A digital enhancement method for measuring the maximum and minimum of a high-power field source

Abstract

The application discloses a kind of polar polar extreme measurement digital enhancement methods for high-power field source, belong to communication, detection and electromagnetic measurement field;First, for radar detection signal, wavelet function is selected to decompose signal layer by layer, and wavelet coefficient is obtained;Then, wavelet coefficient is eliminated and scaled, as far as possible to retain approximate coefficient and reduce detail coefficient, and layer by layer with the reconstruction filter coefficient of wavelet function is made convolution, and wavelet reconstruction obtains the signal y (n) after denoising;Again, singular value decomposition is carried out, and large signal L (n) is separated;The signal y (n) after denoising and the large signal L (n) separated are used as training set, and the weight and bias parameter of deep neural network are trained;Finally, the trained weight and bias parameter are loaded into deep neural network as post-compensator, and the nonlinear distortion signal after denoising is reasonably corrected by post-compensator.The application can eliminate the influence of receiving end nonlinearity and noise simultaneously, and greatly improve signal quality.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of communication, detection and electromagnetic measurement, and particularly relates to a digital enhancement method for measuring the maximum and minimum of a high-power field source. BACKGROUND

[0002] A digital receiving front end generally comprises nonlinear devices such as a low-noise amplifier, a high-suppression-ratio filter, a mixer and an analog-to-digital converter. When receiving a wideband signal, the nonlinear distortion caused by the cascading of the devices will lead to oscillation, so that the system is difficult to ensure a large SFDR. In addition, the noise at the receiving end destroys the signal in a significant way, and in order to perform further data analysis, the noise must be removed from the data.

[0003] A pulse signal is a typical radar detection signal. As a non-stationary random signal with low signal-to-noise ratio, the pulse signal is easily disturbed by noise in the actual acquisition and transmission process, which greatly affects the subsequent correction processing of the pulse signal. Therefore, before the nonlinear correction of the pulse signal, denoising is an important work.

[0004] Existing denoising methods include low-pass filtering denoising, singular value decomposition and wavelet decomposition denoising. Among them, wavelet decomposition denoising is a time-scale signal analysis method with multi-resolution characteristics, and has the ability to represent the characteristics of the signal in the time domain and the frequency domain. It is a time-frequency localization analysis method with a fixed window size, but the shape can be changed, and the time window and the frequency window can be changed. In the low-frequency part, it has high frequency resolution and low time resolution, and in the high-frequency part, it has high time resolution and low frequency resolution. Due to this characteristic, the wavelet transform has the ability of signal adaptation, and is more suitable for signals with mutations in the time domain. Theoretically, wavelet decomposition and reconstruction for pulse signal denoising will obtain good results.

[0005] The principle of wavelet decomposition is to use the Mallat algorithm to decompose the signal by order, and the algorithm decomposes the signal into an approximate component and a detail component at each scale. The approximate component represents the high scale of the signal, i.e. the low-frequency information, and the detail component represents the low scale of the signal, i.e. the high-frequency information. For a signal containing noise, the main energy of the noise component is concentrated in the detail component of the wavelet decomposition. Therefore, if a suitable threshold can be found after decomposition, the detail component containing noise can be reduced or eliminated on the premise of ensuring the smoothness of the signal, and then reconstruction can be performed, so that denoising can be completed.

[0006] The de-noised pulse signal only removes noise interference, but still has a problem of non-linear distortion, and the prior art attempts to optimize the analog circuit as much as possible to improve the linearization performance, but the performance improvement is difficult to achieve the desired effect due to the restriction of the existing process technology. Using a certain digital technology to reduce the non-linear problem of the receiver front end is a more common solution at the present stage.

[0007] The linearization technology widely used in the related field at present includes power back-off method, feedback technology, feedforward technology or pre-distortion technology, etc., wherein the digital pre-distortion (DPD) is a common digital compensation method at the transmitting end, and a large number of studies show that this method has an ideal compensation effect on the non-linearity caused by the power amplifier. The basic principle of the digital pre-distortion is to regard the power amplifier (PA) characteristic at the transmitting end as a non-linear model, fit the inverse model opposite to the non-linear characteristic of the PA through collecting data before and after distortion by a certain mathematical model, and place the pre-distortion device before the power amplifier (PA) device, so that the final output signal presents a linear characteristic.

[0008] It is easy to think that this DPD-based technology is followed in the compensation scheme of the receiver, the non-linearity shown by the receiver RF front end and ADC is regarded as a non-linear system similar to the PA, and a model playing the function of the pre-distortion device is tried to be constructed to ensure that the whole presents linearity. However, this scheme will face a problem: for the receiving end, the input signal is an analog signal, and an additional high-speed ADC needs to be added to convert the signal, which will not only increase the cost but also easily introduce new distortion. If the input signal can be approximately obtained without adding an additional ADC, the above shortcomings can be avoided and the DPD idea can be used for non-linear compensation. SUMMARY

[0009] In order to solve the problems of noise interference and non-linear distortion in the process of receiving radar detection signals represented by pulse signals, the present application proposes a digital enhancement method for maximum and minimum value measurement of high-power field sources, which can better guarantee the signal quality by wavelet de-noising and then combining singular value decomposition and a neural network model for post-compensation.

[0010] The specific steps of the digital enhancement method are as follows:

[0011] Step 1: For radar detection signals, select a 4th-order Daubechies wavelet function as a de-noising filter and set the wavelet decomposition layer number, and decompose the signal layer by layer to obtain wavelet coefficients.

[0012] The present application selects 9 layers of decomposition, and when each layer is decomposed, the known filter coefficients of the previous layer are respectively convolved with high and low pass filters and are decimated by point separation to obtain the detail coefficients of each layer approximating the original signal and the approximate coefficients containing noise components as wavelet coefficients.

[0013] Step two, select threshold and threshold function to eliminate and scale the wavelet coefficients, and try to retain the approximate coefficients and reduce the detail coefficients.

[0014] Common threshold estimation methods include Sqtwolog (fixed form) threshold criterion, Rigrsure (adaptive) threshold criterion, Heursure (heuristic) threshold criterion, and Minimaxi (maximin) threshold criterion.

[0015] Step three, the retained approximate coefficients and detail coefficients are convolved with the reconstruction filter coefficients of the wavelet function layer by layer to reconstruct the wavelet denoised signal y(n).

[0016] Wavelet reconstruction is the inverse process of wavelet decomposition, and the wavelet function reconstruction filter coefficients are known;

[0017] Step four, singular value decomposition is performed on the denoised signal y(n) to separate out the large signal L(n) as a component approximating the original signal before distortion.

[0018] For an m x n matrix X, singular value decomposition is defined as:

[0019] X = USV T

[0020] Where U is an n x n orthogonal left singular matrix, V is an m x m orthogonal right singular matrix, and S is an n x m diagonal matrix, and the non-zero elements on the diagonal are singular values.

[0021] The singular values in the diagonal matrix S are sorted in descending order, and the diagonal matrix composed of the first r singular values greater than 1 is multiplied by the corresponding original left / right singular matrix to approximately recover the large signal L(n).

[0022] Step five, input the wavelet denoised signal y(n) into a deep neural network, use the separated large signal L(n) as the target output, and train the weight and bias parameters of the deep neural network.

[0023] The deep neural network has multiple hidden layers and multiple outputs in the neurons of the hidden layers. For each neuron with n inputs and one output, the weighted sum is passed through a nonlinear activation function to establish a mapping between the input and output as follows:

[0024]

[0025] f(·) is an activation function, w i is the weight corresponding to the i-th input of the neuron, x i is the i-th input of the neuron, and b is the bias constant of the neuron.

[0026] The coefficients of the deep neural network are trained by using a DNN back propagation algorithm, initial values of network structure and network parameters, an activation function and a loss function are set, iteration is performed with the loss between the output of the deep neural network and the ideal output as the target, and finally the final values of the weight and bias parameters of the deep neural network are obtained.

[0027] Step six, loading the trained weight and bias parameters into the deep neural network as a post-compensator, for a new radar detection signal, the wavelet coefficients obtained after decomposition are eliminated and scaled, and the wavelet denoising is reconstructed and input into the post-compensator, and finally a signal with good performance is output.

[0028] The advantages of the present application are:

[0029] 1) A digital enhancement method for extreme value measurement of a high-power field source, which comprehensively considers the influence of noise and nonlinear distortion on the detection results when detecting radar detection signals represented by pulse signals, and can well eliminate noise and nonlinear distortion.

[0030] 2) A digital enhancement method for extreme value measurement of a high-power field source, the post-compensation method well solves the problem that the original signal cannot be obtained at the receiving end and it is difficult to train the compensation model.

[0031] 3) A digital enhancement method for extreme value measurement of a high-power field source, a neural network is used to simulate nonlinear behavior, which is more flexible than a traditional memory polynomial model. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 A post-compensation digital enhancement system diagram is established for the present application;

[0033] Figure 2 A flowchart of the digital enhancement method for extreme value measurement of a high-power field source is provided for the present application;

[0034] Figure 3 A wavelet decomposition schematic diagram is used in the present application;

[0035] Figure 4 A wavelet denoising flowchart is used in the present application;

[0036] Figure 5 A singular value separation flowchart is used in the present application. DETAILED DESCRIPTION

[0037] In order to facilitate those skilled in the art to understand and implement the present application, the present application is further described in detail below in combination with the drawings and examples.

[0038] The present application is a digital enhancement method for measuring the maximum and minimum of a high-power field source, as shown in the figure. Figure 1 As shown in the figure, the analog signal x(t) is input to the digital receiver front end, and after passing through the wideband RF front end and high-speed ADC, the nonlinear correction of wavelet denoising and digital post-compensation is combined to finally realize the output of the signal; wherein the post-compensation takes the large signal L(n) obtained by singular value decomposition and the distorted signal y(n) after denoising as the training set to train the DNN neural network, and inputs the trained network parameters into the post-compensation device.

[0039] The present application is a digital enhancement method of first wavelet denoising and then combining singular value decomposition and neural network model for post-compensation. Singular value decomposition is a mathematical method that separates signals only according to the size of signal power without any prior information of known input signals. For the scene to be solved, the energy of the nonlinear distortion component in the receiving end is usually smaller than that of the original signal, and the nonlinear distortion part and the original signal can be well separated by singular value decomposition. Among the many methods of analog device behavior models, neural networks are widely used due to their strong fitting ability. Designers do not have to explicitly consider all possible distortion reasons in the model, but only need to obtain a suitable training set to train a model that meets the requirements. Using neural network as a post-compensation device, the separated large signal and the distorted signal are used to simulate the nonlinear inverse model to complete the post-compensation of the distorted signal. The method first denoises the signal by wavelet to reduce noise interference, then singular value decomposition is used to obtain a component similar to the original signal, and the decomposed signal and the distorted signal are used to construct a training set for neural network training. Finally, the trained parameters are loaded into the post-compensation device to perform online real-time compensation on the receiving front-end signal.

[0040] The present application is a digital enhancement method for measuring radar detection signals represented by pulse signals. Without additional ADC, the digital pre-distortion idea is used for digital enhancement of the receiving end signal. As a post-compensation method, it can eliminate the influence of the receiving end nonlinearity and noise, and greatly improve the signal quality.

[0041] The digital enhancement method for measuring the maximum and minimum of a high-power field source, as shown in the figure, comprises the following specific steps: Figure 2

[0042] Step one, for radar detection signals, select a 4th order Daubechies wavelet function as a filter and set the wavelet decomposition layer number, decompose the signal layer by layer, and obtain the wavelet coefficient.

[0043] ​The selection of wavelet function and order affects the final denoising effect to some extent. When selecting the wavelet function, the regularity, vanishing moment, symmetry and support length should be considered. Among them, the regularity affects the smoothing effect after processing, which can reduce the influence of quantization error on the denoising result; the vanishing moment can be understood as the number of wavelet coefficients that are set to zero after decomposition, the larger the vanishing moment, the more noise components are removed, but at the same time, the original signal components are more likely to be lost; the symmetry means that the filter corresponding to the wavelet function has the characteristic of linear phase, which can effectively avoid phase distortion in the denoising process; the support length determines the influence range of the wavelet function under a certain scale and a certain translation, and is positively correlated with the vanishing moment.

[0044] After comprehensive consideration, the fourth-order Daubechies (db4) wavelet function is selected, which has the above characteristics and can act as a filter when wavelet denoising.

[0045] The selection of decomposition layers during decomposition will affect the final denoising effect. The more the decomposition layers, the more noise components are removed, but too large decomposition layers will affect the quality of the original signal recovered after reconstruction to some extent, so 9 layers of decomposition are considered.

[0046] After determining the type of wavelet function and the number of wavelet decomposition layers used in wavelet decomposition, the parameters are selected layer by layer. Since the wavelet decomposition filter coefficients corresponding to the selected wavelet are known, the approximate coefficients and high-low pass filters of the last layer are convolved and decimated at intervals during each layer of decomposition to obtain the detail coefficients of each layer of approximate original signal and the approximate coefficients containing noise components, that is, the wavelet coefficients.

[0047] Step two, select the threshold value and threshold function to eliminate and scale the wavelet coefficients, and try to retain the approximate coefficients and reduce the detail coefficients.

[0048] Noise signals are mainly contained in high-frequency details, and the values of approximate coefficients are mostly higher than those of detail coefficients. The noise components mixed in the original signal are divided by selecting the threshold value, and then each coefficient is quantized by the threshold value. Common threshold estimation methods include Sqtwolog (fixed form) threshold criterion, Rigrsure (adaptive) threshold criterion, Heursure (heuristic) threshold criterion, and Minimaxi (maximin) threshold criterion.

[0049] Considering that the Sqtwolog threshold criterion is simple to calculate and can process the full frequency band of the signal, the final threshold is determined by the Sqtwolog criterion;

[0050] After determining the threshold, the appropriate threshold function is selected to process the wavelet coefficients, which are divided into soft threshold and hard threshold methods. The hard threshold method is very direct, which directly sets the wavelet coefficients less than the threshold to 0 and keeps the original value of the coefficients greater than the threshold, which is easy to cause signal oscillation and is not conducive to the smoothness of the denoising result. The soft threshold reduces the part higher than the threshold according to certain rules, which ensures the continuity of the processed signal, so the soft threshold is used to process the coefficients.

[0051] Step three, reconstruct the coefficients processed in step two with the reconstruction filter coefficients of the wavelet function to obtain the wavelet denoised signal y(n).

[0052] Wavelet reconstruction can be regarded as the inverse process of wavelet decomposition, and the reconstruction filter coefficients of the wavelet function are also known. According to the last approximation coefficients and detail coefficients, layer by layer convolution is performed with the reconstruction filter coefficients of the wavelet function, and the denoised signal y(n) can be obtained after reconstruction.

[0053] Step four, singular value decomposition is performed on the denoised signal y(n) to separate out the large signal L(n) as a component approximating the original signal before distortion.

[0054] For the distorted signal affected by the nonlinear device at the receiving end, it can be considered that the main distortion components are generated by harmonics and intermodulation, and the energy of these distortion components is usually smaller than that of the original signal. Singular value decomposition is a process of effectively arranging the characteristics of data, which separates signals according to the size of signal power without knowing any prior information of the input signal. Therefore, the singular value decomposition method can be used to separate the large signal, and the separated large signal is considered as a component approximating the original signal before distortion.

[0055] Singular value decomposition is usually applied to matrices, and the singular value decomposition of an m x n matrix X is defined as:

[0056] X = USV T

[0057] Where U is an n x n orthogonal left singular matrix, V is an m x m orthogonal right singular matrix, and S is an n x m diagonal matrix, and the non-zero elements on the diagonal are singular values.

[0058] The singular value often corresponds to the important information hidden in the matrix, and the importance is positively correlated with the size of the singular value. Each matrix A can be represented as the sum of a series of "small matrices", and the singular value measures the weight of these "small matrices" for A. That is, the original matrix is approximated by the largest r singular values and the corresponding left and right singular vectors, and the large signal is obtained by adjusting the size of r.

[0059] The value of the decomposed diagonal matrix S can be obtained. According to experience, when the value of the r+1 singular value is less than 1, it is considered that the diagonal matrix composed of the first r singular values multiplied by the left singular matrix and the right singular matrix can approximately restore the large signal component.

[0060] Step five, using the large signal L(n) obtained after separation in step four and the denoised signal y(n) obtained in step three as a training set, the weight and bias parameters of the deep neural network are trained.

[0061] The basic idea of post-compensation is to regard the radio frequency receiving front end as a nonlinear system. If the inverse function of the system transfer function can be simulated and cascaded after it, the nonlinear distortion component generated by the post-compensation module can completely compensate for the original system distortion component, and the output of the cascaded two-stage nonlinear subsystem is still the original input signal.

[0062] In order to train such a model, the nonlinear distorted signal y(n) after wavelet denoising is used as input, and the large signal L(n) separated is used as target output to train the neural network.

[0063] The basis of the deep neural network (DNN) model comes from neurons. Each neuron has n inputs and an output. After weighted summation operation, the neuron passes through a specific nonlinear activation function, so that there is a certain mapping relationship between the input and the output:

[0064]

[0065] w i is the weight value corresponding to the i-th input of the neuron, f(·) is the activation function, and b is the bias constant of the neuron.

[0066] DNN is an extension based on neurons, which has multiple hidden layers and the neurons in the hidden layer can have multiple outputs. The DNN network coefficients are trained using the DNN back propagation algorithm. The initial value of the network structure and network parameters, the activation function and the loss function are set. The loss between the output of the DNN network and the ideal output is minimized as the target for iteration, and the final value of the weight and bias parameters of the deep neural network is obtained.

[0067] Step six, using the DNN network model with the same structure as the training and loading the trained weight and bias parameters into it as a post-compensator. The new distorted signal received by the radar is first denoised by the wavelet and then passed through the post-compensator to finally output a signal with better performance.

[0068] Embodiment:

[0069] Step one, using a fourth-order Daubechies (db4) wavelet function for wavelet decomposition, the wavelet decomposition layer is selected to be 9, and the wavelet coefficients are obtained.

[0070] The basic principle of wavelet decomposition is based on the fast algorithm of orthogonal wavelet transform (Mallat algorithm), and the essence is to decompose the signal in a series of different frequency spaces. The multi-resolution analysis of a signal at different scales can show the different frequency characteristics of a signal. Each layer of decomposition separates the signal into high-frequency and low-frequency two parts. The high-frequency component after decomposition is the part containing more noise components, which is called detail coefficient, while the low-frequency component after decomposition is the part closer to the original signal, which is called approximation coefficient. When decomposing layer by layer, the decomposition of the next layer is always based on the approximation coefficient of the last layer.

[0071] The schematic diagram of wavelet decomposition is shown in Figure 3 .

[0072] Wherein, S is the signal to be processed, CD1, CD2, CD3 are detail coefficients, and CA1, CA2, CA3 are approximation coefficients.

[0073] The algorithm of wavelet decomposition can be expressed as:

[0074]

[0075]

[0076] Wherein, CA j (k) and CD j (k) represent the kth wavelet decomposition approximation coefficient and detail coefficient at the jth layer of decomposition, respectively, and Ld and Hd are the low-pass filter coefficient and high-pass filter coefficient corresponding to the wavelet function. Z represents the signal length.

[0077] The wavelet decomposition filter coefficient corresponding to db4 wavelet is known. When decomposing each layer, the approximation coefficient of the last layer is convolved with the low-pass filter and the high-pass filter respectively, and then decimated by point, to obtain the detail coefficient of each layer and the final approximation coefficient.

[0078] Step 2: Calculate the threshold according to the Sqtwolog criterion and process the wavelet coefficients according to the soft threshold function.

[0079] The expression of the threshold value th calculated according to the Sqtwolog criterion is:

[0080]

[0081] σ = mid / 0.6745

[0082] Wherein, σ is the standard deviation of noise, mid represents the median of the coefficients after wavelet decomposition, and N is the length of each coefficient.

[0083] Then, the coefficients are quantized according to a soft threshold function, and the threshold rule is as follows:

[0084]

[0085] where w j,k represents the wavelet coefficient after wavelet decomposition.

[0086] Step three, according to the quantized approximation coefficients and detail coefficients, the signal is reconstructed by wavelet.

[0087] As shown in Figure 4 , for the original signal, first determine the wavelet function and the number of decomposition layers, then decompose the wavelet multi-scale, then process the decomposition coefficients of each layer according to the threshold rule, and finally realize wavelet reconstruction. At each layer of wavelet reconstruction, the data needs to be up-sampled first, then the detail component is convolved with the high-frequency reconstruction filter, and the approximation component is convolved with the low-frequency reconstruction filter, and the next layer approximation coefficient is obtained by adding them. Repeat the above process until the 9-layer reconstruction is completed.

[0088] Step four, singular value decomposition is performed on the signal.

[0089] As shown in Figure 5 , first, the denoised signal y(n) is reconstructed to obtain the matrix B to be decomposed; then, the matrix B is singular value decomposed;

[0090] Specifically, the continuous N-point wavelet denoised signal Y N =[y(n),y(n+1),…,y(n+N-1)] is embedded in a p×q matrix row by row:

[0091]

[0092] The to-be-compensated matrix B p×q is SVDed:

[0093] B p×q =U p×v S v×v V T v×q

[0094] where v=min(p,q), U and V are singular matrices, S v×v is a diagonal matrix, and the elements in the matrix are singular values of the matrix B.

[0095] Then, the first r large singular values are kept unchanged and the remaining v-r small singular values are set to zero to construct a new main diagonal matrix The new main diagonal matrix is multiplied back to the original left and right singular matrices to obtain a new signal matrix.

[0096] The new main diagonal matrix Substitute, we get:

[0097]

[0098] Finally, the new signal matrix is expanded into a one-dimensional sequence by connecting the end of the row, and the N-point large signal L(n) is obtained.

[0099] Step five, select the deep neural network structure as the model of the simulation of the nonlinear inverse characteristics, and train the model parameters by the DNN back propagation algorithm.

[0100] The designed DNN has three hidden layers, each layer has 15 neurons, the activation function is sigmoid function, and the mean square error is used as the loss function to update the parameters by gradient descent method. Assuming that the (l-1)th layer has m neurons, the output of the jth neuron in the lth layer is :

[0101]

[0102]

[0103] The loss function is represented as:

[0104]

[0105] Where, W L is an n×m matrix composed of the weight coefficient of the Lth layer output layer, x L-1 is an m×1 vector output by the L-1th layer, b L is an n×1 bias vector, L y is the ideal output, and f(z) is the sigmoid activation function.

[0106] According to the gradient descent method, the recursive relationship of the gradient of the lth layer is derived as:

[0107] δ l = (W l+1 ) T δ l+1 ·f′(W l x l-1 +b l )

[0108] The specific DNN back propagation algorithm is as follows:

[0109] (1) Initialize the weights W and bias parameters b of each layer of the network with a small random number;

[0110] (2) Input the denoised data Y N to the network, calculate the output of each layer of neurons to get the final output of the network.

[0111] (3) Calculate the loss function by the loss function, and get the output layer gradient;

[0112] (4) Calculate δ i,l according to the back propagation algorithm;

[0113] (5) Update the weight and bias coefficient of the lth layer:

[0114]

[0115]

[0116] (6) When the error value calculated by the loss function is less than the minimum error value set, it is considered that the accuracy requirement is met and the training can be ended, and the weight matrix W and the bias vector b are output.

[0117] Step six, construct a DNN with the same structure as the neural network used for training as a post-compensator, load the trained parameters into the model, and input the denoised signal to obtain the compensated signal.

Claims

1. A digital enhancement method for extreme value measurement of high-power field sources, characterized in that, The specific steps are as follows: Step 1: For radar detection signals, select the 4th-order Daubechies wavelet function as the denoising filter and set the wavelet decomposition level to decompose the signal layer by layer to obtain wavelet coefficients. Step 2: Select a threshold and threshold function to eliminate and scale the wavelet coefficients, retaining as many approximate coefficients as possible while reducing detail coefficients; Step 3: Convolve the processed approximation coefficients and detail coefficients with the reconstruction filter coefficients of the wavelet function layer by layer to reconstruct the wavelet-denoised signal y(t); Wavelet reconstruction is the inverse process of wavelet decomposition, and the wavelet function reconstruction filter coefficients are known; Step 4: Perform singular value decomposition on the denoised signal y(t) to separate the large signal L(t), which is used as a component that approximates the original signal before distortion. For an m×n matrix X, the singular value decomposition is defined as: in It is an n×n orthogonal left singular matrix. It is an m×m orthogonal right singular matrix. It is an n×m diagonal matrix, where the non-zero elements on the diagonal are singular values; Diagonal matrix The singular values ​​in the matrix are sorted from largest to smallest. The diagonal matrix formed by the first r singular values ​​with values ​​greater than 1 is multiplied by the corresponding original left / right intersection singular matrix to approximately recover the large signal L(t). Step 5: Input the wavelet-denoised signal y(t) into the deep neural network, and use the separated large signal L(t) as the target output to train the weights and bias parameters of the deep neural network. Deep neural networks have multiple hidden layers, and each neuron in a hidden layer has multiple outputs. For each neuron's I inputs and one output, a weighted sum is obtained and then passed through a non-linear activation function, resulting in the following mapping between the inputs and outputs: It is an activation function. It is the neuron's first Enter the corresponding weights. It is the neuron's first enter, This is the bias constant of this neuron; The coefficients of the deep neural network are trained using the backpropagation algorithm of the deep neural network. The network structure, initial values ​​of network parameters, activation function and loss function are set. The algorithm is iterated with the goal of minimizing the loss between the output of the deep neural network and the ideal output, and the final values ​​of the weights and bias parameters of the deep neural network are obtained. Step 6: Load the trained weights and bias parameters into the deep neural network as a post-compensator. The subsequent radar detection signals will first be denoised by wavelet before being input into the post-compensator, and finally output a high-performance signal.

2. The digital enhancement method for extreme value measurement of high-power field sources as described in claim 1, characterized in that, In step one, the wavelet decomposition layer is selected as 9 layers. During each layer decomposition, the known filter coefficients of the previous layer are convolved with the high-pass and low-pass filters respectively, and downsampled at intervals to obtain the detail coefficients of each layer that approximate the original signal and the approximate coefficients containing noise components, which are used as wavelet coefficients.

3. The digital enhancement method for extreme value measurement of high-power field sources as described in claim 1, characterized in that, In step two, the threshold estimation adopts a fixed threshold criterion, an adaptive threshold criterion, a heuristic threshold criterion, or a maximum / minimum threshold criterion.

4. The digital enhancement method for extreme value measurement of high-power field sources as described in claim 3, characterized in that, The threshold calculated using a fixed-form threshold criterion The expression is: in, The standard deviation of the noise. This represents the median of the coefficients after wavelet decomposition. The length of each coefficient; Then, the coefficients are quantized according to the soft threshold function, with the threshold rules as follows: in, This represents the wavelet coefficients after wavelet decomposition.

5. The digital enhancement method for extreme value measurement of high-power field sources as described in claim 1, characterized in that, Step four specifically involves: First, the denoised signal y(t) is reconstructed to obtain the matrix B to be decomposed; then, singular value decomposition is performed on matrix B. Specifically, it involves denoising the signal after wavelet denoising at N consecutive points. Embedded line by line Within a 3D matrix: Treatment Compensation Matrix Perform SVD: Where v = min(p, q), and U and V are both singular matrices. It is a diagonal matrix, and the elements in the matrix are the singular values ​​of matrix B; Then, retaining the first r large singular values ​​unchanged while setting the remaining vr small singular values ​​to zero, a new main diagonal matrix is ​​constructed. Multiplying the new diagonal matrix back into the original left and right singular matrices yields the new signal matrix. The new main diagonal matrix Substituting into the equation, we get: Finally, the new signal matrix is ​​expanded into a one-dimensional sequence by connecting the rows end to end to obtain the N-point large signal L(t).

Images