A Meteorological Signal Spectral Moment Estimation Method Based on Two-Ended Incremental Extreme Learning Machine

Abstract

This invention discloses a meteorological signal spectral moment estimation method based on B-ELM. The method includes the following steps: acquiring meteorological pulse data; converting the meteorological pulse data into a meteorological power spectrum using the average corrected periodogram method; using the meteorological power spectrum as the input of the network training samples and the spectral moments as the output of the network training samples to construct the training samples; randomly setting the parameters of the hidden nodes, determining the number of input layer nodes and the number of output layer nodes, and selecting the activation function; constructing a prediction model based on the B-ELM algorithm to obtain the optimal meteorological signal radial velocity and spectral width prediction model; substituting the training samples, the set number of input layer nodes, and the number of output layer nodes into the obtained optimal meteorological signal radial velocity and spectral width prediction model to obtain the radial velocity and spectral width of the meteorological signal. This method has a fast convergence speed, small error, improves the accuracy of meteorological signal spectral moment estimation, and has low computational complexity, which is beneficial for engineering implementation.

Description

Technical Field

[0001] This invention belongs to the field of meteorological signal spectral moment estimation, specifically involving a meteorological signal spectral moment estimation method based on a dual-incremental extreme learning machine. Background Technology

[0002] With the further development of meteorological radar, it has become an important tool for atmospheric remote sensing and meteorological detection, playing a vital role in severe weather warnings, short-term weather monitoring and forecasting, public services, and disaster prevention and mitigation. Meteorological target information in meteorological echoes plays a crucial role in meteorological radar systems. The spectral moment information of meteorological signals is an important parameter for meteorological target information, a key basis for determining the type of meteorological target, and also a crucial basis for determining whether a warning is needed based on the weather conditions. The spectral moment mainly includes average power, average Doppler velocity, and Doppler spectral width. Average power corresponds to the reflectivity of the meteorological target, reflecting the liquid water content and rainfall rate in the area; average Doppler velocity is the sum of the average velocities of each scattering particle; and Doppler spectral width represents the degree of dispersion of the scattering particle velocities, indicating the intensity of the meteorological target's motion.

[0003] Traditional algorithms such as the pulse pair method and the fast Fourier transform method suffer from severe performance degradation under conditions of few pulses and low signal-to-noise ratio, failing to meet the high accuracy requirements of meteorological radar systems for spectral moments. In recent years, spectral moment estimation methods based on parameterized models and sparse reconstruction have been applied and researched, significantly improving the accuracy of spectral moment estimation. However, these two algorithms suffer from high computational complexity and poor timeliness. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes a meteorological signal spectral moment estimation method based on a dual-incremental extreme learning machine.

[0005] To achieve the above-mentioned technical objectives, the technical solution of the present invention is as follows:

[0006] A meteorological signal spectral moment estimation method based on dual-incremental extreme learning machine (B-ELM) is characterized by the following steps:

[0007] S1. Obtain meteorological pulse data, convert the meteorological pulse data into meteorological power spectrum using the average corrected periodogram method, use the meteorological power spectrum as the input of network training samples, and use the spectral moments as the output of network training samples to construct training samples.

[0008] S2. Randomly set the parameters of the hidden nodes, determine the number of nodes in the input layer and the number of nodes in the output layer, and select the activation function;

[0009] S3. Construct a prediction model based on the B-ELM algorithm to obtain the optimal prediction model for the radial velocity and spectral width of meteorological signals;

[0010] S4. Substitute the training samples, the set number of input layer nodes, and the number of output layer nodes into the obtained optimal meteorological signal radial velocity and spectral width prediction model to obtain the radial velocity and spectral width of the meteorological signal.

[0011] Furthermore, the steps for estimating the meteorological power spectrum using the average corrected periodogram method in S1 are as follows:

[0012] Assuming the weather radar is for the first Spectral moment estimation is performed on each distance cell, and spectral moments are selected on both sides thereof. 1 distance unit is used as a training sample;

[0013] S1.1, Meteorological pulse data Divided into Segments, each segment is [length missing] , No. The signal segments are:

[0014]

[0015] in, It is the first The starting point of the segment, when When data overlaps and is segmented, take... At that time, 50% data overlap occurred;

[0016] S1.2 Introducing the Hamming window function The correction periodicity chart corresponding to each data segment is calculated using the following formula, the first... The correction periodicity diagram for the segment is as follows:

[0017]

[0018] In the formula, This is called the normalization factor. ;

[0019] S1.3. The meteorological power spectrum can be estimated by averaging using the modified periodogram method:

[0020] .

[0021] Furthermore, the method for constructing the training samples in S1 is as follows:

[0022] ,

[0023] in, For input, For output, It is the number of input layer nodes. The number of output layer nodes is given. For meteorological signals, the samples constructed in this method are as follows:

[0024]

[0025] in, For the first Power spectrum of each training sample This represents the number of training samples.

[0026] Furthermore, the steps for constructing the optimal meteorological signal radial velocity and spectral width prediction model in S3 are as follows:

[0027] Let the maximum number of hidden layer nodes and the expected error be respectively... and When the output is radial velocity and spectral width, set ;

[0028] Initialization phase: setting , The number of hidden layer nodes in the network is denoted as , and the network residual is denoted as . ,in Indicates the expected output;

[0029] Learning stage: When and hour:

[0030] 1) Add a hidden layer node: ;

[0031] 2) When When the number is odd: Randomly select parameters for the newly added hidden layer nodes. ;when When the number is even: Calculate the error feedback matrix Randomly set the parameters of the hidden layer nodes ;

[0032] 3) Calculate the weight output matrix ;

[0033] 4) Calculate the weights between the newly added hidden nodes and the output layer. ;

[0034] 5) Calculate the network residuals of the B-ELM network after adding new nodes. ;

[0035] when or When the algorithm ends, the optimal network structure obtained at this point is the meteorological signal radial velocity and spectral width prediction model. The weight vector connecting the hidden layer nodes and the input layer nodes is calculated: Time correspondence , Time correspondence And record the parameters selected each time. , , .

[0036] Furthermore, the method for predicting and calculating the radial velocity of meteorological signals in S4 is as follows:

[0037]

[0038] The method for calculating spectral width prediction is as follows:

[0039]

[0040] in, , For the first Power spectrum of each distance cell, , The parameters selected in step three. , The parameters obtained in step three, For the first The radial velocity prediction value of meteorological signals for each distance unit. For the first The predicted spectral width of meteorological signals for each distance unit.

[0041] The beneficial effects of adopting the above technical solution are as follows:

[0042] 1. The method of meteorological signal spectral moment estimation based on dual-ended incremental extreme learning machine has fast convergence speed, small error and low computational complexity, which is conducive to engineering implementation.

[0043] 2. The experimental data verification results show that the meteorological signal spectral moment estimation method based on dual-end incremental extreme learning machine can significantly reduce the estimation bias of radial velocity and spectral width of meteorological signals and improve the accuracy of meteorological signal spectral moment estimation. Attached Figure Description

[0044] Figure 1 This is a flowchart of the signal processing of the present invention;

[0045] Figure 2 This is a flowchart of the B-ELM algorithm of the present invention;

[0046] Figure 3To compare the radial velocity estimates with the true radial velocity values ​​obtained by applying the ELM and B-ELM algorithms;

[0047] Figure 4 To compare the spectral width estimates of the ELM and B-ELM algorithms with the true values ​​of the spectral moment and spectral width;

[0048] Figure 5 The absolute error between the radial velocity estimates and the true values ​​of the three algorithms;

[0049] Figure 6 The absolute error between the spectral width estimates of the three algorithms and the true values ​​is given. Detailed Implementation

[0050] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] This invention mainly studies a meteorological signal spectral moment estimation method based on a two-ended incremental extreme learning machine. Figure 1 This is the signal processing flow. Its main steps are as follows:

[0052] Step 1: Constructing training samples, specifically:

[0053] Assuming the weather radar is for the first Spectral moment estimation is performed on each distance cell, and spectral moments are selected on both sides thereof. Each distance cell is used as a training sample. To estimate the spectral moment information of the meteorological signal with high accuracy, the time-domain pulse signal of each distance cell is converted into a meteorological power spectrum. The meteorological signal power spectrum is used as the training sample input of the network, which significantly improves the input signal-to-noise ratio of the meteorological signal.

[0054] This invention employs the Welch method (mean corrected periodogram) for power spectrum estimation. It assumes that the meteorological pulse data of each training sample, with a coherent accumulation pulse number of 64, are used. Divided into Segments, each segment is [length missing] , No. The signal segments are:

[0055]

[0056] in, It is the first The starting point of the segment, when Data overlap and segmentation can occur at this time, which can be taken as... At that time, 50% data overlap will occur.

[0057] Introducing a Hamming window function before calculating the average periodogram The correction periodicity chart corresponding to each data segment is calculated using the following formula, the first... The correction periodicity diagram for the segment is as follows:

[0058]

[0059] In the formula, This is called the normalization factor.

[0060] The power spectrum estimate can be obtained by averaging using the modified periodogram method:

[0061]

[0062] After obtaining the power spectrum estimate, the meteorological power spectrum is used as the input sample, and the spectral moments are used as the output sample to construct training samples. Assume the... Each distance cell requires spectral moment estimation. Different training samples There are the following construction methods:

[0063] ,

[0064] in For input, For output, It is the number of input layer nodes. The number of output layer nodes is given. For meteorological signals, the samples constructed in this method are as follows:

[0065]

[0066] In the formula, For the first The power spectrum of each training sample.

[0067] Step 2, Model Parameter Selection and Optimization, specifically:

[0068] In the B-ELM model, the parameters of the hidden nodes are set randomly. , It is the first The weight vector between each hidden layer node and the input layer For the first The threshold for each hidden layer node sets the number of input layer nodes. The number of output layer nodes is Select activation function as The function is finally calculated using the least squares method to connect the first... The weight vector between each hidden layer node and the output layer node .

[0069] Step 3: Construct the prediction model using the B-ELM algorithm, specifically as follows:

[0070] Figure 2 This is the flowchart of the B-ELM algorithm, for the set... training samples , , ,in, The first Input and output of each sample The number of nodes in the input layer. Let be the number of nodes in the output layer. Let the maximum number of hidden layer nodes and the expected error be respectively... , When the output is radial velocity and spectral width, set The learning process of the B-ELM algorithm is as follows:

[0071] Initialization phase: setting , The number of hidden layer nodes in the network is denoted as , and the network residual is denoted as . ,in Indicates the expected output;

[0072] Learning stage: When and hour:

[0073] 1) Add a hidden layer node: ;

[0074] 2) When When the number is odd: Randomly select parameters for the newly added hidden layer nodes. ;when When the number is even: Calculate the error feedback matrix Randomly set the parameters of the hidden layer nodes ;

[0075] 3) Calculate the weight output matrix ;

[0076] 4) Calculate the weights between the newly added hidden nodes and the output layer. ;

[0077] 5) Calculate the network residuals of the B-ELM network after adding new nodes. ;

[0078] when or When the algorithm ends, the optimal network structure obtained at this point is the meteorological signal radial velocity and spectral width prediction model. The weight vector connecting the hidden layer nodes and the input layer nodes is calculated: Time correspondence , Time correspondence And record the parameters selected each time. , , .

[0079] Step four: Estimation of the radial velocity and spectral width of the meteorological signal, specifically:

[0080] Based on the optimal network structure obtained in step three, the first step in step one... By predicting the radial velocity and spectral width of each distance cell, the estimated values ​​of the radial velocity and spectral width of the meteorological signal can be obtained:

[0081] ,

[0082] in, , For the first Power spectrum of each distance cell, , The parameters selected in step three. , The parameters obtained in step three, For the first Radial velocity estimates of meteorological signals for each distance cell. For the first The estimated spectral width of the meteorological signal for each distance unit.

[0083] The effectiveness of the algorithm of this invention is verified by experimental data. This invention uses experimental data from the IRCTR Rain Radar (IDRA) to compare the spectral moment estimation accuracy of different algorithms. IDRA is a weather radar system designed and developed by the International Centre for Telecommunications and Radar Research (IRCTR) at Delft University of Technology. The IDRA radar parameters are shown in Table 1. The data selected for this experiment is from September 10, 2011, UTC time 21:00–22:00, within the standard range mode, using experimental data from sector 25. To test the effectiveness of the algorithm of this invention, the spectral moments obtained by the Fast Fourier Transform (FFT) of 512 coherent accumulation pulses are used as the true values ​​and compared with the spectral moments obtained by the algorithm of this invention with 64 pulses.

[0084] Table 1 IDRA Radar Parameters

[0085]

[0086] This experiment estimates the radial velocity and spectral width of meteorological signals and analyzes the performance of the B-ELM algorithm under different distance cell conditions. Figure 3 , Figure 4This compares the estimated spectral moments with the true spectral moments obtained using the ELM and B-ELM algorithms, respectively. The true radial velocity varies slowly between 6.6 and 8.3 m / s. Figure 3 It is evident that the radial velocity estimates from both algorithms are very close to the true values, and they also change slowly within this range. The radial velocity gradually decreases between 30 and 60 distance cells, and increases rapidly between 70 and 80 distance cells. This experiment demonstrates that both algorithms can accurately estimate the spectral moment when the radial velocity changes rapidly. Figure 4 It can be seen that the true spectral width changes slowly between 0.3 and 1.7 m / s, with most spectral width values ​​distributed between 0.4 and 0.8 m / s. The deviations of the ELM algorithm and the B-ELM algorithm from the true values ​​are very small.

[0087] Figure 5 The graph compares the absolute values ​​of radial velocity errors. It can be seen from the graph that the estimation error of the off-grid reconstruction algorithm is significantly greater than that of the ELM and B-ELM algorithms in the vast majority of distance cells. The B-ELM algorithm, on the other hand, has an estimation error of less than 0.2 m / s, exhibiting the smallest error and the most accurate estimation. Figure 6 Comparing the absolute values ​​of the spectral width errors, the figure shows that the estimation error of the off-grid reconstruction algorithm is much larger than that of the ELM and B-ELM algorithms in the vast majority of distance cells. The estimation error of the B-ELM algorithm is generally less than 0.5 m / s, and its estimation performance is close to that of the ELM algorithm.

[0088] Table 2 compares the average errors of spectral moment estimation using different algorithms. This experiment used the spectral moments obtained by FFT with 512 pulses as the true values. Table 2 shows that the B-ELM algorithm performs slightly better than the ELM algorithm, while the off-gate reconstruction algorithm performs the worst. The average radial velocity error of the B-ELM algorithm is 0.0766 m / s, and the spectral width estimation error is 0.1063 m / s, representing improvements of 7% and 11% respectively compared to the ELM algorithm. The B-ELM algorithm improves radial velocity estimation performance by 76.3% and spectral width estimation performance by 57.4% compared to the off-gate reconstruction algorithm.

[0089] Table 2 Algorithm Error Comparison

[0090]

[0091] Therefore, the spectral moment estimation method based on dual-ended incremental limit learning machine studied in this invention has significantly better performance than the off-grid reconstruction algorithm and the ELM algorithm, and has lower computational complexity and faster convergence speed, which is beneficial for engineering implementation.

[0092] The embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for estimating the spectrum moments of meteorological signals based on a dual-end incremental extreme learning machine, characterized in that, The method includes the following steps: S1. Obtain meteorological pulse data, convert the meteorological pulse data into meteorological power spectrum using the average corrected periodogram method, use the meteorological power spectrum as the input of network training samples, and use the spectral moments as the output of network training samples to construct training samples. S2. Randomly set the parameters of the hidden nodes, determine the number of nodes in the input layer and the number of nodes in the output layer, and select the activation function; S3. Construct a prediction model based on the B-ELM algorithm to obtain the optimal prediction model for the radial velocity and spectral width of the meteorological signal; specifically: Let the maximum number of hidden layer nodes and the expected error be respectively... and When the output is radial velocity and spectral width, set ; Initialization phase: setting , The number of hidden layer nodes in the network is denoted as , and the network residual is denoted as . ,in Indicates the expected output; Learning stage: When and hour: 1) Add a hidden layer node: ; 2) When When the number is odd: Randomly select parameters for the newly added hidden layer nodes. ;when When the number is even: Calculate the error feedback matrix Randomly set the parameters of the hidden layer nodes ; 3) Calculate the weight output matrix ; in, For input; 4) Calculate the weights between the newly added hidden nodes and the output layer. ; in, The mesh residual of the B-ELM network for the current hidden layer nodes. This is the transpose of the weight output matrix; 5) Calculate the network residuals of the B-ELM network after adding new nodes. ; when or When the algorithm ends, the optimal network structure obtained at this point is the meteorological signal radial velocity and spectral width prediction model. The weight vector connecting the hidden layer nodes and the input layer nodes is calculated: Time correspondence , Time correspondence And record the parameters selected each time. , , ;in, For output; S4. Substitute the training samples, the set number of input layer nodes, and the number of output layer nodes into the obtained optimal meteorological signal radial velocity and spectral width prediction model to obtain the radial velocity and spectral width of the meteorological signal.

2. The meteorological signal spectral moment estimation method based on dual-ended incremental extreme learning machine according to claim 1, characterized in that, The steps for estimating the meteorological power spectrum using the average corrected periodogram method in S1 are as follows: Assuming the weather radar is for the first Spectral moment estimation is performed on each distance cell, and spectral moments are selected on both sides thereof. 1 distance unit is used as a training sample; S1.1, Meteorological pulse data Divided into Segments, each segment is [length missing] , No. The signal segments are: ； in, It is the first The starting point of the segment, when When data overlaps and is segmented, take... At that time, 50% data overlap occurred; S1.2 Introducing the Hamming window function The correction periodicity chart corresponding to each data segment is calculated using the following formula, the first... The correction period diagram for the segment is as follows: ； In the formula, This is called the normalization factor. ; S1.

3. The meteorological power spectrum is estimated by averaging using the modified periodogram method: 。 3. The meteorological signal spectral moment estimation method based on dual-ended incremental extreme learning machine according to claim 1, characterized in that, The method for constructing the training samples in S1 is as follows: ， ； in, For input, For output, It is the number of input layer nodes. The number of output layer nodes is given. For meteorological signals, the samples constructed in this method are as follows: ； in, For the first The power spectrum of each training sample This represents the number of training samples.

4. The meteorological signal spectral moment estimation method based on dual-ended incremental extreme learning machine according to claim 3, characterized in that, The method for predicting and calculating the radial velocity of meteorological signals in S4 is as follows: ； The method for calculating spectral width prediction is as follows: ； in, , For the first Power spectrum of each distance cell, , The parameters selected in step three. , The parameters obtained in step three, For the first The radial velocity prediction value of meteorological signals for each distance unit. For the first The predicted spectral width of meteorological signals for each distance unit.

Images