Laser radar signal processing and wind speed estimation method and system based on coherent detection

Abstract

The present invention provides a laser radar signal processing and wind speed estimation method and system based on coherent detection, which relates to the field of laser measurement technology, and comprises: acquiring an echo signal and a reference signal of a wind measuring laser radar, performing frequency mixing and filtering processing, generating a difference frequency signal, performing noise reduction and background suppression, obtaining difference frequency signal data, a window function and a resolution of spectrum analysis, performing time-frequency analysis, generating a Doppler spectrum and a time-frequency spectrum; extracting Doppler frequency shift, spectrum width and signal-to-noise ratio, estimating wind field parameters, performing refined correction and adaptive compensation, identifying special flow patterns, extracting high-level semantic features, and obtaining refined wind field parameters; constructing a global wind field model, converting adaptive scanning optimization into a Markov decision process, determining an action space and a state space, setting a reward function, solving an optimal scanning strategy in combination with a reinforcement learning algorithm, predicting a wind field state, and generating a state prediction report.

Description

Technical Field

[0001] The present invention relates to the field of laser measurement technology, and in particular to a laser radar signal processing and wind speed estimation method and system based on coherent detection. Background Art

[0002] With the increasing demand for wind field information in fields such as weather forecasting, aviation safety, and wind power development, wind laser radar has become one of the important means of obtaining wind field data. Traditional wind laser radar mainly uses incoherent detection methods to directly process the intensity or power of the echo signal to extract parameters such as wind speed and wind direction. However, incoherent detection methods have problems such as low signal-to-noise ratio, limited measurement accuracy, and susceptibility to atmospheric turbulence and aerosols, making it difficult to meet the increasing demand for wind field detection.

[0003] Coherent detection uses the coherence between the echo signal and the reference signal to accurately measure the Doppler frequency shift, thereby obtaining a high-precision wind speed estimate. However, coherent detection also faces a series of technical challenges, such as unstable echo signal quality, high signal processing complexity, and long coherent accumulation time. The existing wind lidar system based on coherent detection still has deficiencies in signal processing and wind speed estimation.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a laser radar signal processing and wind speed estimation method and system based on coherent detection, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a laser radar signal processing and wind speed estimation method based on coherent detection, comprising:

[0007] Acquire the echo signal and the reference signal of the wind measuring laser radar, adaptively select the mixing mode and the filtering parameters according to the statistical characteristics and noise level of the echo signal and the reference signal, perform mixing and filtering on the echo signal and the reference signal, generate a difference frequency signal, and perform noise reduction and background suppression through an adaptive threshold and an iterative optimization algorithm to obtain difference frequency signal data, adaptively select the window function and resolution of spectrum analysis based on the time-frequency characteristics of the difference frequency signal data, perform time-frequency analysis on the difference frequency signal data through short-time Fourier transform, and generate a Doppler spectrum and a time-frequency spectrum;

[0008] Extracting Doppler frequency shift, spectrum width and signal-to-noise ratio from the Doppler spectrum, estimating wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight as wind field parameters by combining meteorological knowledge and machine learning algorithms, finely correcting the wind field parameters through deep neural networks and adaptively compensating for atmospheric turbulence and aerosol extinction, identifying special flow patterns in the wind field through out-of-control correlation analysis and data mining based on the corrected wind field parameters, extracting high-level semantic features in the special flow patterns through convolutional neural networks, and obtaining refined wind field parameters;

[0009] Based on the refined wind field parameters, a global wind field model is constructed. According to the global wind field model, the adaptive scanning optimization is converted into a Markov decision process. The action space and state space are determined according to the environmental state and task requirements. At the same time, the corresponding reward function is set. The Markov decision process is solved in combination with the reinforcement learning algorithm to obtain the optimal scanning strategy. Based on the optimal scanning strategy, the wind field state is predicted through a long short-term memory network to generate a state prediction report.

[0010] In an optional embodiment,

[0011] Acquire the echo signal and the reference signal of the wind laser radar, adaptively select the mixing mode and the filtering parameters according to the statistical characteristics and noise level of the echo signal and the reference signal, perform mixing and filtering on the echo signal and the reference signal, generate a difference frequency signal, and perform noise reduction and background suppression through an adaptive threshold and an iterative optimization algorithm to obtain difference frequency signal data, adaptively select the window function and resolution of spectrum analysis based on the time-frequency characteristics of the difference frequency signal data, perform time-frequency analysis on the difference frequency signal data through short-time Fourier transform, and generate a Doppler spectrum and a time-frequency spectrum, including:

[0012] Acquire the echo signal and the reference signal of the wind measuring laser radar, pre-process the echo signal and the reference signal, remove the DC component and the trend term, calculate the power spectrum density and the cross-power spectrum density of the echo signal and the reference signal, obtain the frequency domain statistical characteristics, estimate the main frequency and the frequency bandwidth of the echo signal and the reference signal according to the peak position and width of the cross-power spectrum density, judge whether the main frequencies of the echo signal and the reference signal are close and whether the frequency bandwidth is narrow according to the preset judgment conditions, if the main frequencies are close and the frequency bandwidth is narrow, use the average value of the main frequencies of the echo signal and the reference signal as the local oscillator frequency for direct mixing, if the main frequencies are not close or the frequency bandwidth is wide, use the main frequency difference of the echo signal and the reference signal as the local oscillator frequency, mix through the superheterodyne receiving mixing mode, and obtain the mixed signal;

[0013] The mixed signal is divided into a plurality of signal segments of equal length, the mean, variance and skewness are calculated for each signal segment, and the time domain statistical features are obtained by combining them, and the stationarity and Gaussianity of each signal segment are judged according to the statistics and the empirical threshold value, for the stationary Gaussian signal segment, the cutoff frequency is selected by a low-pass filter, and for the non-stationary or non-Gaussian signal segment, the filter order and the noise variance are adaptively determined according to the statistics by an adaptive Wiener filter, and the filtering is repeated until all signals are processed to obtain the difference frequency signal;

[0014] The difference frequency signal is subjected to wavelet decomposition to generate sub-bands of different scales. On each sub-band, the mean and standard deviation of the wavelet coefficients are calculated to obtain the statistical characteristics of noise at different scales. According to the standard deviation and the empirical factor of the wavelet coefficients, the noise reduction threshold of each sub-band is adaptively set and the wavelet coefficients are subjected to soft threshold processing. The processed wavelet coefficients are subjected to wavelet reconstruction to obtain the difference frequency signal after noise reduction. The power spectrum of the difference frequency signal after noise reduction is estimated to generate an initial Doppler spectrum. The initial Doppler spectrum is modeled and matched filtered by a preset Gaussian mixture model to obtain a spectrum residual. The Gaussian mixture model is updated based on the spectrum residual until convergence. The wind speed peak position and width are extracted from the updated Gaussian mixture model to obtain a background suppression signal. The background suppression signal is subjected to short-time Fourier transform to obtain a time-frequency signal and the statistical characteristics of the time-frequency signal are calculated. The type, length and resolution of the spectrum analysis window are selected based on the statistical characteristics and applied to the time-frequency signal to obtain the Doppler spectrum and the time-frequency spectrum.

[0015] In an optional embodiment,

[0016] The probability density function corresponding to the Gaussian mixture model is shown in the following formula:

[0017]

[0018] Where x represents the observation data vector, K represents the total number of Gaussian components in the Gaussian mixture model, k represents the index of the Gaussian component, D represents the dimension of the observation data vector, and α k represents the mixing coefficient, μ k represents the mean vector of the kth Gaussian component, ∑ k represents the covariance matrix of the kth Gaussian component, and T represents the transpose.

[0019] In an optional embodiment,

[0020] Doppler frequency shift, spectrum width and signal-to-noise ratio are extracted from the Doppler spectrum, and the wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight are estimated as wind field parameters by combining meteorological knowledge and machine learning algorithms. The wind field parameters are refined and corrected by deep neural networks, and atmospheric turbulence and aerosol extinction are adaptively compensated. According to the corrected wind field parameters, special flow patterns in the wind field are identified by out-of-control correlation analysis and data mining, and high-level semantic features in the special flow patterns are extracted by convolutional neural networks. The refined wind field parameters include:

[0021] The Doppler spectrum is fitted by a dual-line fitting method to obtain two peak frequencies and calculate the peak frequency difference, and the Doppler frequency shift, spectrum width and signal-to-noise ratio are obtained based on the peak frequency difference and the radar operating wavelength;

[0022] The wind speed in the line of sight direction is determined by combining meteorological knowledge and the Doppler frequency shift, and the mapping relationship between the wind speed error and the turbulence intensity is determined based on the correlation between the wind speed and the refractive index turbulence intensity. Based on the mapping relationship, the wind speed error is predicted according to the observed wind speed and the radial wind speed is compensated to obtain the wind direction, turbulence intensity and atmospheric stability in the line of sight direction. The wind direction, turbulence intensity and atmospheric stability are used as wind field parameters, and the historical wind field parameters of the current area are obtained and added to a preset neural network model to train the neural network model. The trained model is used to perform fine correction on the true value of the wind field parameter, and the correction amount is dynamically adjusted according to the intensity of atmospheric turbulence and aerosol extinction in combination with the adaptive compensation mechanism to obtain the corrected wind field parameters;

[0023] Based on the corrected wind field parameters, the wind field is represented as a spatiotemporal three-dimensional data cube, wherein each position contains a wind field parameter corresponding to the current position, and based on the corrected wind field parameters, the similarity of each position in the spatiotemporal three-dimensional data cube is calculated, and the wind field is clustered by a density clustering algorithm, and the clustered regions are further summarized to obtain a classic flow pattern of the wind field, and the spatiotemporal evolution law between different flow patterns is determined by an association rule mining algorithm, and the special flow pattern is identified by comparing the spatiotemporal evolution law with the classic flow pattern of the wind field;

[0024] The area corresponding to the special flow pattern is represented as a two-dimensional image, in which rows represent heights and columns represent time. Features of the two-dimensional image are extracted by a pre-set convolutional neural network to obtain high-level semantic features. The high-level semantic features are fused with the corrected wind field parameters to obtain the refined wind field parameters.

[0025] In an optional embodiment,

[0026] Based on the corrected wind field parameters, the similarity of each position in the spatiotemporal three-dimensional data cube is calculated as shown in the following formula:

[0027]

[0028] Where DIS represents the wind field parameter x i and x j The similarity between them, θ represents the mean vector of wind field parameters, n represents the number of samples, (x g -θ) represents the sample x g The difference vector relative to the mean vector, g represents the sample index, (x i -x j ) represents the difference vector between the wind field parameter vectors at two locations.

[0029] In an optional embodiment,

[0030] Based on the refined wind field parameters, a global wind field model is constructed. According to the global wind field model, the adaptive scanning optimization is converted into a Markov decision process. The action space and state space are determined according to the environmental state and task requirements. At the same time, the corresponding reward function is set. The Markov decision process is solved in combination with the reinforcement learning algorithm to obtain the optimal scanning strategy. Based on the optimal scanning strategy, the wind field state is predicted through the long short-term memory network, and the state prediction report is generated, including:

[0031] Based on the refined wind field parameters, the wind field is represented as a four-dimensional tensor in combination with a spatiotemporal wind field representation method. Based on the four-dimensional tensor, each spatial dimension and time dimension is linearly interpolated in combination with a trilinear interpolation method to obtain a continuous wind field representation. Each four-dimensional tensor in the continuous wind field representation is reshaped into a two-dimensional matrix, wherein each row corresponds to a spatiotemporal position, and each column corresponds to a refined wind field parameter. The two-dimensional matrix is ​​decentralized and the covariance matrix corresponding to the two-dimensional matrix is ​​calculated. The covariance matrix is ​​eigenvalue decomposed to obtain an eigenvector matrix and an eigenvalue vector. The eigenvalue vector is sorted in descending order, and the eigenvectors corresponding to the eigenvalues ​​in the top 10% are selected to form a principal component matrix. The refined wind field parameters are mapped to the principal component space corresponding to the principal component matrix to obtain the global wind field model.

[0032] Based on the global wind field model, the adaptive scanning problem is modeled as a Markov decision process, scanning tasks and meteorological data are obtained, the wind field is divided into grids, a state vector is generated based on the wind field state corresponding to each grid and the working state of the radar, the state vector is combined to obtain a state space, the scanning action of the radar in the current state is obtained and the pulse repetition frequency is adjusted to generate an action space, and the state transition probability and the corresponding reward function are determined;

[0033] Based on the constructed Markov decision, the objective function is constructed through Q learning in the reinforcement learning algorithm, the optimal scanning strategy is solved based on the objective function and the state transition probability, and the optimal scanning strategy is applied to the radar scanning control. At each decision moment, the radar selects the scanning action and executes the control instruction according to the current state in combination with the optimal scanning strategy to obtain real-time wind field data;

[0034] A wind farm sequence is constructed for the real-time wind farm data, and the wind farm sequence is added as input data to the long short-term memory network to generate a predicted state. The predicted state is compared with the true label corresponding to the wind farm sequence, and the prediction loss is calculated and the hyperparameters of the long short-term memory network are updated according to the prediction loss. The update is repeated until the model converges, and the wind farm sequence is predicted again using the updated long short-term memory network to obtain the predicted wind farm state, and the state prediction report is generated.

[0035] In an optional embodiment,

[0036] The objective function constructed by Q learning in the reinforcement learning algorithm is shown in the following formula:

[0037]

[0038] Among them, J(Q) represents the objective function value, E represents the expectation, r represents the immediate reward, γ represents the discount factor, and s' represents the next state after executing action a. represents the maximum expected cumulative reward after selecting the optimal action a' in state s', and Q(s, a) represents the expected cumulative reward after executing action a in state s.

[0039] A second aspect of an embodiment of the present invention provides a laser radar signal processing and wind speed estimation system based on coherent detection, comprising:

[0040] The first unit is used to obtain the echo signal and the reference signal of the wind measurement laser radar, adaptively select the mixing mode and the filtering parameters according to the statistical characteristics and noise level of the echo signal and the reference signal, mix and filter the echo signal and the reference signal, generate a difference frequency signal, and perform noise reduction and background suppression through an adaptive threshold and an iterative optimization algorithm to obtain difference frequency signal data, adaptively select the window function and resolution of spectrum analysis based on the time-frequency characteristics of the difference frequency signal data, perform time-frequency analysis on the difference frequency signal data through short-time Fourier transform, and generate a Doppler spectrum and a time-frequency spectrum;

[0041] The second unit is used to extract Doppler frequency shift, spectrum width and signal-to-noise ratio from the Doppler spectrum, estimate the wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight as wind field parameters by combining meteorological knowledge and machine learning algorithms, and perform fine correction on the wind field parameters and adaptively compensate for atmospheric turbulence and aerosol extinction through deep neural networks. According to the corrected wind field parameters, special flow patterns in the wind field are identified through out-of-control correlation analysis and data mining, and high-level semantic features in the special flow patterns are extracted through convolutional neural networks to obtain refined wind field parameters;

[0042] The third unit is used to build a global wind field model based on the refined wind field parameters, convert the adaptive scanning optimization into a Markov decision process according to the global wind field model, determine the action space and state space according to the environmental state and task requirements, and set the corresponding reward function. The Markov decision process is solved in combination with the reinforcement learning algorithm to obtain the optimal scanning strategy. Based on the optimal scanning strategy, the wind field state is predicted through a long short-term memory network to generate a state prediction report.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the embodiments of the present invention is:

[0049] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0050] In the present invention, the data-driven adaptive processing method can dynamically adjust the processing strategy according to the actual signal quality, effectively improve the signal-to-noise ratio and noise reduction effect, thereby providing high-quality difference frequency signal data for subsequent wind field parameter extraction. In the wind field parameter extraction link, the basic parameters such as wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight are first estimated from the Doppler spectrum, and then these parameters are refined and corrected through a deep neural network, and the influence of atmospheric turbulence and aerosol extinction is adaptively compensated. The high-dimensional characteristics of the Doppler spectrum can be fully explored, and the complex mapping relationship between wind field parameters and spectral characteristics can be learned, thereby significantly improving the accuracy of parameter estimation. Adaptive atmospheric effect compensation can reduce the interference of turbulence and aerosols on wind field perception, improve the reliability of wind field parameters, and extract convolutional neural networks. High-level semantic features in special flow patterns, such as the intensity, scale, and movement trajectory of vortices, can further enrich the content of wind field perception and provide more interpretable and operational wind field information. The adaptive scanning strategy can intelligently adjust the scanning angle, speed, density and other parameters of the wind measuring lidar according to the real-time status and evolution trend of the wind field, while maximizing the coverage range while ensuring the detection accuracy, thereby significantly improving the timeliness and coverage of wind field detection. In summary, the present invention significantly improves the intelligence level and wind field perception capability of the wind measuring lidar system, and realizes the optimization and upgrading of the entire chain from data collection to wind field early warning. It is expected to promote technological progress and model changes in wind field perception and related application fields, and provide more accurate, efficient and intelligent decision-making support for meteorological forecasting, aviation safety, wind power development, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of a laser radar signal processing and wind speed estimation method based on coherent detection according to an embodiment of the present invention;

[0052] Figure 2 It is a structural schematic diagram of a laser radar signal processing and wind speed estimation system based on coherent detection according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0055] Figure 1 FIG. 1 is a flow chart of a laser radar signal processing and wind speed estimation method based on coherent detection according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] S1. Acquire the echo signal and reference signal of the wind laser radar, adaptively select the mixing mode and filtering parameters according to the statistical characteristics and noise level of the echo signal and the reference signal, perform mixing and filtering on the echo signal and the reference signal, generate a difference frequency signal, and perform noise reduction and background suppression through an adaptive threshold and iterative optimization algorithm to obtain difference frequency signal data, adaptively select the window function and resolution of spectrum analysis based on the time-frequency characteristics of the difference frequency signal data, perform time-frequency analysis on the difference frequency signal data through short-time Fourier transform, and generate a Doppler spectrum and a time-frequency spectrum;

[0057] The echo signal is the signal reflected back after the signal emitted by the laser radar interacts with the target object. The reference signal is a signal compared with the echo signal and is used to extract useful information. The mixing method refers to the process of mixing the echo signal with the reference signal to generate a difference frequency signal. The filtering parameter refers to the parameter used to select or suppress signals within a specific frequency range in signal processing. The background suppression is to remove or reduce irrelevant background noise or interference signals in the echo signal, thereby improving the signal-to-noise ratio of the useful signal. The difference frequency signal is the result of mixing the echo signal with the reference signal, and contains information about the target object, such as Doppler shift. The time-frequency characteristic describes the distribution of the signal in time and frequency. The window function is a function applied to the signal when the signal is locally processed to reduce the boundary effect. The short-time Fourier transform is to divide the signal into short time periods and then perform Fourier transform on each section to obtain the time-frequency representation of the signal. The Doppler spectrum is the spectrum obtained by analyzing the difference frequency signal, which reflects the relative speed information of the target relative to the radar. The time-frequency spectrum is a joint representation of the signal in the time and frequency domains, showing how the frequency of the signal changes over time.

[0058] In an optional embodiment,

[0059] Acquire the echo signal and the reference signal of the wind laser radar, adaptively select the mixing mode and the filtering parameters according to the statistical characteristics and noise level of the echo signal and the reference signal, perform mixing and filtering on the echo signal and the reference signal, generate a difference frequency signal, and perform noise reduction and background suppression through an adaptive threshold and an iterative optimization algorithm to obtain difference frequency signal data, adaptively select the window function and resolution of spectrum analysis based on the time-frequency characteristics of the difference frequency signal data, perform time-frequency analysis on the difference frequency signal data through short-time Fourier transform, and generate a Doppler spectrum and a time-frequency spectrum, including:

[0060] Acquire the echo signal and the reference signal of the wind measuring laser radar, pre-process the echo signal and the reference signal, remove the DC component and the trend term, calculate the power spectrum density and the cross-power spectrum density of the echo signal and the reference signal, obtain the frequency domain statistical characteristics, estimate the main frequency and the frequency bandwidth of the echo signal and the reference signal according to the peak position and width of the cross-power spectrum density, judge whether the main frequencies of the echo signal and the reference signal are close and whether the frequency bandwidth is narrow according to the preset judgment conditions, if the main frequencies are close and the frequency bandwidth is narrow, use the average value of the main frequencies of the echo signal and the reference signal as the local oscillator frequency for direct mixing, if the main frequencies are not close or the frequency bandwidth is wide, use the main frequency difference of the echo signal and the reference signal as the local oscillator frequency, mix through the superheterodyne receiving mixing mode, and obtain the mixed signal;

[0061] The mixed signal is divided into a plurality of signal segments of equal length, the mean, variance and skewness are calculated for each signal segment, and the time domain statistical features are obtained by combining them, and the stationarity and Gaussianity of each signal segment are judged according to the statistics and the empirical threshold value, for the stationary Gaussian signal segment, the cutoff frequency is selected by a low-pass filter, and for the non-stationary or non-Gaussian signal segment, the filter order and the noise variance are adaptively determined according to the statistics by an adaptive Wiener filter, and the filtering is repeated until all signals are processed to obtain the difference frequency signal;

[0062] The difference frequency signal is subjected to wavelet decomposition to generate sub-bands of different scales. On each sub-band, the mean and standard deviation of the wavelet coefficients are calculated to obtain the statistical characteristics of noise at different scales. According to the standard deviation and the empirical factor of the wavelet coefficients, the noise reduction threshold of each sub-band is adaptively set and the wavelet coefficients are subjected to soft threshold processing. The processed wavelet coefficients are subjected to wavelet reconstruction to obtain the difference frequency signal after noise reduction. The power spectrum of the difference frequency signal after noise reduction is estimated to generate an initial Doppler spectrum. The initial Doppler spectrum is modeled and matched filtered by a preset Gaussian mixture model to obtain a spectrum residual. The Gaussian mixture model is updated based on the spectrum residual until convergence. The wind speed peak position and width are extracted from the updated Gaussian mixture model to obtain a background suppression signal. The background suppression signal is subjected to short-time Fourier transform to obtain a time-frequency signal and the statistical characteristics of the time-frequency signal are calculated. The type, length and resolution of the spectrum analysis window are selected based on the statistical characteristics and applied to the time-frequency signal to obtain the Doppler spectrum and the time-frequency spectrum.

[0063] The trend term is the long-term change trend in the signal, which is usually related to the overall direction or growth pattern of the signal. The power spectral density describes how the power of the signal is distributed at different frequencies. The cross-power spectral density is a measure of the relationship between two signals in the frequency domain. The frequency domain statistical characteristics describe the statistical properties of the signal in the frequency domain, such as frequency distribution, average frequency, bandwidth, etc. The local oscillator frequency is the reference frequency generated by the local oscillator and is used for mixing with the input signal. The Gaussianity describes the degree to which the signal or noise obeys the Gaussian distribution. The filter order is a key parameter in the filter design and represents the order of the filter polynomial. The sub-band is formed by dividing the frequency band of the signal into several narrower frequency bands. The empirical factor The sub is a parameter determined based on experience or experimental data, which is used to adjust the performance of the model or algorithm. The noise reduction threshold is a boundary value set in the signal noise reduction processing, which is used to distinguish between signals and noise. The soft threshold processing is a signal noise reduction technology that reduces noise by gradually reducing the wavelet coefficients to zero. The wavelet reconstruction refers to the process of restoring the decomposed wavelet coefficients to the original signal through inverse wavelet transform. The matched filter is a filtering technology for detecting known signals. It maximizes the signal-to-noise ratio of the signal by convolving the received signal with a known template of the expected signal. The spectral residual is a method used in signal processing and anomaly detection. Anomalies are identified by calculating the difference between the spectrum of the signal and its smoothed spectrum.

[0064] Acquire the echo signal and reference signal of the wind measuring laser radar, pre-process the echo signal and the reference signal to remove the DC component and the trend term, calculate the power spectrum density and the cross-power spectrum density of the echo signal and the reference signal, obtain the frequency domain statistical characteristics, estimate the main frequency and frequency bandwidth of the echo signal and the reference signal according to the peak position and width of the cross-power spectrum density, judge whether the main frequencies of the echo signal and the reference signal are close, and whether the frequency bandwidth is narrow; if the main frequencies are close and the frequency bandwidth is narrow, use the average value of the main frequencies of the echo signal and the reference signal as the local oscillator frequency, and perform direct mixing; if the main frequencies are not close or the frequency bandwidth is wide, use the main frequency difference of the echo signal and the reference signal as the local oscillator frequency, perform mixing through the superheterodyne receiving mixing method, and obtain a mixed signal;

[0065] The mixed signal is divided into multiple signal segments of equal length, and the mean, variance and skewness are calculated for each signal segment respectively, and the time domain statistical characteristics are obtained by combining them. The stationarity and Gaussianity of each signal segment are judged according to the statistics and the empirical threshold. For the stationary Gaussian signal segment, the cutoff frequency is selected by a low-pass filter for filtering. For the non-stationary or non-Gaussian signal segment, the filter order and the noise variance are adaptively determined according to the statistics by an adaptive Wiener filter. The filtering is repeated until all signal segments are processed to obtain the difference frequency signal. The difference frequency signal is subjected to wavelet decomposition to generate sub-bands of different scales. On each sub-band, the mean and standard deviation of the wavelet coefficients are calculated to obtain the statistical characteristics of the noise at different scales. According to the standard deviation and the empirical factor of the wavelet coefficients, the noise reduction threshold of each sub-band is adaptively set, the wavelet coefficients are subjected to soft threshold processing, and the processed wavelet coefficients are subjected to wavelet reconstruction to obtain the noise-reduced difference frequency signal.

[0066] Perform power spectrum estimation on the difference frequency signal after noise reduction to generate an initial Doppler spectrum, perform modeling and matched filtering on the initial Doppler spectrum through a preset Gaussian mixture model to obtain a spectrum residual, update the Gaussian mixture model based on the spectrum residual, repeat the update until convergence, extract the wind speed peak position and width in the updated Gaussian mixture model to obtain a background suppression signal, perform short-time Fourier transform on the background suppression signal to obtain a time-frequency signal, calculate the statistical characteristics of the time-frequency signal, select the type, length and resolution of the spectrum analysis window based on the statistical characteristics, apply the selected spectrum analysis window to the time-frequency signal to obtain the final Doppler spectrum and time-frequency spectrum;

[0067] Exemplarily, it is assumed that an echo signal and a reference signal with a duration of 10 seconds are obtained, and the sampling frequency is 10kHz. The mean values ​​of the echo signal and the reference signal are respectively subtracted to remove the DC component. The linear trend term of the signal is removed by using the first-order difference method. The power spectrum density and the cross-power spectrum density are calculated. The window length is set to 1000 samples, and the overlap rate is 50%. By finding the maximum value position of the cross-power spectrum density, it is estimated that the main frequencies of the echo signal and the reference signal are 1500Hz and 1600Hz respectively. The 3dB bandwidth of the cross-power spectrum density is calculated, and the frequency bandwidth is 200Hz. Since the main frequencies differ by 100Hz and the bandwidth is narrow, the average value of the two main frequencies, 1550Hz, is selected as the local oscillator frequency. The echo signal and the reference signal are multiplied by a 1550Hz sine signal by a direct mixing method to obtain a mixed signal.

[0068] The mixed signal is divided into 100 segments, each with a length of 1000 samples. The mean, variance and skewness of each signal segment are calculated to obtain a 100×3 time domain statistical feature matrix. The stability threshold is set to 10% of the variance, and the Gaussian threshold is set to the absolute value of the skewness less than 0.5. It is judged whether the variance of each signal segment is less than the stability threshold and whether the absolute value of the skewness is less than the Gaussian threshold. For the stable Gaussian signal segment that meets the conditions, a 10th-order Butterworth low-pass filter with a cutoff frequency of 500Hz is designed. For the non-stationary or non-Gaussian signal segment, an adaptive Wiener filter is used. The filter order (5-15 orders) and the noise variance (1%-10% of the variance) are adaptively adjusted according to the variance and skewness of each signal segment. The filtering process is repeated for all signal segments to obtain the difference frequency signal.

[0069] Select wavelet basis function, perform 4-layer wavelet decomposition on the difference frequency signal, calculate the mean and standard deviation of wavelet coefficients on each subband, set the empirical factor to 3, set the soft threshold according to 3 times the standard deviation, apply the soft threshold function to the wavelet coefficients of each subband for denoising, use wavelet reconstruction to obtain the denoised difference frequency signal, use the periodogram method to estimate the power spectrum of the denoised difference frequency signal, and obtain the initial Doppler spectrum. Assume that the wind speed obeys a mixture of 3 Gaussian distributions, initialize the Gaussian mixture model parameters, use the expectation maximization algorithm to train the Gaussian mixture model, obtain the spectrum residual, repeatedly update the Gaussian mixture model 5 times to reach the convergence condition, find the mean and standard deviation corresponding to the wind speed peak in the converged Gaussian mixture model, and obtain the background suppression signal;

[0070] The background suppression signal is subjected to short-time Fourier transform, and a Hamming window is set with a window length of 500 samples and an overlap rate of 75%. The mean, variance, skewness and kurtosis of the time-frequency signal in the time and frequency dimensions are calculated. According to the statistical characteristics, the Kaiser window is adaptively selected and the β parameter is set to 5, the window length is 1000 samples, and the spectral resolution is 1 Hz. The Kaiser window is applied to the time-frequency signal for spectrum estimation to obtain the final Doppler spectrum and time-frequency spectrum.

[0071] In this embodiment, by preprocessing the echo signal and the reference signal, estimating the main frequency and adaptively mixing, the DC component and trend term interference in the signal can be eliminated, and the optimal mixing mode can be adaptively selected according to the signal characteristics to improve the signal-to-noise ratio and spectrum resolution of the signal. By adopting a method combining time-domain segmented statistical feature extraction and adaptive filtering, the filter type and parameters can be adaptively selected according to the stationarity and Gaussianity of the signal, and non-stationary and non-Gaussian noise can be effectively removed to improve the quality of the difference frequency signal. By using a Gaussian mixture model and an iterative optimization algorithm, the background noise and wind speed signal components in the Doppler spectrum can be adaptively estimated, and the model can be updated by the spectrum residual to achieve adaptive extraction of wind speed peaks and background suppression, thereby improving the accuracy and robustness of wind speed estimation. By adopting an adaptive time-frequency analysis method, the type, length and resolution of the spectrum analysis window can be adaptively selected according to the statistical characteristics of the time-frequency signal, and both time and frequency resolution can be taken into account to obtain high-quality Doppler spectrum and time-frequency spectrum, providing rich information for wind speed estimation. In summary, this embodiment can effectively improve the accuracy, resolution and adaptability of wind speed estimation, overcome the limitations of traditional methods in complex wind field environments, and has important application value.

[0072] In an optional embodiment,

[0073] The probability density function corresponding to the Gaussian mixture model is shown in the following formula:

[0074]

[0075] Where x represents the observation data vector, K represents the total number of Gaussian components in the Gaussian mixture model, k represents the index of the Gaussian component, D represents the dimension of the observation data vector, and α k represents the mixing coefficient, μ k represents the mean vector of the kth Gaussian component, ∑ k represents the covariance matrix of the kth Gaussian component, and T represents the transpose.

[0076] In this embodiment, the Gaussian mixture model can adapt to different types of Doppler spectrum distributions by adjusting parameters such as the number, mean, covariance and weight of Gaussian components, and has strong flexibility and adaptability. By modeling the Doppler spectrum with the Gaussian mixture model, the probability distribution of the wind speed signal and the background noise can be adaptively estimated to improve the purity of the wind speed signal. The parameters of the Gaussian mixture model can be estimated and optimized by the expectation-maximization algorithm, which can improve the fitting accuracy and convergence speed of the model. In summary, this embodiment improves the background noise suppression effect and realizes the robust extraction of wind speed peaks through iterative optimization and spectrum residual update strategies, thereby improving the wind speed estimation performance of the wind measuring lidar, making full use of the statistical characteristics of the Doppler spectrum, and has a strong theoretical basis and practical value.

[0077] S2. extracting Doppler shift, spectrum width and signal-to-noise ratio from the Doppler spectrum, estimating wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight as wind field parameters by combining meteorological knowledge and machine learning algorithms, refining the wind field parameters through deep neural networks and adaptively compensating for atmospheric turbulence and aerosol extinction, identifying special flow patterns in the wind field through out-of-control correlation analysis and data mining based on the corrected wind field parameters, extracting high-level semantic features in the special flow patterns through convolutional neural networks, and obtaining refined wind field parameters;

[0078] The Doppler shift is a phenomenon in which the frequency of a wave changes due to the relative motion between the wave source and the observer. The spectrum width is the width of the signal distribution in the frequency domain, indicating the range of the signal frequency components. The signal-to-noise ratio is the ratio of signal strength to noise strength, which is used to measure the clarity and quality of the signal. The turbulence intensity is a parameter that measures the intensity of atmospheric turbulence disturbances, usually expressed by the ratio of the standard deviation of wind speed to the average wind speed. The wind field parameters include physical quantities such as wind speed, wind direction, and wind force level that describe the state of the wind field. The atmospheric turbulence refers to the irregular airflow movement in the atmosphere, which causes the fluctuation of physical quantities such as wind speed, temperature, and pressure in time and space. The aerosol extinction refers to the phenomenon that the aerosol particles absorb and scatter the light beam, resulting in the weakening of light intensity. The out-of-control correlation analysis is a data mining technology used to discover uncontrolled correlation relationships in a data set. The special flow pattern refers to an airflow pattern with special morphology or dynamic characteristics that appears in the wind field, such as vortex, wind shear, etc. The refined wind field parameters are wind field parameters obtained through high-precision measurement and advanced algorithm correction, and have higher spatial and temporal resolution.

[0079] In an optional embodiment,

[0080] Doppler frequency shift, spectrum width and signal-to-noise ratio are extracted from the Doppler spectrum, and the wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight are estimated as wind field parameters by combining meteorological knowledge and machine learning algorithms. The wind field parameters are refined and corrected by deep neural networks, and atmospheric turbulence and aerosol extinction are adaptively compensated. According to the corrected wind field parameters, special flow patterns in the wind field are identified by out-of-control correlation analysis and data mining, and high-level semantic features in the special flow patterns are extracted by convolutional neural networks. The refined wind field parameters include:

[0081] The Doppler spectrum is fitted by a dual-line fitting method to obtain two peak frequencies and calculate the peak frequency difference, and the Doppler frequency shift, spectrum width and signal-to-noise ratio are obtained based on the peak frequency difference and the radar operating wavelength;

[0082] The wind speed in the line of sight direction is determined by combining meteorological knowledge and the Doppler frequency shift, and the mapping relationship between the wind speed error and the turbulence intensity is determined based on the correlation between the wind speed and the refractive index turbulence intensity. Based on the mapping relationship, the wind speed error is predicted according to the observed wind speed and the radial wind speed is compensated to obtain the wind direction, turbulence intensity and atmospheric stability in the line of sight direction. The wind direction, turbulence intensity and atmospheric stability are used as wind field parameters, and the historical wind field parameters of the current area are obtained and added to a preset neural network model to train the neural network model. The trained model is used to perform fine correction on the true value of the wind field parameter, and the correction amount is dynamically adjusted according to the intensity of atmospheric turbulence and aerosol extinction in combination with the adaptive compensation mechanism to obtain the corrected wind field parameters;

[0083] Based on the corrected wind field parameters, the wind field is represented as a spatiotemporal three-dimensional data cube, wherein each position contains a wind field parameter corresponding to the current position, and based on the corrected wind field parameters, the similarity of each position in the spatiotemporal three-dimensional data cube is calculated, and the wind field is clustered by a density clustering algorithm, and the clustered regions are further summarized to obtain a classic flow pattern of the wind field, and the spatiotemporal evolution law between different flow patterns is determined by an association rule mining algorithm, and the special flow pattern is identified by comparing the spatiotemporal evolution law with the classic flow pattern of the wind field;

[0084] The area corresponding to the special flow pattern is represented as a two-dimensional image, in which rows represent heights and columns represent time. Features of the two-dimensional image are extracted by a pre-set convolutional neural network to obtain high-level semantic features. The high-level semantic features are fused with the corrected wind field parameters to obtain the refined wind field parameters.

[0085] The dual spectrum fitting method is a method for analyzing nonlinear components in a signal. The phase coupling and nonlinear characteristics in the signal are captured by calculating the dual spectrum (i.e., third-order spectrum) of the signal. The peak frequency difference refers to the difference between the two main frequency components in the signal. The radar operating wavelength refers to the wavelength of the signal emitted by the radar system. The refractive index turbulence intensity describes the degree of random change of the refractive index in the atmosphere, which is caused by atmospheric turbulence. The radial wind speed refers to the wind speed component along the radar observation direction. The refined correction refers to detailed adjustment of the model or data to improve the accuracy of prediction or measurement. The spatiotemporal three-dimensional data cube refers to a cube structure with data in both spatial and temporal dimensions, which is used to represent and analyze three-dimensional spatial data that changes with time, such as the spatial distribution and temporal changes of wind field data. The density clustering algorithm is an unsupervised learning algorithm used to discover dense areas in the data and group these areas. The wind field classic flow mode refers to the common flow structure in the wind field, such as high pressure system, low pressure system, cyclone, anticyclone, etc.

[0086] Apply the dual-line fitting method to the Doppler spectrum to obtain two peak frequencies, calculate the peak frequency difference, calculate the Doppler frequency shift, spectrum width and signal-to-noise ratio according to the radar working wavelength, calculate the line-of-sight wind speed according to the Doppler frequency shift and radar wavelength, establish a mapping relationship between wind speed error and turbulence intensity based on the correlation between wind speed and refractive index turbulence intensity, predict the wind speed error based on the observed wind speed, compensate the radial wind speed, and obtain the corrected line-of-sight wind speed. Estimate the wind direction, turbulence intensity and atmospheric stability in the line-of-sight direction based on the corrected wind speed as the initial wind field parameters;

[0087] Obtain historical wind field parameter data for the current area, add it to the pre-set neural network model for training, input the estimated wind field parameters into the trained neural network model, make fine corrections to the wind field parameters, design an adaptive compensation mechanism, and dynamically adjust the correction amount of the wind field parameters according to the intensity of atmospheric turbulence and aerosol extinction to obtain the corrected wind field parameters;

[0088] The corrected wind field parameters are represented as a spatiotemporal three-dimensional data cube, each position contains the corresponding wind field parameters, the similarity of each position in the data cube is calculated, the wind field is clustered by a density clustering algorithm, the clustered areas are analyzed and summarized, and the classic flow pattern of the wind field is obtained. The association rule mining algorithm is used to determine the spatiotemporal evolution law between different flow patterns, and the spatiotemporal evolution law is compared with the classic flow pattern to identify the special flow pattern. The area corresponding to the special flow pattern is represented as a two-dimensional image, with the row representing the height and the column representing the time. The pre-set convolutional neural network is used to extract features from the two-dimensional image to obtain high-level semantic features, and the extracted semantic features are fused with the corrected wind field parameters to obtain refined wind field parameters.

[0089] For example, assuming that a period of wind laser radar data is acquired in a certain area, the Doppler spectrum at one moment is processed, and dual-line fitting is applied to the Doppler spectrum to obtain peak frequencies f1=1.2MHz, f2=1.0MHz, and radar working wavelength λ=1550nm. The Doppler frequency shift, spectrum width, signal-to-noise ratio, and wind speed in the line of sight direction are calculated. Assuming that the refractive index turbulence intensity Cn^2 in the area is Cn^2=10^(-15)m^(-2 / 3), the wind speed error is predicted to be 1.5m / s through the pre-established mapping relationship, and the radial wind speed is compensated. The wind direction in the line of sight direction is estimated to be 270°, the turbulence intensity is 0.1, and the atmospheric stability is -2 (stable).

[0090] The historical wind field parameters of the area in the past week were obtained for training the neural network model. The currently estimated wind field parameters were input into the model for correction, and the corrected wind direction was 272°, the corrected turbulence intensity was 0.12, and the corrected atmospheric stability was -1.8. According to the current intensity of atmospheric turbulence and aerosol extinction, the correction amount was fine-tuned through the adaptive compensation mechanism, and finally the corrected wind direction was 273°, the corrected turbulence intensity was 0.11, and the corrected atmospheric stability was -1.9. The spatiotemporal data cube of the wind field in the area in the past week was constructed, and three classic flow modes were obtained through density clustering: southwest wind, northerly wind and calm wind. The association rule mining found that the southwest wind mode often occurs in the afternoon on sunny days, the northerly wind often occurs on rainy days, and the calm wind often occurs at night. The current wind field was compared with the classic mode and identified as belonging to the southwest wind mode, but the wind speed was obviously too large, so it was determined to be a special mode.

[0091] The area corresponding to the special mode is represented as a 72x24 image (covering 72 hours and an altitude range of 24km). The image features are extracted using a convolutional neural network. The extracted semantic features are fused with the corrected wind field parameters to obtain refined wind field parameters with a wind direction of 274°, a turbulence intensity of 0.12, and an atmospheric stability of -1.7.

[0092] In this embodiment, the Doppler spectrum is processed by dual-line fitting, combined with wind speed error compensation and adaptive correction strategies, the influence of atmospheric turbulence and aerosol extinction on wind speed estimation can be effectively overcome, and more accurate and stable wind field parameters can be obtained. Combined with association rule mining, the spatiotemporal evolution laws between different modes can be revealed, and intelligent recognition and prediction of wind field modes can be realized. The semantic features of wind field areas are extracted by combining convolutional neural networks, which can better characterize the spatiotemporal dynamics of special modes and provide a basis for extreme weather monitoring and early warning. Machine learning and pattern recognition technology are fully utilized, and wind field parameters are adaptively corrected by training neural network models. Wind field modes are mined by density clustering and association rules, and wind field area features are extracted by convolutional neural networks. Intelligent and automated wind field analysis is realized, and the influence of manual intervention and subjective judgment is reduced. In summary, this embodiment improves the intelligence level of wind field analysis, provides important support for disastrous weather warning, aviation safety and atmospheric science research, and has good innovation and practical value.

[0093] In an optional embodiment,

[0094] Based on the corrected wind field parameters, the similarity of each position in the spatiotemporal three-dimensional data cube is calculated as shown in the following formula:

[0095]

[0096] Where DIS represents the wind field parameter x i and x j The similarity between them, θ represents the mean vector of wind field parameters, n represents the number of samples, (x g -θ) represents the sample x g The difference vector relative to the mean vector, g represents the sample index, (x i -x j ) represents the difference vector between the wind field parameter vectors at two locations.

[0097] In this embodiment, by calculating the difference vector between the wind field parameter vector and the mean vector and normalizing it with the covariance matrix, the similarity of wind field parameters at different locations can be effectively measured, and the distribution characteristics of the wind field parameters are taken into account. It has better discrimination for differences in parameters such as wind direction, wind speed, and turbulence intensity. By normalizing the covariance matrix of the difference vector between the wind field parameter vector and the mean vector, the influence of the differences in dimensions and scales of different wind field parameters can be eliminated, the rationality and reliability of the similarity measurement can be improved, and it can help identify the internal structure and pattern of the wind field, and obtain a more reasonable and explainable The clustering results can discover the evolution path of wind field patterns in time and space by setting appropriate similarity thresholds, extract meaningful pattern conversion rules, and provide support for wind field prediction and trend analysis. In summary, this embodiment fully considers the statistical distribution characteristics of wind field parameters, eliminates the dimensional and scale differences of different physical quantities, improves the effects of wind field clustering, pattern recognition, and evolution law mining, enhances the recognition ability of special patterns, and facilitates the fusion of multi-source heterogeneous data. It is of great significance to improve the accuracy, stability, and interpretability of wind field analysis, and is an effective data-driven wind field pattern analysis technology.

[0098] S3. Based on the refined wind field parameters, a global wind field model is constructed. According to the global wind field model, the adaptive scanning optimization is converted into a Markov decision process. The action space and state space are determined according to the environmental state and task requirements. Meanwhile, a corresponding reward function is set. The Markov decision process is solved in combination with a reinforcement learning algorithm to obtain an optimal scanning strategy. Based on the optimal scanning strategy, the wind field state is predicted through a long short-term memory network to generate a state prediction report.

[0099] The global wind field model is a mathematical model used to describe and predict the distribution of wind speed and wind direction in a larger range. The Markov decision process is a mathematical framework for decision-making problems. The action space is a component of the Markov decision process, representing the set of all possible actions that can be taken in each state. The state space is another component of the Markov decision process, representing the set of all possible states of the system. The reward function is used to define the immediate feedback or reward obtained by taking an action in a specific state. The reinforcement learning algorithm is used to solve the problem of optimizing the decision-making process by interacting with the environment. The optimal scanning strategy involves selecting the most effective scanning or observation method to achieve the desired goal.

[0100] In an optional embodiment,

[0101] Based on the refined wind field parameters, a global wind field model is constructed. According to the global wind field model, the adaptive scanning optimization is converted into a Markov decision process. The action space and state space are determined according to the environmental state and task requirements. At the same time, the corresponding reward function is set. The Markov decision process is solved in combination with the reinforcement learning algorithm to obtain the optimal scanning strategy. Based on the optimal scanning strategy, the wind field state is predicted through the long short-term memory network, and the state prediction report is generated, including:

[0102] Based on the refined wind field parameters, the wind field is represented as a four-dimensional tensor in combination with a spatiotemporal wind field representation method. Based on the four-dimensional tensor, each spatial dimension and time dimension is linearly interpolated in combination with a trilinear interpolation method to obtain a continuous wind field representation. Each four-dimensional tensor in the continuous wind field representation is reshaped into a two-dimensional matrix, wherein each row corresponds to a spatiotemporal position, and each column corresponds to a refined wind field parameter. The two-dimensional matrix is ​​decentralized and the covariance matrix corresponding to the two-dimensional matrix is ​​calculated. The covariance matrix is ​​eigenvalue decomposed to obtain an eigenvector matrix and an eigenvalue vector. The eigenvalue vector is sorted in descending order, and the eigenvectors corresponding to the eigenvalues ​​in the top 10% are selected to form a principal component matrix. The refined wind field parameters are mapped to the principal component space corresponding to the principal component matrix to obtain the global wind field model.

[0103] Based on the global wind field model, the adaptive scanning problem is modeled as a Markov decision process, scanning tasks and meteorological data are obtained, the wind field is divided into grids, a state vector is generated based on the wind field state corresponding to each grid and the working state of the radar, the state vector is combined to obtain a state space, the scanning action of the radar in the current state is obtained and the pulse repetition frequency is adjusted to generate an action space, and the state transition probability and the corresponding reward function are determined;

[0104] Based on the constructed Markov decision, the objective function is constructed through Q learning in the reinforcement learning algorithm, the optimal scanning strategy is solved based on the objective function and the state transition probability, and the optimal scanning strategy is applied to the radar scanning control. At each decision moment, the radar selects the scanning action and executes the control instruction according to the current state in combination with the optimal scanning strategy to obtain real-time wind field data;

[0105] A wind farm sequence is constructed for the real-time wind farm data, and the wind farm sequence is added as input data to the long short-term memory network to generate a predicted state. The predicted state is compared with the true label corresponding to the wind farm sequence, and the prediction loss is calculated and the hyperparameters of the long short-term memory network are updated according to the prediction loss. The update is repeated until the model converges, and the wind farm sequence is predicted again using the updated long short-term memory network to obtain the predicted wind farm state, and the state prediction report is generated.

[0106] The four-dimensional tensor is a data structure with four dimensions. The spatiotemporal wind field representation method is used to describe the distribution of the wind field in space and time. The trilinear interpolation is a method for estimating data values ​​in three-dimensional space. The data value at any three-dimensional coordinate point is estimated by linearly interpolating the eight corner points of the three-dimensional data grid. The continuous wind field representation method is used to convert wind field data from discrete points into a continuous function representation. The decentralization refers to distributing decisions, controls or calculations to multiple nodes or systems instead of concentrating them on a central node. The eigenvalue decomposition is a matrix decomposition technique that decomposes a matrix into the product of eigenvectors and eigenvalues. The principal component matrix is ​​a matrix obtained by principal component analysis (PCA), which contains the main components or features of the data. The principal component space is a space defined by the principal component matrix, which is used to represent the main feature directions of the data. The real-time wind field data refers to wind speed and direction data collected in real time by sensors or radar systems. The predicted loss value refers to the error measure between the actual value and the predicted value in the prediction model.

[0107] Using the refined wind field parameters obtained above, the wind field is represented as a four-dimensional tensor in combination with the spatiotemporal wind field representation method. The dimensions of the four-dimensional tensor are three spatial dimensions (longitude, latitude, and altitude) and one temporal dimension. The trilinear interpolation method is used to perform linear interpolation on each spatial dimension and temporal dimension to obtain a continuous wind field representation. Each four-dimensional tensor in the continuous wind field representation is reshaped into a two-dimensional matrix. Each row of the matrix corresponds to a spatiotemporal position, and each column corresponds to a refined wind field parameter, such as wind speed, wind direction, turbulence intensity, etc. Decentralization is performed, that is, the mean of each parameter is subtracted to eliminate the influence of different parameter magnitudes, and the covariance matrix of the matrix is ​​calculated. The covariance matrix is ​​decomposed by eigenvalue to obtain an eigenvector matrix and an eigenvalue vector. The eigenvalue vectors are arranged in descending order, and the eigenvectors corresponding to the eigenvalues ​​in the top 10% are selected to form a principal component matrix. The refined wind field parameters are mapped to the principal component space corresponding to the principal component matrix to obtain a low-dimensional global wind field model.

[0108] Obtain scanning tasks and meteorological data, divide the wind field into grids, generate a state vector for each grid according to its corresponding wind field state (such as average wind speed, turbulence intensity, etc.) and the working state of the radar (such as current scanning angle, pulse repetition frequency, etc.), combine the state vectors of all grids to form a state space, define the optional scanning actions of the radar in the current state (such as adjusting the scanning angle, changing the pulse repetition frequency, etc.), generate an action space, determine the transition probability between states according to the dynamic characteristics of the wind field and the impact of radar scanning on wind field observation, and design a reward function to evaluate the effect of each state-action combination, such as observation accuracy, scanning coverage, etc.;

[0109] Using the Q-learning algorithm in reinforcement learning, an objective function is constructed to estimate the long-term cumulative reward of each state-action combination. The update rule of the objective function is based on the Bellman equation, and it is iteratively optimized by combining the state transition probability and the immediate reward. By continuously interacting with the environment and updating the Q value, the optimal scanning strategy is finally obtained. At each decision moment, the radar selects the next scanning action based on the current state using the learned optimal strategy and executes corresponding control instructions, such as adjusting the scanning angle, transmitting pulses, etc., to obtain real-time wind field data;

[0110] Serialize the wind farm data acquired in real time to construct a wind farm time series. Use the wind farm series as input and send it to a long short-term memory network for training and prediction. Compare the predicted wind farm state with the actual wind farm label and calculate the prediction loss. Based on the prediction loss, use the back propagation algorithm to update the weights and other hyperparameters of the long short-term memory network and continuously optimize the model. Repeat this process until the model converges. Use the trained long short-term memory network model to predict the new wind farm series, generate the predicted wind farm state in the future, and generate the corresponding state prediction report.

[0111] For example, assuming that a wind laser radar is deployed in a certain area, the wind field in the area needs to be adaptively scanned and predicted. The refined wind direction, wind speed, turbulence intensity and other parameters obtained above are used to divide the area into a 100x100x20 three-dimensional grid, with a time span of the past hour and a time resolution of 1 minute. A 100x100x20x60 four-dimensional wind field tensor is constructed, and trilinear interpolation is performed on the tensor to obtain a continuous wind field representation. The four-dimensional tensor is reshaped into a 12000000x10 two-dimensional matrix, where each row represents a spatiotemporal position and each column represents a wind field parameter. The matrix is ​​decentralized, the covariance matrix is ​​calculated and eigenvalue decomposition is performed, and the eigenvectors corresponding to the first 10 largest eigenvalues ​​are selected to form a principal component matrix, and the wind field parameters are mapped to the principal component space to obtain a global wind field model after dimensionality reduction.

[0112] The wind field is divided into 10x10 grids. The state of each grid includes the average wind speed, wind direction, turbulence intensity, and the current scanning angle, elevation angle, pulse repetition frequency, etc. of the radar, with a total of 8 dimensions. The action space includes the adjustment of the scanning angle in the horizontal and vertical directions (±1°, ±2°, ±5°), as well as the adjustment of the pulse repetition frequency (±100Hz, ±500Hz), with a total of 13 actions. The state transition probability is set so that the probability of the grid wind field state scanned by the radar to transfer to the direction of high wind speed and high turbulence is greater. The reward function comprehensively considers the scanning coverage rate and observation accuracy, and gives higher weight to the coverage of high-risk areas.

[0113] Apply the Q learning algorithm, initialize the Q value table, set the learning rate α = 0.1, the discount factor γ = 0.9, and the exploration probability ε = 0.1. At each decision moment, the radar selects the action with the largest Q value with a probability of 1-ε, and randomly explores with a probability of ε. After executing the action, the Q value is updated according to the observed wind field state and reward. The update is repeated 10,000 times, and each update contains 60 time steps, corresponding to a 1-hour scanning process. Finally, a converged Q value table is obtained, that is, the optimal scanning strategy.

[0114] The wind field state sequence (10x10x60) of the most recent hour is input into a LSTM network with 3 layers and 128 neurons in each layer. The wind field state of the previous 59 time steps is used as input and the wind field state of the 60th time step is used as the label. The LSTM network is trained to predict the wind field state at the next moment. After each prediction, the prediction loss is calculated and the Adam optimizer is used to update the network parameters. The learning rate is set to 0.001 and the training is repeated 100 times to obtain the wind field prediction model. The trained model is used to predict the wind field state in the next hour and a wind field prediction report is generated. The report shows that in the next hour, gusts of more than 10m / s may occur in the southwest direction and the turbulence intensity in the corresponding area will also increase significantly, requiring strengthened monitoring.

[0115] In this embodiment, by performing four-dimensional tensor representation and trilinear interpolation on the refined wind field parameters, a continuous wind field distribution in time and space can be obtained, the detailed structure and dynamic changes of the wind field can be captured, and high-quality input data can be provided for subsequent scanning strategy optimization and prediction analysis. The main variation mode of the wind field data is retained, and noise and redundant information are filtered out, which is helpful to improve the computational efficiency and generalization ability of the subsequent decision-making process. It can find the optimal balance between observation accuracy and scanning efficiency, dynamically adapt to the changing characteristics of the wind field, and continuously improve the accuracy of wind field prediction by assimilating radar observation data in real time and using historical data for model training and updating. The adaptive scanning strategy can focus on scanning and analyzing high-risk areas such as strong wind shear and high turbulence according to the setting of the reward function. Combined with the wind field prediction model, signs of extreme weather can be discovered early, the warning time can be extended, and precious time can be won for disaster prevention and emergency response. In summary, this embodiment can significantly improve the spatiotemporal resolution, accuracy and timeliness of wind field perception and warning, optimize the utilization efficiency of meteorological sensor resources, and provide important support for extreme weather monitoring, disaster prevention and weather and climate research.

[0116] In an optional embodiment,

[0117] The objective function constructed by Q learning in the reinforcement learning algorithm is shown in the following formula:

[0118]

[0119] Among them, J(Q) represents the objective function value, E represents the expectation, r represents the immediate reward, γ represents the discount factor, and s' represents the next state after executing action a. represents the maximum expected cumulative reward after selecting the optimal action a' in state s', and Q(s, a) represents the expected cumulative reward after executing action a in state s.

[0120] In this embodiment, when optimizing the strategy, both the immediate effect of the current action and the long-term benefits of the future state are taken into consideration, thereby achieving a globally optimal scanning strategy. It can process continuous state space and achieve generalization through function approximation, so that the strategy can make reasonable decisions for unseen states. When the wind farm environment changes, such as new meteorological conditions or scanning task requirements, the algorithm can adjust the strategy in real time to adapt to the new situation. In summary, this embodiment provides a flexible, efficient, and intelligent solution for wind farm perception, which is expected to significantly improve the performance and practical value of wind farm perception.

[0121] Figure 2 FIG. 1 is a schematic diagram of the structure of a laser radar signal processing and wind speed estimation system based on coherent detection according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0122] The first unit is used to obtain the echo signal and the reference signal of the wind measurement laser radar, adaptively select the mixing mode and the filtering parameters according to the statistical characteristics and noise level of the echo signal and the reference signal, mix and filter the echo signal and the reference signal, generate a difference frequency signal, and perform noise reduction and background suppression through an adaptive threshold and an iterative optimization algorithm to obtain difference frequency signal data, adaptively select the window function and resolution of spectrum analysis based on the time-frequency characteristics of the difference frequency signal data, perform time-frequency analysis on the difference frequency signal data through short-time Fourier transform, and generate a Doppler spectrum and a time-frequency spectrum;

[0123] The second unit is used to extract Doppler frequency shift, spectrum width and signal-to-noise ratio from the Doppler spectrum, estimate the wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight as wind field parameters by combining meteorological knowledge and machine learning algorithms, and perform fine correction on the wind field parameters and adaptively compensate for atmospheric turbulence and aerosol extinction through deep neural networks. According to the corrected wind field parameters, special flow patterns in the wind field are identified through out-of-control correlation analysis and data mining, and high-level semantic features in the special flow patterns are extracted through convolutional neural networks to obtain refined wind field parameters;

[0124] The third unit is used to build a global wind field model based on the refined wind field parameters, convert the adaptive scanning optimization into a Markov decision process according to the global wind field model, determine the action space and state space according to the environmental state and task requirements, and set the corresponding reward function. The Markov decision process is solved in combination with the reinforcement learning algorithm to obtain the optimal scanning strategy. Based on the optimal scanning strategy, the wind field state is predicted through a long short-term memory network to generate a state prediction report.

[0125] According to a third aspect of the embodiments of the present invention,

[0126] An electronic device is provided, comprising:

[0127] processor;

[0128] a memory for storing processor-executable instructions;

[0129] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0130] A fourth aspect of the embodiments of the present invention is:

[0131] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0132] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A laser radar signal processing and wind speed estimation method based on coherent detection, characterized in that: include: Acquire the echo signal and the reference signal of the wind measuring laser radar, adaptively select the mixing mode and the filtering parameters according to the statistical characteristics and noise level of the echo signal and the reference signal, perform mixing and filtering on the echo signal and the reference signal, generate a difference frequency signal, and perform noise reduction and background suppression through an adaptive threshold and an iterative optimization algorithm to obtain difference frequency signal data, adaptively select the window function and resolution of spectrum analysis based on the time-frequency characteristics of the difference frequency signal data, perform time-frequency analysis on the difference frequency signal data through short-time Fourier transform, and generate a Doppler spectrum and a time-frequency spectrum; Extracting Doppler frequency shift, spectrum width and signal-to-noise ratio from the Doppler spectrum, estimating wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight as wind field parameters by combining meteorological knowledge and machine learning algorithms, finely correcting the wind field parameters through deep neural networks and adaptively compensating for atmospheric turbulence and aerosol extinction, identifying special flow patterns in the wind field through out-of-control correlation analysis and data mining based on the corrected wind field parameters, extracting high-level semantic features in the special flow patterns through convolutional neural networks, and obtaining refined wind field parameters; Based on the refined wind field parameters, a global wind field model is constructed, and the adaptive scanning optimization is converted into a Markov decision process according to the global wind field model. The action space and the state space are determined according to the environmental state and task requirements, and the corresponding reward function is set. The Markov decision process is solved in combination with the reinforcement learning algorithm to obtain the optimal scanning strategy. Based on the optimal scanning strategy, the wind field state is predicted through the long short-term memory network to generate a state prediction report; Based on the refined wind field parameters, a global wind field model is constructed. According to the global wind field model, the adaptive scanning optimization is converted into a Markov decision process. The action space and state space are determined according to the environmental state and task requirements. At the same time, the corresponding reward function is set. The Markov decision process is solved in combination with the reinforcement learning algorithm to obtain the optimal scanning strategy. Based on the optimal scanning strategy, the wind field state is predicted through the long short-term memory network, and the state prediction report is generated, including: Based on the refined wind field parameters, the wind field is represented as a four-dimensional tensor in combination with a spatiotemporal wind field representation method. Based on the four-dimensional tensor, each spatial dimension and time dimension is linearly interpolated in combination with a trilinear interpolation method to obtain a continuous wind field representation. Each four-dimensional tensor in the continuous wind field representation is reshaped into a two-dimensional matrix, wherein each row corresponds to a spatiotemporal position, and each column corresponds to a refined wind field parameter. The two-dimensional matrix is ​​decentralized and the covariance matrix corresponding to the two-dimensional matrix is ​​calculated. The covariance matrix is ​​eigenvalue decomposed to obtain an eigenvector matrix and an eigenvalue vector. The eigenvalue vector is sorted in descending order, and the eigenvectors corresponding to the eigenvalues ​​in the top 10% are selected to form a principal component matrix. The refined wind field parameters are mapped to the principal component space corresponding to the principal component matrix to obtain the global wind field model. Based on the global wind field model, the adaptive scanning problem is modeled as a Markov decision process, scanning tasks and meteorological data are obtained, the wind field is divided into grids, a state vector is generated based on the wind field state corresponding to each grid and the working state of the radar, the state vector is combined to obtain a state space, the scanning action of the radar in the current state is obtained and the pulse repetition frequency is adjusted to generate an action space, and the state transition probability and the corresponding reward function are determined; Based on the constructed Markov decision, the objective function is constructed through Q learning in the reinforcement learning algorithm, the optimal scanning strategy is solved based on the objective function and the state transition probability, and the optimal scanning strategy is applied to the radar scanning control. At each decision moment, the radar selects the scanning action and executes the control instruction according to the current state in combination with the optimal scanning strategy to obtain real-time wind field data; A wind farm sequence is constructed for the real-time wind farm data, and the wind farm sequence is added as input data to the long short-term memory network to generate a predicted state. The predicted state is compared with the true label corresponding to the wind farm sequence, and the prediction loss is calculated and the hyperparameters of the long short-term memory network are updated according to the prediction loss. The update is repeated until the model converges, and the wind farm sequence is predicted again using the updated long short-term memory network to obtain the predicted wind farm state, and the state prediction report is generated.

2. The method according to claim 1, characterized in that Acquire the echo signal and the reference signal of the wind laser radar, adaptively select the mixing mode and the filtering parameters according to the statistical characteristics and noise level of the echo signal and the reference signal, perform mixing and filtering on the echo signal and the reference signal, generate a difference frequency signal, and perform noise reduction and background suppression through an adaptive threshold and an iterative optimization algorithm to obtain difference frequency signal data, adaptively select the window function and resolution of spectrum analysis based on the time-frequency characteristics of the difference frequency signal data, perform time-frequency analysis on the difference frequency signal data through short-time Fourier transform, and generate a Doppler spectrum and a time-frequency spectrum, including: Acquire the echo signal and the reference signal of the wind measuring laser radar, pre-process the echo signal and the reference signal, remove the DC component and the trend term, calculate the power spectrum density and the cross-power spectrum density of the echo signal and the reference signal, obtain the frequency domain statistical characteristics, estimate the main frequency and the frequency bandwidth of the echo signal and the reference signal according to the peak position and width of the cross-power spectrum density, judge whether the main frequencies of the echo signal and the reference signal are close and whether the frequency bandwidth is narrow according to the preset judgment conditions, if the main frequencies are close and the frequency bandwidth is narrow, use the average value of the main frequencies of the echo signal and the reference signal as the local oscillator frequency for direct mixing, if the main frequencies are not close or the frequency bandwidth is wide, use the main frequency difference of the echo signal and the reference signal as the local oscillator frequency, mix through the superheterodyne receiving mixing mode, and obtain the mixed signal; The mixed signal is divided into a plurality of signal segments of equal length, the mean, variance and skewness are calculated for each signal segment, and the time domain statistical features are obtained by combining them, and the stationarity and Gaussianity of each signal segment are judged according to the statistics and the empirical threshold value, for the stationary Gaussian signal segment, the cutoff frequency is selected by a low-pass filter, and for the non-stationary or non-Gaussian signal segment, the filter order and the noise variance are adaptively determined according to the statistics by an adaptive Wiener filter, and the filtering is repeated until all signals are processed to obtain the difference frequency signal; The difference frequency signal is subjected to wavelet decomposition to generate sub-bands of different scales. On each sub-band, the mean and standard deviation of the wavelet coefficients are calculated to obtain the statistical characteristics of noise at different scales. According to the standard deviation and the empirical factor of the wavelet coefficients, the noise reduction threshold of each sub-band is adaptively set and the wavelet coefficients are subjected to soft threshold processing. The processed wavelet coefficients are subjected to wavelet reconstruction to obtain the difference frequency signal after noise reduction. The power spectrum of the difference frequency signal after noise reduction is estimated to generate an initial Doppler spectrum. The initial Doppler spectrum is modeled and matched filtered by a preset Gaussian mixture model to obtain a spectrum residual. The Gaussian mixture model is updated based on the spectrum residual until convergence. The wind speed peak position and width are extracted from the updated Gaussian mixture model to obtain a background suppression signal. The background suppression signal is subjected to short-time Fourier transform to obtain a time-frequency signal and the statistical characteristics of the time-frequency signal are calculated. The type, length and resolution of the spectrum analysis window are selected based on the statistical characteristics and applied to the time-frequency signal to obtain the Doppler spectrum and the time-frequency spectrum.

3. The method according to claim 2, characterized in that The probability density function corresponding to the Gaussian mixture model is shown in the following formula: Where x represents the observation data vector, K represents the total number of Gaussian components in the Gaussian mixture model, k represents the index of the Gaussian component, D represents the dimension of the observation data vector, and α k represents the mixing coefficient, μ k represents the mean vector of the kth Gaussian component, ∑ k represents the covariance matrix of the kth Gaussian component, and T represents the transpose.

4. The method according to claim 1, characterized in that Doppler frequency shift, spectrum width and signal-to-noise ratio are extracted from the Doppler spectrum, and the wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight are estimated as wind field parameters by combining meteorological knowledge and machine learning algorithms. The wind field parameters are refined and corrected by deep neural networks, and atmospheric turbulence and aerosol extinction are adaptively compensated. According to the corrected wind field parameters, special flow patterns in the wind field are identified by out-of-control correlation analysis and data mining, and high-level semantic features in the special flow patterns are extracted by convolutional neural networks. The refined wind field parameters include: The Doppler spectrum is fitted by a dual-line fitting method to obtain two peak frequencies and calculate the peak frequency difference, and the Doppler frequency shift, spectrum width and signal-to-noise ratio are obtained based on the peak frequency difference and the radar operating wavelength; The wind speed in the line of sight direction is determined by combining meteorological knowledge and the Doppler frequency shift, and the mapping relationship between the wind speed error and the turbulence intensity is determined based on the correlation between the wind speed and the refractive index turbulence intensity. Based on the mapping relationship, the wind speed error is predicted according to the observed wind speed and the radial wind speed is compensated to obtain the wind direction, turbulence intensity and atmospheric stability in the line of sight direction. The wind direction, turbulence intensity and atmospheric stability are used as wind field parameters, and the historical wind field parameters of the current area are obtained and added to a preset neural network model to train the neural network model. The trained model is used to perform fine correction on the true value of the wind field parameter, and the correction amount is dynamically adjusted according to the intensity of atmospheric turbulence and aerosol extinction in combination with the adaptive compensation mechanism to obtain the corrected wind field parameters; Based on the corrected wind field parameters, the wind field is represented as a spatiotemporal three-dimensional data cube, wherein each position contains a wind field parameter corresponding to the current position, and based on the corrected wind field parameters, the similarity of each position in the spatiotemporal three-dimensional data cube is calculated, and the wind field is clustered by a density clustering algorithm, and the clustered regions are further summarized to obtain a classic flow pattern of the wind field, and the spatiotemporal evolution law between different flow patterns is determined by an association rule mining algorithm, and the special flow pattern is identified by comparing the spatiotemporal evolution law with the classic flow pattern of the wind field; The area corresponding to the special flow pattern is represented as a two-dimensional image, in which rows represent heights and columns represent time. Features of the two-dimensional image are extracted through a preset convolutional neural network to obtain high-level semantic features. The high-level semantic features are fused with the corrected wind field parameters to obtain the refined wind field parameters.

5. The method according to claim 4, characterized in that Based on the corrected wind field parameters, the similarity of each position in the spatiotemporal three-dimensional data cube is calculated as shown in the following formula: Where DIS represents the wind field parameter x i and x j The similarity between them, θ represents the mean vector of wind field parameters, n represents the number of samples, (x g -θ) represents the sample x g The difference vector relative to the mean vector, g represents the sample index, (x i -x j ) represents the difference vector between the wind field parameter vectors at two locations.

6. The method according to claim 1, characterized in that The objective function constructed by Q learning in the reinforcement learning algorithm is shown in the following formula: Among them, J(Q) represents the objective function value, E represents the expectation, r represents the immediate reward, γ represents the discount factor, and s' represents the next state after executing action a. represents the maximum expected cumulative reward after selecting the optimal action a' in state s', and Q(s, a) represents the expected cumulative reward after executing action a in state s.

7. A laser radar signal processing and wind speed estimation system based on coherent detection, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain the echo signal and the reference signal of the wind measurement laser radar, adaptively select the mixing mode and the filtering parameters according to the statistical characteristics and noise level of the echo signal and the reference signal, mix and filter the echo signal and the reference signal, generate a difference frequency signal, and perform noise reduction and background suppression through an adaptive threshold and an iterative optimization algorithm to obtain difference frequency signal data, adaptively select the window function and resolution of spectrum analysis based on the time-frequency characteristics of the difference frequency signal data, perform time-frequency analysis on the difference frequency signal data through short-time Fourier transform, and generate a Doppler spectrum and a time-frequency spectrum; The second unit is used to extract Doppler frequency shift, spectrum width and signal-to-noise ratio from the Doppler spectrum, estimate the wind speed, wind direction, turbulence intensity and atmospheric stability in the line of sight as wind field parameters by combining meteorological knowledge and machine learning algorithms, and perform fine correction on the wind field parameters and adaptively compensate for atmospheric turbulence and aerosol extinction through deep neural networks. According to the corrected wind field parameters, special flow patterns in the wind field are identified through out-of-control correlation analysis and data mining, and high-level semantic features in the special flow patterns are extracted through convolutional neural networks to obtain refined wind field parameters; The third unit is used to build a global wind field model based on the refined wind field parameters, convert the adaptive scanning optimization into a Markov decision process according to the global wind field model, determine the action space and state space according to the environmental state and task requirements, and set the corresponding reward function. The Markov decision process is solved in combination with the reinforcement learning algorithm to obtain the optimal scanning strategy. Based on the optimal scanning strategy, the wind field state is predicted through a long short-term memory network to generate a state prediction report.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images