Millimeter wave large dynamic adaptive time-frequency modeling method based on deep learning
By adopting a deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method, the subjectivity and wear problems of traditional gun ballistic performance evaluation are solved, achieving more accurate and reliable evaluation and possessing good generalization ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-27
AI Technical Summary
Traditional methods for evaluating the ballistic performance of guns and cannons rely on manual measurement, which is subjective and lacks generality, resulting in inaccurate evaluation results and wear and tear on the guns and cannons.
A deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method is adopted. By cleaning and filtering the external radar signal data, a dataset is constructed and a deep neural network model is trained to predict the ballistic performance of guns.
It improves the accuracy and reliability of ballistic performance assessment within guns and cannons, reduces wear and tear on guns and cannons, and has better generalization capabilities.
Smart Images
Figure CN121744532A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a millimeter wave large dynamic adaptive time-frequency modeling method based on deep learning. BACKGROUND
[0002] Gun interior ballistic performance evaluation technology plays an important role in predicting the muzzle velocity of the projectile and improving the accuracy of the strike, however, due to the difficulty in obtaining data and the large parameter measurement error, the gun interior ballistic performance evaluation brings not small challenges.
[0003] The traditional gun interior ballistic performance evaluation method determines the change of the interior ballistic performance under any state by measuring the change of the inner diameter of the barrel characteristic section, and gives an interior ballistic performance evaluation empirical formula. Although the traditional measurement method can evaluate the gun performance to a certain extent, it is strongly subjective and empirical in judging the measurement results, and is not general. SUMMARY
[0004] The application aims to provide a millimeter wave large dynamic adaptive time-frequency modeling method based on deep learning, which solves the problems of subjectivity and lack of generality in the existing gun interior ballistic performance evaluation, and makes the evaluation method more accurate and general.
[0005] In order to achieve the purpose of the application, the application provides a millimeter wave large dynamic adaptive time-frequency modeling method based on deep learning, comprising the following steps:
[0006] Step 1, data cleaning is performed on the measured baseband Doppler signal data of the outfield radar to obtain a barrel projectile velocity curve, and then a polynomial fitting is performed to obtain a barrel projectile motion model; a return signal mathematical model is obtained by using the barrel projectile motion model, and a noise model is superimposed to obtain an original return signal model of the simulated outfield test environment;
[0007] Step 2, the original return signal model of the simulated outfield test environment is down-converted to obtain an original baseband Doppler signal, and then an adaptive filter is used to obtain a baseband Doppler signal filtered of noise;
[0008] Step 3, the baseband Doppler signal filtered of noise is converted into a projectile velocity curve, input features and output features are extracted from the projectile velocity curve, part of the samples are obtained, the input feature parameters are changed on the basis of the part of the sample set, a data set containing the input features and the output features is obtained, and the data set includes a training set and a test set;
[0009] Step 4, input the training set into the built deep neural network model, train the neural network model, obtain the trained deep neural network model, test the trained neural network model by using the test set, and finally obtain the deep neural network model for predicting the gun interior ballistic performance.
[0010] Compared with the prior art, the significant progress of the present application is that:
[0011] 1) In the consideration of constructing the data set, the present application pushes the gun interior ballistic performance index characteristics through the projectile muzzle velocity characteristic curve, avoids the error caused by manual measurement of the gun, and improves the accuracy of the gun interior ballistic performance evaluation;
[0012] 2) The present application avoids the related processing of the gun barrel in the traditional measurement method, maximally reduces the wear degree of the experiment itself on the gun interior ballistic, and makes the research content and result reliable;
[0013] 3) The present application builds a deep neural network, carries out deep learning based on the training data set, tests the reliability and accuracy of the deep neural network model based on the test data set, learns the velocity curve characteristics, predicts the gun interior ballistic performance classification, each step is strictly followed by a formula, avoids the subjectivity caused by manual measurement of the gun interior ballistic physical parameters, and can also well predict the gun interior ballistic performance characteristics of the sample data which has not been learned, so that the research result has more generalization ability.
[0014] In order to more clearly illustrate the function characteristics and structure parameters of the present application, the following further illustrates the present application in combination with the drawings and specific embodiments. DETAILED DESCRIPTION
[0015] The drawings described herein are used to provide further understanding of the present application, constitute a part of the present application, and the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:
[0016] Figure 1 is a step flow chart of the present application;
[0017] Figure 2 is an external field speed measurement system of the present application;
[0018] Figure 3 is a step flow chart of the analog signal source design of the present application;
[0019] Figure 4 is a velocity curve diagram of the analog signal source process fitting of the present application;
[0020] Figure 5 is a self-adaptive filtering result diagram of the present application;
[0021] Figure 6 is a deep neural network model flowchart based on the Softmax classifier of the present application;
[0022] Figure 7 is a deep neural network training result diagram of the present application. DETAILED DESCRIPTION
[0023] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0024] A millimeter wave large dynamic self-adaptive time-frequency modeling method based on deep learning of the present application combines Figure 1 , comprising the following steps:
[0025] Step 1, data cleaning is performed on the measured baseband Doppler signal data of the outfield radar to obtain a velocity curve of the projectile in the bore, and then a polynomial fitting is performed to obtain a projectile motion model in the bore; a mathematical model of the echo signal is obtained by using the projectile motion model in the bore, and a noise model is superimposed to obtain an original echo signal model of the simulated outfield test environment;
[0026] Step 2, the original echo signal model of the simulated outfield test environment is down-converted to obtain an original baseband Doppler signal (containing noise), and then an adaptive filter is used to obtain a baseband Doppler signal after filtering noise;
[0027] Step 3, based on the Doppler principle, the baseband Doppler signal after filtering noise is converted into a projectile velocity curve, input features and output features are extracted from the projectile velocity curve, and part of the sample is obtained; on the basis of the part of the sample set, the input feature parameters are changed to quickly obtain a large number of data sets containing input features and output features, the data sets including a training set and a test set;
[0028] Step 4, the training set is input into the built deep neural network model, the neural network model is trained to obtain a trained deep neural network model, the trained neural network model is tested by using the test set, the parameters of the deep neural network model are modified until the accuracy reaches the standard, and finally a deep neural network model for predicting the interior ballistic performance of a gun is obtained.
[0029] In combination with Figure 2 and Figure 3 , the data acquisition system parameters include: a sampling rate of 10MHz and a sampling time of 0 to 10ms, and the step 1 includes the following steps:
[0030] Step 1-1, data cleaning is performed on the measured baseband Doppler signal data of the external field radar, measurement error data is removed, the cleaned Doppler baseband signal is converted into a discrete in-bore projectile velocity curve based on the Doppler principle, the velocity curve is fitted by using a polynomial, and an in-bore projectile motion model is obtained;
[0031] Specifically as follows:
[0032] ;
[0033] Wherein, is the interior ballistic velocity of the projectile, A is the fitting order, is the i-th A-order polynomial fitting coefficient, is the i-order independent variable;
[0034] Step 1-2, based on the in-bore projectile motion model and the radar transmitting signal model, an echo signal model is obtained, and then a projectile echo signal model is obtained;
[0035] The radar transmitting signal model is a single-frequency continuous wave signal :
[0036] ;
[0037] Wherein, is the signal amplitude, is the millimeter wave radar operating frequency, is the initial phase of the signal, is the imaginary unit, is the transmission time;
[0038] The radar receiving projectile echo signal is obtained through the radar transmitting signal model :
[0039] ;
[0040] Wherein, is the amplitude attenuation, is a delay function related to the velocity of the projectile: , wherein is the initial distance between the projectile and the radar, is the speed of light;
[0041] Then the echo signal is obtained :
[0042] ;
[0043] Let the carrier signal be :
[0044] ;
[0045] The final projectile echo signal model is:
[0046] ;
[0047] Step 1-3, superimpose the projectile echo signal model on a noise mathematical model,
[0048] The echo signal received by the radar inevitably exists thermal noise signal in the receiver, and the noise of the continuous wave speed measurement radar is derived from atmospheric noise, phase noise and amplitude quantization noise. The atmospheric thermal noise is reflected in the noise coefficient of the radar front-end receiver, and is specifically shown in the following formula:
[0049]
[0050] In the formula, is the Boltzmann constant, , is the equivalent thermal noise temperature 290K, is the radar obtained working bandwidth, is the receiver noise of the radar;
[0051] The phase noise is reflected in the noise coefficient of the radar transmitter, and is specifically shown in the following formula:
[0052]
[0053] In the formula, is the transmission signal leakage degree, is the noise accumulation;
[0054] The amplitude quantization noise is mainly determined by the AD sampling; the present application adopts 10MHz, 14-bit AD for sampling, and the amplitude quantization noise is specifically shown in the following formula:
[0055]
[0056] According to the calculation, the noise influence can be ignored;
[0057] Therefore, the Gaussian distribution model is adopted to simulate the probability density function noise model of the system noise:
[0058] ;
[0059] In the formula, is the standard deviation of the noise signal, is the noise random variable;
[0060] The projectile echo signal model is superimposed on the noise model to obtain the original echo signal model of the simulated field test environment.
[0061] Finally, the above projectile echo signal model is superimposed with a noise mathematical model to obtain an actual echo signal, the echo signal is down-converted to an intermediate frequency signal, i.e. an actual Doppler signal, and based on the Doppler principle, an actual speed curve is obtained, as shown in Figure 4
[0062] The step 2 includes the following steps:
[0063] Step 2-1, converting the original echo signal model of the simulated outfield test environment into a noise-containing baseband Doppler signal through down-conversion;
[0064] Step 2-2, performing windowing processing on the noise-containing baseband Doppler signal to truncate the signal and obtain a noise-containing segmented Doppler signal;
[0065] Step 2-3, acquiring a pure noise signal having coherence with the original noise signal through another same antenna, obtaining an estimated noise signal through normalized least mean square filtering (NLMS), and obtaining a baseband Doppler signal after filtering noise by the difference between the original signal and the estimated noise signal.
[0066] The windowing processing of the step 2-2 adopts a rectangular window function, the Doppler signal output by the analog signal source has a sampling frequency of 10 MHz and a sampling time of 10 ms, i.e. a sampling number of 10 5 ; before spectrum analysis, the analog radar signal needs to be truncated to realize segmented analysis of the signal spectrum data, and the frequency resolution needs to be considered in the spectrum analysis process, as shown in the following formula:
[0067] ;
[0068] wherein, is the frequency resolution, is the sampling frequency, is the sampling number, is the effective length.
[0069] Considering the number of truncated segments and the frequency resolution of the Doppler signal, if the length of the rectangular window function is selected as 0.2 ms, the frequency resolution is 5 kHz, which can effectively reduce the influence of low-frequency aliasing.
[0070] The step 2-3 is specifically as shown in the following formula:
[0071] ;
[0072] ;
[0073] ;
[0074] ;
[0075] in, For Normalized Least Mean Square (NLMS) filter output, For the filter weight vector, This is the transpose of the weight vector. The filter input vector, For error signals, Let z be the reference noise signal, and z be the sampling point. The length of the weight vector. This refers to the index in the corresponding row vector. The filter step size, Step size factor Here is the regularization constant;
[0076] The adaptive filtering results are shown below. Figure 5 The first column of images shows the time-domain and frequency-domain images of a segmented Doppler signal, respectively. The second column shows the time-domain and frequency-domain images after adding noise. The third column shows the time-domain and frequency-domain images after filtering. By comparison, this method achieves better adaptive filtering for signals with poor signal-to-noise ratio.
[0077] Ultimately, by performing down-conversion, windowing, and NLMS filtering on the original echo signal, adaptive filtering of the original echo signal was achieved, effectively removing noise interference while preserving the original large dynamic characteristics of the Doppler signal.
[0078] The specific input features for step 3 are shown in the following formula:
[0079] ;
[0080] ;
[0081] ;
[0082] ;
[0083] in, When the projectile's velocity is greater than zero, The moment when the gunpowder begins to burn. For the time it takes for gunpowder to burn, The moment when the projectile's velocity reaches its maximum. To increase the duration of the bullet, The maximum acceleration of the projectile. This is the maximum sustained velocity of the projectile;
[0084] The input features are obtained based on the above four parameters.
[0085] Based on the input features of the velocity curve of each gun, the internal ballistic performance of the gun is analyzed, the output features are extracted, the internal ballistic performance classification corresponding to the projectile velocity curve is obtained, and one-hot encoding is performed. Each projectile velocity curve is labeled to obtain a structured dataset. The output features include: whether the gunpowder combustion rate is normal, whether the maximum chamber pressure of the internal ballistics meets the standard, and whether the internal ballistic friction coefficient is normal.
[0086] The logic for determining the output feature is as follows:
[0087] (1) Burning time of gunpowder If the gunpowder burning speed is less than or equal to the average value, the burning time is considered normal. If the speed is greater than the average, the gunpowder combustion speed is considered abnormal.
[0088] (2) Bullet acceleration time Less than or equal to the average value, and maximum holding speed When the maximum internal ballistic chamber pressure is greater than or equal to the average value, it is considered to meet the standard. When only one of the two conditions is met or neither condition is met, the maximum internal ballistic chamber pressure is considered to be substandard.
[0089] (3) Maximum acceleration of the projectile If the internal ballistic friction coefficient is greater than or equal to the average value, it is considered normal, and the maximum acceleration of the projectile is... If the coefficient of friction is less than the average value, the internal ballistic friction coefficient is considered abnormal.
[0090] The above describes the logic for determining the output metrics. Each metric outputs True if it meets the standard or is normal, and False if it does not meet the standard or is abnormal. Combining the output results of each metric yields the final result. Classification, using These eight numbers represent the relationships as follows:
[0091] 0: The gunpowder combustion rate is normal, the maximum chamber pressure of the internal ballistics meets the standard, and the internal ballistic friction coefficient is normal.
[0092] 1: The gunpowder combustion rate is normal, the maximum chamber pressure of the internal ballistics meets the standard, but the internal ballistic friction coefficient is abnormal.
[0093] 2: The gunpowder combustion rate is normal, the maximum chamber pressure of the internal ballistics is below standard, and the internal ballistic friction coefficient is normal.
[0094] 3: The gunpowder combustion rate is normal, but the maximum chamber pressure of the internal ballistics is below standard, and the internal ballistic friction coefficient is abnormal.
[0095] 4: The gunpowder combustion rate is abnormal, but the maximum chamber pressure of the internal ballistics meets the standard, and the internal ballistics friction coefficient is normal.
[0096] 5: The gunpowder combustion rate is abnormal, the maximum chamber pressure of the internal ballistics meets the standard, but the internal ballistics friction coefficient is abnormal.
[0097] 6: The gunpowder combustion rate is abnormal, the maximum chamber pressure of the internal ballistics is below standard, and the internal ballistic friction coefficient is normal.
[0098] 7: Abnormal gunpowder combustion rate, insufficient maximum chamber pressure, and abnormal internal ballistic friction coefficient.
[0099] Each projectile velocity curve corresponds to one type of output. The output is encoded using 8-bit one-hot encoding. If the condition is met, the output is 1; otherwise, the output is 0. Each velocity curve is labeled, and a structured dataset is obtained. The dataset includes a training set and a test set.
[0100] The training set in step 4 is used to calculate the output through forward propagation. The loss function is selected as cross-entropy loss. The weights and biases are updated by backpropagation based on the gradient descent algorithm. The iteration continues until the loss function converges, and a trained deep neural network is obtained.
[0101] The constructed neural network includes an input layer, three linear layers, two activation function layers, and an output layer; the input layer has 4 inputs and the output layer has 8 outputs.
[0102] like Figure 6 As shown, the Linear Layer is a linear layer, the Sigmoid Layer is an activation function layer, the Input Layer is the input layer, and the Softmax Layer is a Softmax classifier layer. The output layer uses a Softmax classifier and outputs a distribution whose features include:
[0103] ;
[0104] ;
[0105] in, This is the output of the Softmax Layer. For specific category numbers, The total number of categories;
[0106] The final output features include: outputting the probability of each category, all of which are greater than or equal to 0, and the sum of all probabilities is 1.
[0107] The construction of the deep neural network is shown in the following formula:
[0108] ;
[0109] ;
[0110] ;
[0111] ;
[0112] in, For the first The first activation function layer One parameter, , For the corresponding weights and biases, For the first The first linear layer One parameter, For the first The number of parameters in each activation function layer For the first The output of each activation function layer corresponds to the output of a complete linear layer and its activation function layer. This is the output of the last linear layer. To ensure that the probability of each category in the output is greater than zero, we take the probability of each output... The index is calculated and normalized to obtain the probability of belonging to each category. For loss function, This is the result of one-hot encoding. This represents the probability corresponding to the j-th category.
[0113] The test set samples are input into the trained deep neural network model to obtain the predicted output, which is compared with the theoretical value to calculate the accuracy. The parameters of the deep neural network model are then modified until the final accuracy reaches more than 90%. The output is a deep neural network model that can be used to evaluate the internal ballistic performance of guns and cannons with new velocity curves.
[0114] The first linear layer has 4 inputs and 8 outputs, the second linear layer has 8 inputs and 16 outputs, and the third linear layer has 16 inputs and 8 outputs.
[0115] Furthermore, in each epoch, the input features are fed into the deep neural network for forward propagation to obtain the loss function. This loss function is then used to iterate through the hidden layer parameters of the deep neural network via backpropagation until it converges. Figure 7 As shown, the accuracy is as high as 96.65%, which can accurately assess the performance of guns and cannons without losing generality.
[0116] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method, characterized in that, Includes the following steps: Step 1: Clean the measured baseband Doppler signal data of the field radar to obtain the velocity curve of the projectile in the barrel, and then obtain the motion model of the projectile in the barrel through polynomial fitting; use the motion model of the projectile in the barrel to obtain the mathematical model of the echo signal, and superimpose the noise model to obtain the original echo signal model simulating the field test environment. Step 2: Down-convert the original echo signal model of the simulated field test environment to obtain the original baseband Doppler signal, and then pass it through an adaptive filter to obtain the baseband Doppler signal after noise removal. Step 3: Convert the noise-filtered baseband Doppler signal into a projectile velocity curve, extract input and output features from the projectile velocity curve, obtain a partial sample, and on the basis of the partial sample set, change the input feature parameters to obtain a dataset containing input and output features. The dataset includes a training set and a test set. Step 4: Input the training set into the constructed deep neural network model, train the neural network model, obtain the trained deep neural network model, test the trained neural network model using the test set, and finally obtain a deep neural network model for predicting the ballistic performance of guns and cannons.
2. The deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1-1: Clean the measured baseband Doppler signal data of the field radar to remove erroneous data. Based on the Doppler principle, convert the cleaned Doppler baseband signal into a discrete velocity curve of the projectile in the barrel. Use a polynomial to fit the velocity curve to obtain the motion model of the projectile in the barrel. The specific formula is as follows: ； in, Let A be the projectile velocity within the trajectory, and A be the fitting order. The fitting coefficients are those of the i-th order A polynomial. Let i be the independent variable of order i; Steps 1-2: Based on the in-bore projectile motion model and the radar transmission signal model, the echo signal model is obtained, and then the projectile echo signal model is obtained. The radar transmission signal model is a single-frequency continuous wave signal. : ； in, The signal amplitude, For millimeter-wave radar operating frequency, The initial phase of the signal. The imaginary unit, Launch time; The echo signal of the radar-received projectile is obtained through the radar transmission signal model. : ； in, For amplitude attenuation, The delay function is related to the projectile velocity: ,in The initial distance between the projectile and the radar. The speed of light; The echo signal was then processed and obtained. : ； Assume carrier signal : ； The final projectile echo signal model is as follows: ； Steps 1-3: Superimpose the projectile echo signal model with the noise mathematical model, and use a Gaussian distribution model to simulate the probability density function noise model of the system noise: ； in, The standard deviation of the noise signal. It is a noise random variable; The original echo signal model simulating the field test environment is obtained by superimposing the projectile echo signal model with the noise model.
3. The deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method according to claim 2, characterized in that, Step 2 includes the following steps: Step 2-1: Convert the original echo signal model of the simulated field test environment into a noisy baseband Doppler signal through down-conversion. Step 2-2: Window the noisy baseband Doppler signal to truncate the signal and obtain a noisy segmented Doppler signal; Steps 2-3: Acquire a pure noise signal that is coherent with the original noise signal through another identical antenna, obtain the estimated noise signal by normalized least mean square filtering, and obtain the baseband Doppler signal after noise removal by the difference between the estimated noise signal and the original signal.
4. The deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method according to claim 3, characterized in that, The windowing process in step 2-2 uses a rectangular window function, and its frequency resolution is shown in the following formula: ； in, For frequency resolution, Sampling frequency, The number of samples, This is the effective length.
5. The deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method according to claim 4, characterized in that, Choosing a rectangular window function length of 0.2ms results in a frequency resolution of 5kHz.
6. The deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method according to claim 3, characterized in that, The specific steps 2-3 are shown in the following formula: ； ； ； ； in, This is the output of the normalized minimum mean square filter. For the filter weight vector, This is the transpose of the weight vector. The filter input vector, For error signals, Let z be the reference noise signal, and z be the sampling point. The length of the weight vector. This refers to the index in the corresponding row vector. The filter step size, Step size factor is the regularization constant.
7. The deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method according to claim 6, characterized in that, The specific input features for step 3 are shown in the following formula: ； ； ； ； in, When the projectile's velocity is greater than zero, The moment when the gunpowder begins to burn. For the time it takes for gunpowder to burn, The moment when the projectile's velocity reaches its maximum. To increase the duration of the bullet, The maximum acceleration of the projectile. This is the maximum sustained velocity of the projectile; The input features are obtained based on the above four parameters.
8. The deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method according to claim 7, characterized in that, Based on the input features of the velocity curve of each gun, the internal ballistic performance of the gun is analyzed, the output features are extracted, the internal ballistic performance classification corresponding to the projectile velocity curve is obtained, and one-hot encoding is performed. A label is set for each projectile velocity curve to obtain a structured dataset. The output characteristics include: whether the gunpowder combustion rate is normal, whether the maximum chamber pressure of the internal ballistics meets the standard, and whether the internal ballistics friction coefficient is normal.
9. The deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method according to claim 7, characterized in that, The training set in step 4 is used to calculate the output through forward propagation. The loss function is selected as cross-entropy loss. The weights and biases are updated by backpropagation based on the gradient descent algorithm. The iteration continues until the loss function converges, and a trained deep neural network is obtained. The constructed neural network includes an input layer, an output layer, linear layers, and activation function layers; The output layer employs a Softmax classifier, outputting a distribution whose features include: ； ； in, This is the output of the Softmax Layer. For specific category numbers, The total number of categories; The final output features include: outputting the probability of each category, all of which are greater than or equal to 0, and the sum of all probabilities is 1.
10. A deep learning-based millimeter-wave large dynamic adaptive time-frequency modeling method according to claim 9, characterized in that, The deep neural network is specifically shown in the following formula: ； ； ； ； in, For the first The first activation function layer One parameter, , For the corresponding weights and biases, For the first The first linear layer One parameter, For the first The number of parameters in each activation function layer For the first The output of each activation function layer corresponds to the output of a complete linear layer and its activation function layer. This is the output of the last linear layer. To ensure that the probability of each category in the output is greater than zero, we take the probability of each output... The index is calculated and normalized to obtain the probability of belonging to each category. For loss function, This is the result of one-hot encoding. This represents the probability corresponding to the j-th category.