of noise prediction.

[0005] In the field of noise prediction, the most popular method currently is to predict noise using nonparametric models based on machine learning. This method has high prediction accuracy based on a large amount of continuous data, but it requires complex parameter estimation, and the calculated parameters are not portable. In practice, data omissions often occur due to various reasons, leading to a decrease in model accuracy. Current noise prediction algorithms still rely heavily on historical data, and building models based on large amounts of historical data results in huge computational loads, leading to low noise prediction efficiency. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides an airport noise prediction method based on backpropagation neural networks and support vector machines, which solves the problems of existing noise prediction algorithms relying heavily on historical data, having a huge computational load, and low noise prediction efficiency.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: an airport noise prediction method based on backpropagation neural network and support vector machine, comprising the following steps: S1. Collect historical airport noise data, and obtain historical flight data of aircraft with different power based on the historical flight database. Perform principal component analysis on the historical flight data to obtain the predicted input variables, and obtain the predicted output variables based on the historical airport noise data. S2. Based on the predicted input variables, predicted output variables, and elastic backpropagation algorithm, and combined with gradient time-series dynamic sparsification method and adaptive structure adjustment method, establish a backpropagation neural network prediction model. S3. Based on the physical characteristics of fused noise and prior knowledge of signal processing, an airport noise kernel function is constructed, and a support vector machine prediction model is established according to the predicted input variables and predicted output variables, combined with the cross-validation method. S4. Input the parameters of aircraft takeoff and landing at the current airport into the backpropagation neural network prediction model and the support vector machine prediction model respectively, and obtain the backpropagation neural network prediction results and the support vector machine prediction results. Combine the preset evaluation index, compare the backpropagation neural network prediction results and the support vector machine prediction results, and select the optimal prediction result as the airport noise prediction result.

[0008] The beneficial effects of this invention are as follows: By using an elastic backpropagation neural network and a support vector machine based on the airport noise kernel function, this invention can accurately predict airport noise, realize noise prediction without relying on historical data, reduce the amount of computation for noise prediction, and improve the efficiency, accuracy and precision of noise prediction. It can provide strong support for subsequent intelligent monitoring of airport noise, scheduling of aircraft flights and control of flight paths.

[0009] Further, S1 includes the following steps: S101. Use the airport noise sensor monitoring information system to collect historical airport noise data, preprocess the historical airport noise data using a normalization method, and obtain historical flight data of aircraft with different power based on the historical flight database. S102. Based on the flight history data of aircraft with different power, principal component analysis is used to extract the aircraft's flight altitude and engine power as prediction input variables, and based on the preprocessed historical airport noise data, airport noise is used as the prediction output variable.

[0010] The beneficial effects of the above-mentioned further scheme are as follows: The present invention uses principal component analysis to extract principal components from the flight history data of aircraft with different power, and uses the principal components as the prediction input variables and the normalized airport noise history data as the prediction output variables, thereby improving the efficiency and accuracy of the prediction method.

[0011] Furthermore, S2 specifically includes: Based on the backpropagation neural network model that includes an input layer, hidden layer and output layer, the backpropagation neural network model is trained according to the predicted input variables, predicted output variables and elastic backpropagation algorithm to obtain the trained backpropagation neural network model. The computation path of the trained backpropagation neural network model is adjusted by gradient temporal dynamic sparsification, and the number of nodes and hidden layers in the trained backpropagation neural network model is adjusted by adaptive structure adjustment method to obtain the backpropagation neural network prediction model.

[0012] Furthermore, the gradient temporal dynamic sparsification method specifically includes: By utilizing a dynamic sparse gating mechanism and taking advantage of gradient temporal features, redundant computational paths are dynamically closed, while high-contribution neuron activations are preserved. The gating coefficient expression for the dynamic sparse gating mechanism is as follows: ; ; in, Indicates the first l Layer i The gating coefficient of each neuron, express S Type function, The window size represents the time step or the number of training iterations. Indicates the first l Layer i The original activation value of each neuron. Represents the loss function. Indicates dynamic threshold. Indicates the first l The mean of the gradients of all neurons in the layer. This represents the adjustment coefficient.

[0013] Furthermore, the structural adjustment rule expression of the adaptive structural adjustment method is as follows: ; ; in, Indicates the first l Decision score for hierarchical restructuring Indicates the first l The number of neurons or channels in a layer Indicates the first l Layer i The importance score of each neuron The dynamic threshold representing the importance score. Indicates the learning rate. Represents the median function. Indicates the first l The set of importance scores for all neurons in a layer.

[0014] The beneficial effects of the above-mentioned further solutions are as follows: This invention constructs a backpropagation neural network prediction model through the elastic backpropagation algorithm, which improves the convergence speed of the neural network prediction model, reduces the complexity of the algorithm, and reduces the memory consumed when performing noise prediction; it uses the gradient temporal dynamic sparsification method to adjust the calculation path, and retains the activation of high-contribution neurons through the dynamic sparse gating mechanism, thereby accelerating the calculation speed of the model; it adopts an adaptive structure adjustment method to determine the optimal number of hidden layer nodes, making the model simpler and easier to use for prediction; and by using the elastic backpropagation algorithm, it achieves the function of not needing to manually set the global learning rate, reducing the difficulty of adjusting parameters.

[0015] Furthermore, step S3 includes the following steps: S301. Based on the physical characteristics of fused noise, a frequency band kernel function is designed, and an adaptive weighting mechanism is introduced to dynamically adjust the weighting coefficients to obtain the voiceprint decoupling kernel. S302. Introduce prior knowledge of airport layout to construct a runway distance attenuation term; S303. Construct meteorological correction terms based on wind speed, temperature, and humidity; S304. Construct the airport noise kernel function based on the acoustic decoupling kernel, runway distance attenuation term, and meteorological correction term; S305. Based on the predicted input variables and predicted output variables, perform nonlinear fitting in an infinite-dimensional feature space. Using cross-validation, take the penalty coefficient with the minimum mean square error and the minimum model complexity as the support vector machine model parameters to establish a support vector machine prediction model.

[0016] Furthermore, the airport noise kernel function is expressed as follows: ; in, Represents the airport noise kernel function. Indicates the first j Airport noise, This represents the runway distance decay term. It represents the Hadamah accumulation. This indicates meteorological influencing factors.

[0017] The beneficial effects of the above-mentioned further solutions are as follows: This invention decomposes airport noise through acoustic features, designs frequency band kernel functions, introduces an adaptive weighting mechanism to dynamically adjust the weighting coefficients, introduces prior knowledge of airport layout, constructs attenuation terms, and finally obtains an airport noise kernel function that covers acoustic decoupling kernel, runway distance attenuation and weather correction, thereby improving the prediction ability and accuracy of support vector machine. By utilizing the airport noise kernel function, the original sample space is mapped to an infinite-dimensional feature space. Through cross-validation, the support vector product prediction model is trained and validated to determine the optimal values ​​of the support vector product parameters. This improves the nonlinear fitting ability of the support vector product prediction model, achieves global optimum, increases convergence speed, and reduces training time.

[0018] Furthermore, step S4 includes the following steps: S401. Input the parameters of aircraft takeoff and landing at the current airport into the backpropagation neural network prediction model and the support vector machine prediction model respectively, and obtain the backpropagation neural network prediction results and the support vector machine prediction results. S402. Based on the prediction results of the backpropagation neural network and the support vector machine, calculate the mean absolute percentage error and the mean square error respectively to obtain the first mean absolute percentage error, the second mean absolute percentage error, the first mean square error and the second mean square error. S403. Based on the first mean absolute percentage error, the second mean absolute percentage error, the first mean square error, and the second mean square error, combined with the training sample time and convergence speed of the backpropagation neural network prediction model and the support vector machine prediction model, the prediction results are analyzed, and the optimal prediction result is selected as the airport noise prediction result.

[0019] The beneficial effects of the above-mentioned further solutions are as follows: By introducing mean absolute percentage error and mean square error, and combining the training sample time and convergence speed of the backpropagation neural network prediction model and the support vector machine prediction model, the present invention accurately compares the prediction effects, and comprehensively judges the optimal noise prediction method based on the actual situation, thereby improving the versatility of the object of the present invention and the accuracy of the prediction results. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention.

[0021] Figure 2 This is a diagram of the BP neural network structure in this embodiment.

[0022] Figure 3 This is a schematic diagram illustrating the principle of the support vector machine in this embodiment. Detailed Implementation

[0023] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0024] Before describing this embodiment, the following terms will be explained: BP neural network: Backpropagation neural network; SVM model: Support Vector Product model; Sigmoid: S-shaped growth curve; MATLAB: A software for advanced technical computing and visualization, widely used in engineering and scientific fields; MAPE: Mean Absolute Percentage Error; MSE: Mean Squared Error.

[0025] Example like Figure 1 As shown, this invention provides an airport noise prediction method based on backpropagation neural network and support vector machine, the implementation method of which is as follows: S1. Collect historical airport noise data, and based on the historical flight database, obtain historical flight data of aircraft with different power levels. Perform principal component analysis on the historical aircraft flight data to obtain the predicted input variables, and obtain the predicted output variables based on the historical airport noise data. The specific steps are as follows: S101. Use the airport noise sensor monitoring information system to collect historical airport noise data, and use the normalization method to preprocess the historical airport noise data, and obtain historical flight data of aircraft with different power based on the historical flight database. S102. Based on the flight history data of aircraft with different power, principal component analysis is used to extract the aircraft's flight altitude and engine power as prediction input variables, and based on the preprocessed historical airport noise data, airport noise is used as the prediction output variable.

[0026] In this embodiment, an airport noise sensor monitoring information system is used to collect historical airport noise data, and a normalization method is used for preprocessing. The preprocessing method is as follows: The minimum (min) and maximum (max) values ​​in a given sample of historical airport noise data are calculated separately. The min-max standardization method is then used to normalize the data, mapping the normalized airport noise data to an interval. In the middle, that is, based on the airport noise data set , This represents the total number of airport noise data samples. Indicates the first t Given several airport noise data points, we calculate the minimum and maximum values ​​(min and max) of the airport noise data, and then perform normalization calculations. The calculation expressions are shown below: ; in, Indicates the normalized first t Airport noise data is collected from individual airports and integrated with normalized airport noise data to obtain preprocessed historical airport noise data. Based on the historical flight database, flight history data of aircraft with different power were obtained, and noise monitoring values ​​of A320 aircraft at Tianfu Airport at different flight altitudes and with different engine power were selected. Based on historical flight data of aircraft with different power ratings, principal component analysis was used to extract two principal components: aircraft altitude and aircraft engine power. These principal components were then used as the input variables for prediction. Based on preprocessed historical airport noise data, airport noise was used as the output variable for prediction.

[0027] S2. Based on the predicted input variables, predicted output variables, and elastic backpropagation algorithm, and combined with gradient temporal dynamic sparsification and adaptive structure adjustment methods, a backpropagation neural network prediction model is established, specifically as follows: Based on the backpropagation neural network model that includes an input layer, hidden layer and output layer, the backpropagation neural network model is trained according to the predicted input variables, predicted output variables and elastic backpropagation algorithm to obtain the trained backpropagation neural network model. The computation path of the trained backpropagation neural network model is adjusted by gradient temporal dynamic sparsification, and the number of nodes and hidden layers in the trained backpropagation neural network model is adjusted by adaptive structure adjustment method to obtain the backpropagation neural network prediction model.

[0028] In this embodiment, as Figure 2 The backpropagation neural network structure shown is used to construct a backpropagation neural network prediction model, where... Indicates input variables, This represents the total number of input variables, and also the number of input layers. Indicates the number of hidden layers. This indicates the number of output layers, and introduces node errors into the output layers. , Indicates the first j Hidden layers and the first k The weights of the output layer; In selecting noise monitoring indicators, this embodiment referenced relevant standards to ensure that the selected indicators comprehensively reflect the characteristics and impact of airport noise. During noise calculation and evaluation, the equivalent continuous sound level (…) was used. L eq ), maximum sound level ( L max ), minimum sound level ( L min Indicators such as ( ), as well as the exposure sound level of aviation noise events and the calculation of individual noise events, are all important contents; specifically, L eq (Equivalent continuous sound level) is used to represent the average noise level over a period of time and can reflect the total energy of the noise; L max (Maximum sound level) and L min (Minimum sound level) is used to assess the peak and background levels of noise, which helps to identify sudden noise events and continuous noise. L 10 , L 50 and L 90 Cumulative percentage sound level provides the distribution of noise at different time percentages, which helps to assess the persistence and variability of noise; such as Figure 2 As shown, Y1, Y2, ..., Y M They represent the equivalent continuous sound level ( L eq ), maximum sound level ( L max ), minimum sound level ( L min Indicators such as ) L AE (Exposure sound level): Exposure sound level refers to the A-weighted sound pressure level of a noise event within a specified measurement period or for a single noise event, which is equivalent to the sound energy of the event for 1 second. It is used to assess the total energy of a noise event and is an important indicator for measuring the degree of noise exposure. L eq (Equivalent continuous sound level): The equivalent continuous sound level refers to the time average of the sound pressure level over a specified time interval. It converts the sound pressure level at different times into an equivalent continuous sound pressure level through energy averaging, and is used to represent the average noise level over a period of time. L max (Maximum sound level): The maximum sound level refers to the maximum A-weighted sound level measured during a specified measurement period or for a specific noise event. It reflects the highest noise level in the noise event and is very important for assessing the peak impact of noise. L min (Minimum sound level): The minimum sound level refers to the minimum A-weighted sound level measured during a specified measurement period or for a specific noise event. It reflects the lowest noise level in the noise event and is of great significance for assessing the lower limit of noise impact. L 10 (10% Exceeding Sound Level): The 10% exceeding sound level refers to the A-weighted sound pressure level that is exceeded for 10% of the time within the considered time T. It indicates that the noise level is higher than this value for 10% of the time during the measurement period, and is used to assess the distribution of higher noise levels. L 50 (Median sound level): The median sound level is the A-weighted sound pressure level that is exceeded for 50% of the time within the time period T under consideration. It means that the noise level is higher than this value for 50% of the time during the measurement period, and it is the median of the noise level. L 90 (90% Exceeding Sound Level): The 90% exceeding sound level refers to the A-weighted sound pressure level that is exceeded for 90% of the time within the considered time T. It indicates that the noise level is higher than this value for 90% of the time during the measurement period and is used to assess the low-level distribution of noise.

[0029] In this embodiment, the training process of a BP neural network typically includes two stages: forward propagation and backward propagation. In the forward propagation stage, the input signal is processed from the input layer through the hidden layer and then transmitted to the output layer. If the output of the output layer does not match the expected output, the error signal will propagate backward through the network to adjust the connection weights between the layers in order to reduce the error. Therefore, the training speed and accuracy of the BP neural network model are related to the number of nodes and the number of hidden layers. By using principal component analysis to obtain two principal components as input variables for a backpropagation neural network and airport noise level as the output variable, the backpropagation neural network model is trained using MATLAB in conjunction with the elastic backpropagation algorithm. The gradient time-series dynamic sparsity method is used to improve the model's computation speed by dynamically adjusting the computation path while maintaining model performance. Furthermore, the number of nodes and hidden layers are adjusted using an adaptive structure adjustment method based on dynamic gradient sparsity and path importance to obtain the optimal BP neural network prediction model. The gradient temporal dynamic sparsification method specifically involves: using a dynamic sparsity gating mechanism, through gradient temporal features, dynamically closing redundant computation paths and retaining the activation of high-contribution neurons; The expression for the gating coefficient is as follows: ; ; in, Indicates the first l Layer i The gating coefficient of each neuron, This represents the Sigmoid function, which compresses the input to [0,1]. The window size represents the time step or the number of training iterations, used to calculate the temporal average of the gradient. Indicates the first l Layer i The original activation value of each neuron. The loss function can be represented by cross-entropy or mean squared error. This represents a dynamic threshold used to distinguish between neurons with high or low contribution levels. Indicates the first l The mean of the gradients of all neurons in the layer. Indicates the adjustment coefficient; The forward propagation correction is performed using the gating coefficient, as shown in the following expression: ; Among them, if Skip the current neuron's computation. The corrected output is the gated weighted result of the original activations, adjusted by gating the original activation values ​​using gating coefficients. Weighted, low-contribution neurons were suppressed ( (Small) noise, while amplifying high-contribution neurons ( Selective correction of large signals allows the network to focus more on important feature representations; By using gating mechanisms to distinguish the importance of neurons, high-contribution activations are preserved and enhanced while low-contribution activations are suppressed, thereby improving the efficiency and generalization of the model and achieving dynamic sparsity.

[0030] In this embodiment, the adaptive structure adjustment method based on dynamic gradient sparsity and path importance achieves intelligent structure adjustment of the model by quantifying the criticality of the gradient propagation path. Define a path importance index to measure the importance of a path. The specific expression is as follows: ; in, Representing a path p Importance index Indicates the first l Layer activation value The derivative, This represents the gradient variance, used to calculate the variability of the gradient of the same layer parameters across different samples or time steps. Indicates the first l Layer parameters; By calculating node importance scores, structural adjustment rules are constructed to achieve intelligent structural adjustment of the model; The expression for the structural adjustment rule is as follows: ; ; in, Indicates the first l The decision score for layer structure adjustment is used to determine whether to prune or expand the layer. Indicates the first l The number of neurons or channels in a layer Indicates the first l Layer i The importance score of each neuron The dynamic threshold representing the importance score. Indicates the learning rate. Represents the median function. Indicates the first l The set of importance scores for all neurons in a layer is used to adjust the number of nodes and hidden layers in the trained backpropagation neural network model.

[0031] In this embodiment, the elastic backpropagation algorithm independently adapts to gradient changes by adjusting the update step size of each parameter, avoiding the need to manually set the global learning rate. In elastic backpropagation, each parameter has an independent learning step size, which is adaptively adjusted according to gradient changes, eliminating the need to manually set the global learning rate and reducing the difficulty of parameter tuning. Furthermore, the weight update rules are different. The weight update is based on the step size and gradient sign, and is not sensitive to the gradient magnitude. The update step size is more stable, and the adaptive step size adjustment mechanism accelerates convergence, especially when the gradient changes drastically. Because there are fewer hyperparameters, parameter tuning is relatively simple. The training error (loss function) is a quadratic function of the input weights and biases. Taking the partial derivatives with respect to the weights and biases yields the gradient vector. The direction of the gradient vector is where the training error increases the fastest, while the opposite direction shows the fastest decrease in gradient. It is easier to find the minimum value of the training error function (loss function) in this direction. Experiments in this embodiment show that, under aircraft takeoff conditions, the optimal BP neural network prediction model contains one hidden layer. The first hidden layer of the model has 1 node, and the model structure is 2-1-1 (2 input nodes, 1 hidden node). With one output node, the average error between the true and predicted values ​​of the model is 0.070, and the correlation is 0.740. In the case of aircraft landing, the model contains one hidden layer with two nodes in the first hidden layer. The model structure is 2-1-2 (2 input nodes, 1 hidden node, 2 output nodes). The average error between the true and predicted values ​​of the BP neural network test samples is 0.046, and the correlation is 0.769. The backpropagation neural network prediction model is obtained. It can be seen that the BP neural network still has high prediction accuracy under different conditions of aircraft takeoff and landing.

[0032] S3. Based on the fusion of noise physical characteristics and prior knowledge of signal processing, an airport noise kernel function is constructed. Then, according to the predicted input and output variables, and combined with cross-validation, a support vector machine prediction model is established, specifically: S301. Based on the physical characteristics of fused noise, a frequency band kernel function is designed, and an adaptive weighting mechanism is introduced to dynamically adjust the weighting coefficients to obtain the voiceprint decoupling kernel. S302. Introduce prior knowledge of airport layout to construct a runway distance attenuation term; S303. Construct meteorological correction terms based on wind speed, temperature, and humidity; S304. Construct the airport noise kernel function based on the acoustic decoupling kernel, runway distance attenuation term, and meteorological correction term; S305. Based on the predicted input variables and predicted output variables, perform nonlinear fitting in an infinite-dimensional feature space. Using cross-validation, take the penalty coefficient with the minimum mean square error and the minimum model complexity as the support vector machine model parameters to establish a support vector machine prediction model.

[0033] In this embodiment, an airport noise kernel function is constructed by fusing the physical characteristics of noise with prior knowledge of signal processing; Based on acoustic characteristics, airport noise is decomposed into: steady-state components (engine fundamental frequency), transient components (takeoff and landing impact noise), and broadband noise (aerodynamic noise). The frequency band kernel function is designed, and its specific expression is as follows: ; in, Represents the frequency band kernel function. and Both represent vectors. Indicates the first j Airport noise, Indicates the first j Weighting coefficients for each frequency band Indicates the first j Sub-kernel functions for each frequency band, which can be linear kernels or Gaussian kernels. This indicates wavelet packet energy feature extraction, which represents transient components. This indicates cepstral coefficient feature extraction, representing steady-state components. This represents the feature extraction of the Mel filter bank, which is broadband noise; An adaptive weighting mechanism is introduced to dynamically adjust the weighting coefficients, resulting in a voiceprint decoupling kernel. ; The weighting coefficients The expression for dynamic adjustment is shown below: ; in, Subkernel function , Indicates the frequency band index. Indicates the first j Signal-to-noise ratio of each frequency band Indicates the total signal-to-noise ratio; Furthermore, prior knowledge of airport layout is introduced to construct a runway distance decay term, the expression of which is shown below: ; in, This represents the runway distance decay term. Indicates the sound source x spatial coordinates, Indicates the observation point y spatial coordinates, Indicates the spatial attenuation coefficient. The adjustment factor representing the runway distance. This indicates the vertical distance from the sound source to the runway; Meteorological correction terms are constructed based on wind speed, temperature, and humidity. ; ; in, This indicates the wind speed-wind direction term. This represents the reference speed of sound, i.e., the current speed of sound. This represents the wind speed vector, which is a vector containing the magnitude and direction of the wind speed. Indicates from the sound source point x To the receiving pointy The direction vector of the (predicted point) Represents vector d The modulus, i.e., the propagation distance between the sound source and the receiver. Represents the temperature gradient term. This represents the empirical proportionality coefficient for the temperature gradient term. Indicates the amount of temperature change. Indicates humidity. The empirical proportionality coefficient representing the humidity term. Indicates relative humidity; Based on the acoustic decoupling kernel, runway distance attenuation term, and weather correction term, the airport noise kernel function is constructed, and its expression is shown below: ; in, It represents the Hadamah accumulation. The meteorological influencing factor is expressed as follows: ; in, This represents the Sigmoid function, which maps wind speed to [0,1]. Indicates wind speed. Represents the hyperbolic tangent function. This indicates the humidity level.

[0034] In this embodiment, as Figure 3 Based on the principle shown, a support vector machine model based on the airport noise kernel function is constructed. Specifically, linearly inseparable points in the low-dimensional space represented by two-dimensional coordinate axes X1 and X2 are mapped to linearly separable points in the high-dimensional feature space represented by three-dimensional coordinate axes Z1, Z2, and Z3. The support vector machine model uses the airport noise kernel function to decompose airport noise and achieve optimal segmentation of airport noise, minimizing the "total deviation" of all samples from the hyperplane, which is the optimal solution of the support vector machine model. This invention selects the airport noise kernel function and integrates the physical characteristics of noise with prior knowledge of signal processing to realize a support vector machine model for airport noise prediction. Based on the predicted input and output variables, a nonlinear fitting is performed in an infinite-dimensional feature space, and the penalty coefficient that minimizes the mean squared error is selected using cross-validation. and minimum model complexity As parameters of the Support Vector Machine (SVM) model, the penalty coefficient that minimizes the mean squared error of the SVM model during aircraft takeoff. and minimum model complexity The values ​​are 0.001 and 1 respectively, the model's mean squared error is 0.070, and the correlation is 0.739; under the aircraft landing condition, the SVM model has the smallest penalty coefficient for mean squared error. and minimum model complexity The values ​​are 0.1 and 1 respectively. The mean squared error of the model prediction is 0.047, and the correlation is 0.759. A support vector machine prediction model is established.

[0035] S4. Input the current airport's aircraft takeoff and landing parameters into the backpropagation neural network prediction model and the support vector machine prediction model, respectively, to obtain the backpropagation neural network prediction results and the support vector machine prediction results. Combine these with preset evaluation indicators, compare the backpropagation neural network prediction results and the support vector machine prediction results, and select the optimal prediction result as the airport noise prediction result. Specifically: S401. Input the parameters of aircraft takeoff and landing at the current airport into the backpropagation neural network prediction model and the support vector machine prediction model respectively, and obtain the backpropagation neural network prediction results and the support vector machine prediction results. S402. Based on the prediction results of the backpropagation neural network and the support vector machine, calculate the mean absolute percentage error and the mean square error respectively to obtain the first mean absolute percentage error, the second mean absolute percentage error, the first mean square error and the second mean square error. S403. Based on the first mean absolute percentage error, the second mean absolute percentage error, the first mean square error, and the second mean square error, combined with the training sample time and convergence speed of the backpropagation neural network prediction model and the support vector machine prediction model, the prediction results are analyzed, and the optimal prediction result is selected as the airport noise prediction result.

[0036] In this embodiment, the real-time parameters of aircraft takeoff and landing at the current airport are input into the backpropagation neural network prediction model and the support vector machine prediction model, respectively, to obtain the backpropagation neural network prediction results and the support vector machine prediction results. Based on the prediction results of the backpropagation neural network and the support vector machine, the mean absolute percentage error and the mean square error are calculated respectively, resulting in the first mean absolute percentage error, the second mean absolute percentage error, the first mean square error, and the second mean square error. Based on the first mean absolute percentage error, the second mean absolute percentage error, the first mean square error, and the second mean square error, combined with the training sample time and convergence speed of the backpropagation neural network prediction model and the support vector machine prediction model, the prediction results are analyzed, as shown in Table 1. The optimal prediction result is selected as the airport noise prediction result.

[0037] Table 1

[0038] As shown in the table, under takeoff conditions, the prediction accuracy of the support vector machine (SVM) prediction model is slightly higher than that of the backpropagation (BP) neural network (BNN) prediction model, while the opposite is true for landing conditions. Compared to the BNN prediction model, the SVM prediction model is faster in computation, has an advantage for large datasets, and does not get stuck in local convergence during computation, thus obtaining the global optimal solution. Considering all aspects, the overall performance of the SVM prediction model is better. Therefore, the SVM prediction results were selected for this airport noise prediction.

Claims

1. An airport noise prediction method based on backpropagation neural network and support vector machine, characterized in that, Includes the following steps: S1. Collect historical airport noise data, and obtain historical flight data of aircraft with different power based on the historical flight database. Perform principal component analysis on the historical flight data to obtain the predicted input variables, and obtain the predicted output variables based on the historical airport noise data. S2. Based on the predicted input variables, predicted output variables, and elastic backpropagation algorithm, and combined with gradient time-series dynamic sparsification method and adaptive structure adjustment method, establish a backpropagation neural network prediction model. S3. Based on the physical characteristics of fused noise and prior knowledge of signal processing, an airport noise kernel function is constructed, and a support vector machine prediction model is established according to the predicted input variables and predicted output variables, combined with the cross-validation method. S4. Input the parameters of aircraft takeoff and landing at the current airport into the backpropagation neural network prediction model and the support vector machine prediction model respectively, and obtain the backpropagation neural network prediction results and the support vector machine prediction results. Combine the preset evaluation index, compare the backpropagation neural network prediction results and the support vector machine prediction results, and select the optimal prediction result as the airport noise prediction result.

2. The airport noise prediction method based on backpropagation neural network and support vector machine according to claim 1, characterized in that, S1 includes the following steps: S101. Use the airport noise sensor monitoring information system to collect historical airport noise data, preprocess the historical airport noise data using a normalization method, and obtain historical flight data of aircraft with different power based on the historical flight database. S102. Based on the flight history data of aircraft with different power, principal component analysis is used to extract the aircraft's flight altitude and engine power as prediction input variables, and based on the preprocessed historical airport noise data, airport noise is used as the prediction output variable.

3. The airport noise prediction method based on backpropagation neural network and support vector machine according to claim 1, characterized in that, S2 specifically includes: Based on the backpropagation neural network model that includes an input layer, hidden layer and output layer, the backpropagation neural network model is trained according to the predicted input variables, predicted output variables and elastic backpropagation algorithm to obtain the trained backpropagation neural network model. The computation path of the trained backpropagation neural network model is adjusted by gradient temporal dynamic sparsification, and the number of nodes and hidden layers in the trained backpropagation neural network model is adjusted by adaptive structure adjustment method to obtain the backpropagation neural network prediction model.

4. The airport noise prediction method based on backpropagation neural network and support vector machine according to claim 3, characterized in that, The gradient temporal dynamic sparsification method is specifically as follows: By utilizing a dynamic sparse gating mechanism and taking advantage of gradient temporal features, redundant computational paths are dynamically closed, while high-contribution neuron activations are preserved. The gating coefficient expression for the dynamic sparse gating mechanism is as follows: in, Indicates the first l Layer i The gating coefficient of each neuron, express S Type function, The window size represents the time step or the number of training iterations. Indicates the first l Layer i The original activation value of each neuron. Represents the loss function. Indicates a dynamic threshold. Indicates the first l The mean of the gradients of all neurons in the layer. This represents the adjustment coefficient.

5. The airport noise prediction method based on backpropagation neural network and support vector machine according to claim 3, characterized in that, The structural adjustment rule expression of the adaptive structural adjustment method is as follows: in, Indicates the first l Decision score for hierarchical restructuring Indicates the first l The number of neurons or channels in a layer Indicates the first l Layer i The importance score of each neuron The dynamic threshold representing the importance score. Indicates the learning rate. Represents the median function. Indicates the first l The set of importance scores for all neurons in a layer.

6. The airport noise prediction method based on backpropagation neural network and support vector machine according to claim 1, characterized in that, S3 includes the following steps: S301. Based on the physical characteristics of fused noise, a frequency band kernel function is designed, and an adaptive weighting mechanism is introduced to dynamically adjust the weighting coefficients to obtain the voiceprint decoupling kernel. S302. Introduce prior knowledge of airport layout to construct a runway distance attenuation term; S303. Construct meteorological correction terms based on wind speed, temperature, and humidity; S304. Construct the airport noise kernel function based on the acoustic decoupling kernel, runway distance attenuation term, and meteorological correction term; S305. Based on the predicted input variables and predicted output variables, perform nonlinear fitting in an infinite-dimensional feature space. Using cross-validation, take the penalty coefficient with the minimum mean square error and the minimum model complexity as the support vector machine model parameters to establish a support vector machine prediction model.

7. The airport noise prediction method based on backpropagation neural network and support vector machine according to claim 6, characterized in that, The airport noise kernel function is expressed as follows: in, Represents the airport noise kernel function. Indicates the first j Airport noise, This represents the runway distance decay term. It represents the Hadamah accumulation. This indicates meteorological influencing factors.

8. The airport noise prediction method based on backpropagation neural network and support vector machine according to claim 1, characterized in that, S4 includes the following steps: S401. Input the parameters of aircraft takeoff and landing at the current airport into the backpropagation neural network prediction model and the support vector machine prediction model respectively, and obtain the backpropagation neural network prediction results and the support vector machine prediction results. S402. Based on the prediction results of the backpropagation neural network and the support vector machine, calculate the mean absolute percentage error and the mean square error respectively to obtain the first mean absolute percentage error, the second mean absolute percentage error, the first mean square error and the second mean square error. S403. Based on the first mean absolute percentage error, the second mean absolute percentage error, the first mean square error, and the second mean square error, combined with the training sample time and convergence speed of the backpropagation neural network prediction model and the support vector machine prediction model, the prediction results are analyzed, and the optimal prediction result is selected as the airport noise prediction result.

Images